diff --git "a/ParakeetEncoder_v2.mlmodelc/model.mil" "b/ParakeetEncoder_v2.mlmodelc/model.mil"
--- "a/ParakeetEncoder_v2.mlmodelc/model.mil"
+++ "b/ParakeetEncoder_v2.mlmodelc/model.mil"
@@ -1,14 +1,14 @@
 program(1.0)
-[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3405.2.1"}})]
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}})]
 {
-    func main<ios16>(tensor<fp32, [1, 128, 1001]> melspectogram, tensor<int32, [1]> melspectogram_length) {
+    func main<ios16>(tensor<fp32, [1, 128, 1001]> audio_signal, tensor<int32, [1]> length) {
             tensor<int32, []> var_25 = const()[name = tensor<string, []>("op_25"), val = tensor<int32, []>(-1)];
             tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
-            tensor<string, []> melspectogram_to_fp16_dtype_0 = const()[name = tensor<string, []>("melspectogram_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<string, []> audio_signal_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_signal_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
             tensor<string, []> cast_0_to_fp16_dtype_0 = const()[name = tensor<string, []>("cast_0_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
             tensor<fp16, []> var_82_promoted_to_fp16 = const()[name = tensor<string, []>("op_82_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
-            tensor<fp16, [1]> melspectogram_length_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = melspectogram_length)[name = tensor<string, []>("cast_3")];
-            tensor<fp16, [1]> var_83_cast_fp16 = add(x = melspectogram_length_to_fp16, y = var_82_promoted_to_fp16)[name = tensor<string, []>("op_83_cast_fp16")];
+            tensor<fp16, [1]> length_to_fp16 = cast(dtype = cast_0_to_fp16_dtype_0, x = length)[name = tensor<string, []>("cast_3")];
+            tensor<fp16, [1]> var_83_cast_fp16 = add(x = length_to_fp16, y = var_82_promoted_to_fp16)[name = tensor<string, []>("op_83_cast_fp16")];
             tensor<fp16, []> _inversed_85_y_0_to_fp16 = const()[name = tensor<string, []>("_inversed_85_y_0_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1]> _inversed_85_cast_fp16 = mul(x = var_83_cast_fp16, y = _inversed_85_y_0_to_fp16)[name = tensor<string, []>("_inversed_85_cast_fp16")];
             tensor<fp16, []> var_86_to_fp16 = const()[name = tensor<string, []>("op_86_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
@@ -29,8 +29,8 @@ program(1.0)
             tensor<fp16, [1]> lengths_13_cast_fp16 = add(x = _inversed_101_cast_fp16, y = var_102_to_fp16)[name = tensor<string, []>("lengths_13_cast_fp16")];
             tensor<fp16, [1]> lengths_cast_fp16 = floor(x = lengths_13_cast_fp16)[name = tensor<string, []>("lengths_cast_fp16")];
             tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1, 128, 1001]> melspectogram_to_fp16 = cast(dtype = melspectogram_to_fp16_dtype_0, x = melspectogram)[name = tensor<string, []>("cast_2")];
-            tensor<fp16, [1, 1001, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = melspectogram_to_fp16)[name = tensor<string, []>("transpose_314")];
+            tensor<fp16, [1, 128, 1001]> audio_signal_to_fp16 = cast(dtype = audio_signal_to_fp16_dtype_0, x = audio_signal)[name = tensor<string, []>("cast_2")];
+            tensor<fp16, [1, 1001, 128]> x_1_cast_fp16 = transpose(perm = x_1_perm_0, x = audio_signal_to_fp16)[name = tensor<string, []>("transpose_314")];
             tensor<fp16, [1, 1, 1001, 128]> input_1_cast_fp16 = expand_dims(axes = input_1_axes_0, x = x_1_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
             tensor<string, []> input_3_pad_type_0 = const()[name = tensor<string, []>("input_3_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [4]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
@@ -131,14 +131,14 @@ program(1.0)
             tensor<int32, [4]> var_252 = const()[name = tensor<string, []>("op_252"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_1_cast_fp16 = reshape(shape = var_252, x = linear_5_cast_fp16)[name = tensor<string, []>("v_1_cast_fp16")];
             tensor<int32, [4]> value_1_perm_0 = const()[name = tensor<string, []>("value_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_0_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_0_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15918464)))];
-            tensor<fp16, [1, 126, 8, 128]> var_264_cast_fp16 = add(x = q_1_cast_fp16, y = model_layers_0_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_264_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_0_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_0_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15920576)))];
-            tensor<fp16, [1, 126, 8, 128]> var_266_cast_fp16 = add(x = q_1_cast_fp16, y = model_layers_0_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_266_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_0_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15918464))), scale = tensor<fp16, [8]>([0x1.674p-8, 0x1.208p-7, 0x1.778p-8, 0x1.d7cp-8, 0x1.a3p-9, 0x1.02cp-7, 0x1.308p-8, 0x1.4fcp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_264_cast_fp16 = add(x = q_1_cast_fp16, y = model_layers_0_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_264_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_0_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15919552))), scale = tensor<fp16, [8]>([0x1.7bp-10, 0x1.ff4p-9, 0x1.4d4p-9, 0x1.0fcp-8, 0x1.71p-10, 0x1.64cp-9, 0x1.f6cp-10, 0x1.06p-9]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_266_cast_fp16 = add(x = q_1_cast_fp16, y = model_layers_0_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_266_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_268_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_268_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15922688))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16180096))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_268_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_268_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15920640))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16178048))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_1_cast_fp16 = transpose(perm = q_with_bias_v_1_perm_0, x = var_266_cast_fp16)[name = tensor<string, []>("transpose_311")];
             tensor<fp16, [1, 8, 126, 251]> x_7_cast_fp16 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1_cast_fp16, y = op_268_to_fp16_quantized)[name = tensor<string, []>("x_7_cast_fp16")];
             tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -182,12 +182,12 @@ program(1.0)
             tensor<int32, [3]> var_301 = const()[name = tensor<string, []>("op_301"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_300_cast_fp16 = transpose(perm = var_300_perm_0, x = x_13_cast_fp16)[name = tensor<string, []>("transpose_307")];
             tensor<fp16, [1, 126, 1024]> input_35_cast_fp16 = reshape(shape = var_301, x = var_300_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16180672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17229312))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16178624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17227264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_7_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_0_self_attn_linear_out_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_39_cast_fp16 = add(x = input_31_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
             tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17231424)))];
-            tensor<fp16, [1024]> model_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17233536)))];
+            tensor<fp16, [1024]> model_layers_0_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17229376)))];
+            tensor<fp16, [1024]> model_layers_0_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17231488)))];
             tensor<fp16, [1, 126, 1024]> x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = model_layers_0_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_0_norm_conv_weight_to_fp16, x = input_39_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
             tensor<int32, [3]> input_41_perm_0 = const()[name = tensor<string, []>("input_41_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_43_pad_type_0 = const()[name = tensor<string, []>("input_43_pad_type_0"), val = tensor<string, []>("valid")];
@@ -195,7 +195,7 @@ program(1.0)
             tensor<int32, [2]> input_43_pad_0 = const()[name = tensor<string, []>("input_43_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_43_dilations_0 = const()[name = tensor<string, []>("input_43_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_43_groups_0 = const()[name = tensor<string, []>("input_43_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17235648))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19334976))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17233600))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332928))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_41_cast_fp16 = transpose(perm = input_41_perm_0, x = x_17_cast_fp16)[name = tensor<string, []>("transpose_306")];
             tensor<fp16, [1, 2048, 126]> input_43_cast_fp16 = conv(dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = model_layers_0_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
             tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -215,8 +215,8 @@ program(1.0)
             tensor<int32, [1]> input_49_strides_0 = const()[name = tensor<string, []>("input_49_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_49_pad_0 = const()[name = tensor<string, []>("input_49_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_49_dilations_0 = const()[name = tensor<string, []>("input_49_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_248_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_248_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19339136))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19348416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_249_to_fp16 = const()[name = tensor<string, []>("const_249_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19350528)))];
+            tensor<fp16, [1024, 1, 9]> const_248_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_248_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19337088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19346368))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_249_to_fp16 = const()[name = tensor<string, []>("const_249_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19348480)))];
             tensor<fp16, [1, 1024, 126]> input_51_cast_fp16 = conv(bias = const_249_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = const_248_to_fp16_quantized, x = input_47_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_53_cast_fp16 = silu(x = input_51_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")];
             tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
@@ -224,64 +224,64 @@ program(1.0)
             tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<string, []>("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19352640))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20401280))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19350592))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20399232))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_21_cast_fp16 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = model_layers_0_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
             tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_55_cast_fp16 = transpose(perm = input_55_perm_0, x = x_21_cast_fp16)[name = tensor<string, []>("transpose_305")];
             tensor<fp16, [1, 126, 1024]> input_57_cast_fp16 = add(x = input_39_cast_fp16, y = input_55_cast_fp16)[name = tensor<string, []>("input_57_cast_fp16")];
             tensor<int32, [1]> input_59_axes_0 = const()[name = tensor<string, []>("input_59_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20403392)))];
-            tensor<fp16, [1024]> model_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20405504)))];
+            tensor<fp16, [1024]> model_layers_0_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20401344)))];
+            tensor<fp16, [1024]> model_layers_0_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20403456)))];
             tensor<fp16, [1, 126, 1024]> input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = model_layers_0_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_0_norm_feed_forward2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20407616))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24601984))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_0_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20405568))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24599936))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_feed_forward2_linear1_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_63_cast_fp16 = silu(x = linear_8_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24610240))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28804608))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_0_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_0_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24608192))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28802560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_9_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_0_feed_forward2_linear2_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
             tensor<fp16, []> var_361_to_fp16 = const()[name = tensor<string, []>("op_361_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_362_cast_fp16 = mul(x = linear_9_cast_fp16, y = var_361_to_fp16)[name = tensor<string, []>("op_362_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_69_cast_fp16 = add(x = input_57_cast_fp16, y = var_362_cast_fp16)[name = tensor<string, []>("input_69_cast_fp16")];
             tensor<int32, [1]> input_71_axes_0 = const()[name = tensor<string, []>("input_71_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28806720)))];
-            tensor<fp16, [1024]> model_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28808832)))];
+            tensor<fp16, [1024]> model_layers_0_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28804672)))];
+            tensor<fp16, [1024]> model_layers_0_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_0_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28806784)))];
             tensor<fp16, [1, 126, 1024]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = model_layers_0_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_0_norm_out_weight_to_fp16, x = input_69_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
             tensor<int32, [1]> input_73_axes_0 = const()[name = tensor<string, []>("input_73_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28810944)))];
-            tensor<fp16, [1024]> model_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28813056)))];
+            tensor<fp16, [1024]> model_layers_1_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28808896)))];
+            tensor<fp16, [1024]> model_layers_1_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28811008)))];
             tensor<fp16, [1, 126, 1024]> input_73_cast_fp16 = layer_norm(axes = input_73_axes_0, beta = model_layers_1_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_1_norm_feed_forward1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor<string, []>("input_73_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28815168))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33009536))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_1_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28813120))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33007488))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_feed_forward1_linear1_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_77_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor<string, []>("input_77_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33017792))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37212160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_1_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33015744))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37210112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_11_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_1_feed_forward1_linear2_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
             tensor<fp16, []> var_390_to_fp16 = const()[name = tensor<string, []>("op_390_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_391_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_390_to_fp16)[name = tensor<string, []>("op_391_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_83_cast_fp16 = add(x = input_71_cast_fp16, y = var_391_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
             tensor<int32, [1]> query_3_axes_0 = const()[name = tensor<string, []>("query_3_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37214272)))];
-            tensor<fp16, [1024]> model_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37216384)))];
+            tensor<fp16, [1024]> model_layers_1_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37212224)))];
+            tensor<fp16, [1024]> model_layers_1_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37214336)))];
             tensor<fp16, [1, 126, 1024]> query_3_cast_fp16 = layer_norm(axes = query_3_axes_0, beta = model_layers_1_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_1_norm_self_att_weight_to_fp16, x = input_83_cast_fp16)[name = tensor<string, []>("query_3_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37218496))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38267136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37216448))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38265088))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_12_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_1_self_attn_linear_q_weight_to_fp16_quantized, x = query_3_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
             tensor<int32, [4]> var_407 = const()[name = tensor<string, []>("op_407"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_7_cast_fp16 = reshape(shape = var_407, x = linear_12_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38269248))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39317888))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38267200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39315840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_13_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_1_self_attn_linear_k_weight_to_fp16_quantized, x = query_3_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
             tensor<int32, [4]> var_411 = const()[name = tensor<string, []>("op_411"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_5_cast_fp16 = reshape(shape = var_411, x = linear_13_cast_fp16)[name = tensor<string, []>("k_5_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39320000))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40368640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39317952))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40366592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_1_self_attn_linear_v_weight_to_fp16_quantized, x = query_3_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
             tensor<int32, [4]> var_415 = const()[name = tensor<string, []>("op_415"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_3_cast_fp16 = reshape(shape = var_415, x = linear_14_cast_fp16)[name = tensor<string, []>("v_3_cast_fp16")];
             tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_1_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_1_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40370752)))];
-            tensor<fp16, [1, 126, 8, 128]> var_427_cast_fp16 = add(x = q_7_cast_fp16, y = model_layers_1_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_427_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_1_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_1_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40372864)))];
-            tensor<fp16, [1, 126, 8, 128]> var_429_cast_fp16 = add(x = q_7_cast_fp16, y = model_layers_1_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_429_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_1_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40368704))), scale = tensor<fp16, [8]>([0x1.5ep-8, 0x1.294p-7, 0x1.72cp-7, 0x1.868p-8, 0x1.49cp-7, 0x1.618p-8, 0x1.eecp-8, 0x1.e5p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_427_cast_fp16 = add(x = q_7_cast_fp16, y = model_layers_1_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_427_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_1_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40369792))), scale = tensor<fp16, [8]>([0x1.d78p-8, 0x1.5e8p-8, 0x1.9fcp-9, 0x1.3a4p-9, 0x1.a6p-9, 0x1.934p-11, 0x1.09cp-8, 0x1.5a4p-9]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_429_cast_fp16 = add(x = q_7_cast_fp16, y = model_layers_1_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_429_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_29_transpose_x_0 = const()[name = tensor<string, []>("x_29_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_29_transpose_y_0 = const()[name = tensor<string, []>("x_29_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_431_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_431_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40374976))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40632064))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_431_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_431_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40370880))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40627968))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_3_cast_fp16 = transpose(perm = q_with_bias_v_3_perm_0, x = var_429_cast_fp16)[name = tensor<string, []>("transpose_304")];
             tensor<fp16, [1, 8, 126, 251]> x_29_cast_fp16 = matmul(transpose_x = x_29_transpose_x_0, transpose_y = x_29_transpose_y_0, x = q_with_bias_v_3_cast_fp16, y = op_431_to_fp16_quantized)[name = tensor<string, []>("x_29_cast_fp16")];
             tensor<int32, [8]> x_31_pad_0 = const()[name = tensor<string, []>("x_31_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -321,12 +321,12 @@ program(1.0)
             tensor<int32, [3]> var_464 = const()[name = tensor<string, []>("op_464"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_463_cast_fp16 = transpose(perm = var_463_perm_0, x = x_35_cast_fp16)[name = tensor<string, []>("transpose_300")];
             tensor<fp16, [1, 126, 1024]> input_87_cast_fp16 = reshape(shape = var_464, x = var_463_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40632640))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41681280))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40628544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41677184))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_16_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_1_self_attn_linear_out_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_91_cast_fp16 = add(x = input_83_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
             tensor<int32, [1]> x_39_axes_0 = const()[name = tensor<string, []>("x_39_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41683392)))];
-            tensor<fp16, [1024]> model_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41685504)))];
+            tensor<fp16, [1024]> model_layers_1_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41679296)))];
+            tensor<fp16, [1024]> model_layers_1_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41681408)))];
             tensor<fp16, [1, 126, 1024]> x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = model_layers_1_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_1_norm_conv_weight_to_fp16, x = input_91_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
             tensor<int32, [3]> input_93_perm_0 = const()[name = tensor<string, []>("input_93_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_95_pad_type_0 = const()[name = tensor<string, []>("input_95_pad_type_0"), val = tensor<string, []>("valid")];
@@ -334,7 +334,7 @@ program(1.0)
             tensor<int32, [2]> input_95_pad_0 = const()[name = tensor<string, []>("input_95_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_95_dilations_0 = const()[name = tensor<string, []>("input_95_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_95_groups_0 = const()[name = tensor<string, []>("input_95_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41687616))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43784832))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41683520))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43780736))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_93_cast_fp16 = transpose(perm = input_93_perm_0, x = x_39_cast_fp16)[name = tensor<string, []>("transpose_299")];
             tensor<fp16, [1, 2048, 126]> input_95_cast_fp16 = conv(dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = model_layers_1_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
             tensor<int32, []> x_41_split_num_splits_0 = const()[name = tensor<string, []>("x_41_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -352,8 +352,8 @@ program(1.0)
             tensor<int32, [1]> input_101_strides_0 = const()[name = tensor<string, []>("input_101_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_101_pad_0 = const()[name = tensor<string, []>("input_101_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_101_dilations_0 = const()[name = tensor<string, []>("input_101_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_250_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_250_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43788992))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43798272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_251_to_fp16 = const()[name = tensor<string, []>("const_251_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43800384)))];
+            tensor<fp16, [1024, 1, 9]> const_250_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_250_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43784896))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43794176))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_251_to_fp16 = const()[name = tensor<string, []>("const_251_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43796288)))];
             tensor<fp16, [1, 1024, 126]> input_103_cast_fp16 = conv(bias = const_251_to_fp16, dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = const_250_to_fp16_quantized, x = input_99_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_105_cast_fp16 = silu(x = input_103_cast_fp16)[name = tensor<string, []>("input_105_cast_fp16")];
             tensor<string, []> x_43_pad_type_0 = const()[name = tensor<string, []>("x_43_pad_type_0"), val = tensor<string, []>("valid")];
@@ -361,64 +361,64 @@ program(1.0)
             tensor<int32, [2]> x_43_pad_0 = const()[name = tensor<string, []>("x_43_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_43_dilations_0 = const()[name = tensor<string, []>("x_43_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_43_groups_0 = const()[name = tensor<string, []>("x_43_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43802496))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44851136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43798400))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44847040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_43_cast_fp16 = conv(dilations = x_43_dilations_0, groups = x_43_groups_0, pad = x_43_pad_0, pad_type = x_43_pad_type_0, strides = x_43_strides_0, weight = model_layers_1_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
             tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_107_cast_fp16 = transpose(perm = input_107_perm_0, x = x_43_cast_fp16)[name = tensor<string, []>("transpose_298")];
             tensor<fp16, [1, 126, 1024]> input_109_cast_fp16 = add(x = input_91_cast_fp16, y = input_107_cast_fp16)[name = tensor<string, []>("input_109_cast_fp16")];
             tensor<int32, [1]> input_111_axes_0 = const()[name = tensor<string, []>("input_111_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44853248)))];
-            tensor<fp16, [1024]> model_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44855360)))];
+            tensor<fp16, [1024]> model_layers_1_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44849152)))];
+            tensor<fp16, [1024]> model_layers_1_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44851264)))];
             tensor<fp16, [1, 126, 1024]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = model_layers_1_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_1_norm_feed_forward2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44857472))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49051840))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_1_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44853376))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49047744))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_feed_forward2_linear1_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_115_cast_fp16 = silu(x = linear_17_cast_fp16)[name = tensor<string, []>("input_115_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49060096))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53254464))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_1_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_1_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(49056000))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53250368))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_1_feed_forward2_linear2_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
             tensor<fp16, []> var_524_to_fp16 = const()[name = tensor<string, []>("op_524_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_525_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_524_to_fp16)[name = tensor<string, []>("op_525_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_121_cast_fp16 = add(x = input_109_cast_fp16, y = var_525_cast_fp16)[name = tensor<string, []>("input_121_cast_fp16")];
             tensor<int32, [1]> input_123_axes_0 = const()[name = tensor<string, []>("input_123_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53256576)))];
-            tensor<fp16, [1024]> model_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53258688)))];
+            tensor<fp16, [1024]> model_layers_1_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53252480)))];
+            tensor<fp16, [1024]> model_layers_1_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_1_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53254592)))];
             tensor<fp16, [1, 126, 1024]> input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = model_layers_1_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_1_norm_out_weight_to_fp16, x = input_121_cast_fp16)[name = tensor<string, []>("input_123_cast_fp16")];
             tensor<int32, [1]> input_125_axes_0 = const()[name = tensor<string, []>("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53260800)))];
-            tensor<fp16, [1024]> model_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53262912)))];
+            tensor<fp16, [1024]> model_layers_2_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53256704)))];
+            tensor<fp16, [1024]> model_layers_2_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53258816)))];
             tensor<fp16, [1, 126, 1024]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = model_layers_2_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_2_norm_feed_forward1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor<string, []>("input_125_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53265024))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57459392))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_2_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53260928))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57455296))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_feed_forward1_linear1_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_129_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57467648))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61662016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_2_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57463552))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61657920))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_20_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_2_feed_forward1_linear2_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
             tensor<fp16, []> var_553_to_fp16 = const()[name = tensor<string, []>("op_553_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_554_cast_fp16 = mul(x = linear_20_cast_fp16, y = var_553_to_fp16)[name = tensor<string, []>("op_554_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_135_cast_fp16 = add(x = input_123_cast_fp16, y = var_554_cast_fp16)[name = tensor<string, []>("input_135_cast_fp16")];
             tensor<int32, [1]> query_5_axes_0 = const()[name = tensor<string, []>("query_5_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61664128)))];
-            tensor<fp16, [1024]> model_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61666240)))];
+            tensor<fp16, [1024]> model_layers_2_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61660032)))];
+            tensor<fp16, [1024]> model_layers_2_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61662144)))];
             tensor<fp16, [1, 126, 1024]> query_5_cast_fp16 = layer_norm(axes = query_5_axes_0, beta = model_layers_2_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_2_norm_self_att_weight_to_fp16, x = input_135_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61668352))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62716992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61664256))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62712896))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_21_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_2_self_attn_linear_q_weight_to_fp16_quantized, x = query_5_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
             tensor<int32, [4]> var_570 = const()[name = tensor<string, []>("op_570"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_13_cast_fp16 = reshape(shape = var_570, x = linear_21_cast_fp16)[name = tensor<string, []>("q_13_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62719104))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63767744))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62715008))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63763648))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_2_self_attn_linear_k_weight_to_fp16_quantized, x = query_5_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
             tensor<int32, [4]> var_574 = const()[name = tensor<string, []>("op_574"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_9_cast_fp16 = reshape(shape = var_574, x = linear_22_cast_fp16)[name = tensor<string, []>("k_9_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63769856))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64818496))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63765760))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64814400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_23_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_2_self_attn_linear_v_weight_to_fp16_quantized, x = query_5_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
             tensor<int32, [4]> var_578 = const()[name = tensor<string, []>("op_578"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_5_cast_fp16 = reshape(shape = var_578, x = linear_23_cast_fp16)[name = tensor<string, []>("v_5_cast_fp16")];
             tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_2_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_2_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64820608)))];
-            tensor<fp16, [1, 126, 8, 128]> var_590_cast_fp16 = add(x = q_13_cast_fp16, y = model_layers_2_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_590_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_2_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_2_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64822720)))];
-            tensor<fp16, [1, 126, 8, 128]> var_592_cast_fp16 = add(x = q_13_cast_fp16, y = model_layers_2_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_592_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_2_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64816512))), scale = tensor<fp16, [8]>([0x1.784p-7, 0x1.cacp-8, 0x1.97p-8, 0x1.00cp-7, 0x1.8b4p-7, 0x1.9c4p-7, 0x1.3d4p-8, 0x1.accp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_590_cast_fp16 = add(x = q_13_cast_fp16, y = model_layers_2_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_590_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_2_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64817600))), scale = tensor<fp16, [8]>([0x1.2bp-9, 0x1.da4p-11, 0x1.cecp-9, 0x1.9cp-9, 0x1.9bcp-9, 0x1.664p-8, 0x1.438p-9, 0x1.008p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_592_cast_fp16 = add(x = q_13_cast_fp16, y = model_layers_2_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_592_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_51_transpose_x_0 = const()[name = tensor<string, []>("x_51_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_51_transpose_y_0 = const()[name = tensor<string, []>("x_51_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_594_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_594_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64824832))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65081920))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_594_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_594_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64818688))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65075776))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_5_cast_fp16 = transpose(perm = q_with_bias_v_5_perm_0, x = var_592_cast_fp16)[name = tensor<string, []>("transpose_297")];
             tensor<fp16, [1, 8, 126, 251]> x_51_cast_fp16 = matmul(transpose_x = x_51_transpose_x_0, transpose_y = x_51_transpose_y_0, x = q_with_bias_v_5_cast_fp16, y = op_594_to_fp16_quantized)[name = tensor<string, []>("x_51_cast_fp16")];
             tensor<int32, [8]> x_53_pad_0 = const()[name = tensor<string, []>("x_53_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -458,12 +458,12 @@ program(1.0)
             tensor<int32, [3]> var_627 = const()[name = tensor<string, []>("op_627"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_626_cast_fp16 = transpose(perm = var_626_perm_0, x = x_57_cast_fp16)[name = tensor<string, []>("transpose_293")];
             tensor<fp16, [1, 126, 1024]> input_139_cast_fp16 = reshape(shape = var_627, x = var_626_cast_fp16)[name = tensor<string, []>("input_139_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65082496))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66131136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65076352))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66124992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_25_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_2_self_attn_linear_out_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_143_cast_fp16 = add(x = input_135_cast_fp16, y = linear_25_cast_fp16)[name = tensor<string, []>("input_143_cast_fp16")];
             tensor<int32, [1]> x_61_axes_0 = const()[name = tensor<string, []>("x_61_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66133248)))];
-            tensor<fp16, [1024]> model_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66135360)))];
+            tensor<fp16, [1024]> model_layers_2_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66127104)))];
+            tensor<fp16, [1024]> model_layers_2_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66129216)))];
             tensor<fp16, [1, 126, 1024]> x_61_cast_fp16 = layer_norm(axes = x_61_axes_0, beta = model_layers_2_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_2_norm_conv_weight_to_fp16, x = input_143_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")];
             tensor<int32, [3]> input_145_perm_0 = const()[name = tensor<string, []>("input_145_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_147_pad_type_0 = const()[name = tensor<string, []>("input_147_pad_type_0"), val = tensor<string, []>("valid")];
@@ -471,7 +471,7 @@ program(1.0)
             tensor<int32, [2]> input_147_pad_0 = const()[name = tensor<string, []>("input_147_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_147_dilations_0 = const()[name = tensor<string, []>("input_147_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_147_groups_0 = const()[name = tensor<string, []>("input_147_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66137472))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68234688))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(66131328))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68228544))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_145_cast_fp16 = transpose(perm = input_145_perm_0, x = x_61_cast_fp16)[name = tensor<string, []>("transpose_292")];
             tensor<fp16, [1, 2048, 126]> input_147_cast_fp16 = conv(dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = model_layers_2_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = tensor<string, []>("input_147_cast_fp16")];
             tensor<int32, []> x_63_split_num_splits_0 = const()[name = tensor<string, []>("x_63_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -489,8 +489,8 @@ program(1.0)
             tensor<int32, [1]> input_153_strides_0 = const()[name = tensor<string, []>("input_153_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_153_pad_0 = const()[name = tensor<string, []>("input_153_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_153_dilations_0 = const()[name = tensor<string, []>("input_153_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_252_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_252_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68238848))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68248128))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_253_to_fp16 = const()[name = tensor<string, []>("const_253_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68250240)))];
+            tensor<fp16, [1024, 1, 9]> const_252_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_252_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68232704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68241984))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_253_to_fp16 = const()[name = tensor<string, []>("const_253_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68244096)))];
             tensor<fp16, [1, 1024, 126]> input_155_cast_fp16 = conv(bias = const_253_to_fp16, dilations = input_153_dilations_0, groups = input_153_groups_0, pad = input_153_pad_0, pad_type = input_153_pad_type_0, strides = input_153_strides_0, weight = const_252_to_fp16_quantized, x = input_151_cast_fp16)[name = tensor<string, []>("input_155_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_157_cast_fp16 = silu(x = input_155_cast_fp16)[name = tensor<string, []>("input_157_cast_fp16")];
             tensor<string, []> x_65_pad_type_0 = const()[name = tensor<string, []>("x_65_pad_type_0"), val = tensor<string, []>("valid")];
@@ -498,64 +498,64 @@ program(1.0)
             tensor<int32, [2]> x_65_pad_0 = const()[name = tensor<string, []>("x_65_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_65_dilations_0 = const()[name = tensor<string, []>("x_65_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_65_groups_0 = const()[name = tensor<string, []>("x_65_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68252352))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69300992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(68246208))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69294848))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_65_cast_fp16 = conv(dilations = x_65_dilations_0, groups = x_65_groups_0, pad = x_65_pad_0, pad_type = x_65_pad_type_0, strides = x_65_strides_0, weight = model_layers_2_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
             tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_159_cast_fp16 = transpose(perm = input_159_perm_0, x = x_65_cast_fp16)[name = tensor<string, []>("transpose_291")];
             tensor<fp16, [1, 126, 1024]> input_161_cast_fp16 = add(x = input_143_cast_fp16, y = input_159_cast_fp16)[name = tensor<string, []>("input_161_cast_fp16")];
             tensor<int32, [1]> input_163_axes_0 = const()[name = tensor<string, []>("input_163_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69303104)))];
-            tensor<fp16, [1024]> model_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69305216)))];
+            tensor<fp16, [1024]> model_layers_2_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69296960)))];
+            tensor<fp16, [1024]> model_layers_2_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69299072)))];
             tensor<fp16, [1, 126, 1024]> input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = model_layers_2_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_2_norm_feed_forward2_weight_to_fp16, x = input_161_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_2_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69307328))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73501696))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_2_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69301184))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73495552))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_feed_forward2_linear1_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_167_cast_fp16 = silu(x = linear_26_cast_fp16)[name = tensor<string, []>("input_167_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_2_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73509952))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77704320))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_2_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_2_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73503808))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77698176))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_27_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_2_feed_forward2_linear2_weight_to_fp16_quantized, x = input_167_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
             tensor<fp16, []> var_687_to_fp16 = const()[name = tensor<string, []>("op_687_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_688_cast_fp16 = mul(x = linear_27_cast_fp16, y = var_687_to_fp16)[name = tensor<string, []>("op_688_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_173_cast_fp16 = add(x = input_161_cast_fp16, y = var_688_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")];
             tensor<int32, [1]> input_175_axes_0 = const()[name = tensor<string, []>("input_175_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77706432)))];
-            tensor<fp16, [1024]> model_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77708544)))];
+            tensor<fp16, [1024]> model_layers_2_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77700288)))];
+            tensor<fp16, [1024]> model_layers_2_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_2_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77702400)))];
             tensor<fp16, [1, 126, 1024]> input_175_cast_fp16 = layer_norm(axes = input_175_axes_0, beta = model_layers_2_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_2_norm_out_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("input_175_cast_fp16")];
             tensor<int32, [1]> input_177_axes_0 = const()[name = tensor<string, []>("input_177_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77710656)))];
-            tensor<fp16, [1024]> model_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77712768)))];
+            tensor<fp16, [1024]> model_layers_3_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77704512)))];
+            tensor<fp16, [1024]> model_layers_3_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77706624)))];
             tensor<fp16, [1, 126, 1024]> input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = model_layers_3_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_3_norm_feed_forward1_weight_to_fp16, x = input_175_cast_fp16)[name = tensor<string, []>("input_177_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_3_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77714880))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81909248))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_3_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77708736))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81903104))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_28_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_feed_forward1_linear1_weight_to_fp16_quantized, x = input_177_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_181_cast_fp16 = silu(x = linear_28_cast_fp16)[name = tensor<string, []>("input_181_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_3_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81917504))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86111872))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_3_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81911360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86105728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_29_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_3_feed_forward1_linear2_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
             tensor<fp16, []> var_716_to_fp16 = const()[name = tensor<string, []>("op_716_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_717_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_716_to_fp16)[name = tensor<string, []>("op_717_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_187_cast_fp16 = add(x = input_175_cast_fp16, y = var_717_cast_fp16)[name = tensor<string, []>("input_187_cast_fp16")];
             tensor<int32, [1]> query_7_axes_0 = const()[name = tensor<string, []>("query_7_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86113984)))];
-            tensor<fp16, [1024]> model_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86116096)))];
+            tensor<fp16, [1024]> model_layers_3_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86107840)))];
+            tensor<fp16, [1024]> model_layers_3_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86109952)))];
             tensor<fp16, [1, 126, 1024]> query_7_cast_fp16 = layer_norm(axes = query_7_axes_0, beta = model_layers_3_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_3_norm_self_att_weight_to_fp16, x = input_187_cast_fp16)[name = tensor<string, []>("query_7_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86118208))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87166848))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86112064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87160704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_3_self_attn_linear_q_weight_to_fp16_quantized, x = query_7_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
             tensor<int32, [4]> var_733 = const()[name = tensor<string, []>("op_733"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_19_cast_fp16 = reshape(shape = var_733, x = linear_30_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87168960))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88217600))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87162816))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88211456))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_31_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_3_self_attn_linear_k_weight_to_fp16_quantized, x = query_7_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
             tensor<int32, [4]> var_737 = const()[name = tensor<string, []>("op_737"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_13_cast_fp16 = reshape(shape = var_737, x = linear_31_cast_fp16)[name = tensor<string, []>("k_13_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88219712))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89268352))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88213568))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89262208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_32_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_3_self_attn_linear_v_weight_to_fp16_quantized, x = query_7_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
             tensor<int32, [4]> var_741 = const()[name = tensor<string, []>("op_741"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_7_cast_fp16 = reshape(shape = var_741, x = linear_32_cast_fp16)[name = tensor<string, []>("v_7_cast_fp16")];
             tensor<int32, [4]> value_7_perm_0 = const()[name = tensor<string, []>("value_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_3_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_3_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89270464)))];
-            tensor<fp16, [1, 126, 8, 128]> var_753_cast_fp16 = add(x = q_19_cast_fp16, y = model_layers_3_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_753_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_3_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_3_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89272576)))];
-            tensor<fp16, [1, 126, 8, 128]> var_755_cast_fp16 = add(x = q_19_cast_fp16, y = model_layers_3_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_755_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_3_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89264320))), scale = tensor<fp16, [8]>([0x1.cf4p-8, 0x1.144p-7, 0x1.e6p-8, 0x1.d4p-9, 0x1.948p-8, 0x1.f5cp-8, 0x1.048p-7, 0x1.9dcp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_753_cast_fp16 = add(x = q_19_cast_fp16, y = model_layers_3_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_753_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_3_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89265408))), scale = tensor<fp16, [8]>([0x1.1dcp-8, 0x1.3ep-9, 0x1.f88p-9, 0x1.45p-8, 0x1.5f4p-8, 0x1.388p-8, 0x1.e14p-8, 0x1.42p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_755_cast_fp16 = add(x = q_19_cast_fp16, y = model_layers_3_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_755_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_73_transpose_x_0 = const()[name = tensor<string, []>("x_73_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_73_transpose_y_0 = const()[name = tensor<string, []>("x_73_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_757_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_757_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89274688))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89531776))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_757_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_757_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89266496))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89523584))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_7_cast_fp16 = transpose(perm = q_with_bias_v_7_perm_0, x = var_755_cast_fp16)[name = tensor<string, []>("transpose_290")];
             tensor<fp16, [1, 8, 126, 251]> x_73_cast_fp16 = matmul(transpose_x = x_73_transpose_x_0, transpose_y = x_73_transpose_y_0, x = q_with_bias_v_7_cast_fp16, y = op_757_to_fp16_quantized)[name = tensor<string, []>("x_73_cast_fp16")];
             tensor<int32, [8]> x_75_pad_0 = const()[name = tensor<string, []>("x_75_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -595,12 +595,12 @@ program(1.0)
             tensor<int32, [3]> var_790 = const()[name = tensor<string, []>("op_790"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_789_cast_fp16 = transpose(perm = var_789_perm_0, x = x_79_cast_fp16)[name = tensor<string, []>("transpose_286")];
             tensor<fp16, [1, 126, 1024]> input_191_cast_fp16 = reshape(shape = var_790, x = var_789_cast_fp16)[name = tensor<string, []>("input_191_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89532352))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90580992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89524160))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90572800))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_3_self_attn_linear_out_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_195_cast_fp16 = add(x = input_187_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")];
             tensor<int32, [1]> x_83_axes_0 = const()[name = tensor<string, []>("x_83_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90583104)))];
-            tensor<fp16, [1024]> model_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90585216)))];
+            tensor<fp16, [1024]> model_layers_3_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90574912)))];
+            tensor<fp16, [1024]> model_layers_3_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90577024)))];
             tensor<fp16, [1, 126, 1024]> x_83_cast_fp16 = layer_norm(axes = x_83_axes_0, beta = model_layers_3_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_3_norm_conv_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("x_83_cast_fp16")];
             tensor<int32, [3]> input_197_perm_0 = const()[name = tensor<string, []>("input_197_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_199_pad_type_0 = const()[name = tensor<string, []>("input_199_pad_type_0"), val = tensor<string, []>("valid")];
@@ -608,7 +608,7 @@ program(1.0)
             tensor<int32, [2]> input_199_pad_0 = const()[name = tensor<string, []>("input_199_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_199_dilations_0 = const()[name = tensor<string, []>("input_199_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_199_groups_0 = const()[name = tensor<string, []>("input_199_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90587328))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92684544))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90579136))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92676352))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_197_cast_fp16 = transpose(perm = input_197_perm_0, x = x_83_cast_fp16)[name = tensor<string, []>("transpose_285")];
             tensor<fp16, [1, 2048, 126]> input_199_cast_fp16 = conv(dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = model_layers_3_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_197_cast_fp16)[name = tensor<string, []>("input_199_cast_fp16")];
             tensor<int32, []> x_85_split_num_splits_0 = const()[name = tensor<string, []>("x_85_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -626,8 +626,8 @@ program(1.0)
             tensor<int32, [1]> input_205_strides_0 = const()[name = tensor<string, []>("input_205_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_205_pad_0 = const()[name = tensor<string, []>("input_205_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_205_dilations_0 = const()[name = tensor<string, []>("input_205_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_254_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_254_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92688704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92697984))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_255_to_fp16 = const()[name = tensor<string, []>("const_255_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92700096)))];
+            tensor<fp16, [1024, 1, 9]> const_254_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_254_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92680512))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92689792))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_255_to_fp16 = const()[name = tensor<string, []>("const_255_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92691904)))];
             tensor<fp16, [1, 1024, 126]> input_207_cast_fp16 = conv(bias = const_255_to_fp16, dilations = input_205_dilations_0, groups = input_205_groups_0, pad = input_205_pad_0, pad_type = input_205_pad_type_0, strides = input_205_strides_0, weight = const_254_to_fp16_quantized, x = input_203_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_209_cast_fp16 = silu(x = input_207_cast_fp16)[name = tensor<string, []>("input_209_cast_fp16")];
             tensor<string, []> x_87_pad_type_0 = const()[name = tensor<string, []>("x_87_pad_type_0"), val = tensor<string, []>("valid")];
@@ -635,64 +635,64 @@ program(1.0)
             tensor<int32, [2]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_87_dilations_0 = const()[name = tensor<string, []>("x_87_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_87_groups_0 = const()[name = tensor<string, []>("x_87_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92702208))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93750848))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92694016))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93742656))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_87_cast_fp16 = conv(dilations = x_87_dilations_0, groups = x_87_groups_0, pad = x_87_pad_0, pad_type = x_87_pad_type_0, strides = x_87_strides_0, weight = model_layers_3_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
             tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_211_cast_fp16 = transpose(perm = input_211_perm_0, x = x_87_cast_fp16)[name = tensor<string, []>("transpose_284")];
             tensor<fp16, [1, 126, 1024]> input_213_cast_fp16 = add(x = input_195_cast_fp16, y = input_211_cast_fp16)[name = tensor<string, []>("input_213_cast_fp16")];
             tensor<int32, [1]> input_215_axes_0 = const()[name = tensor<string, []>("input_215_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93752960)))];
-            tensor<fp16, [1024]> model_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93755072)))];
+            tensor<fp16, [1024]> model_layers_3_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93744768)))];
+            tensor<fp16, [1024]> model_layers_3_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93746880)))];
             tensor<fp16, [1, 126, 1024]> input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = model_layers_3_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_3_norm_feed_forward2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor<string, []>("input_215_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_3_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93757184))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97951552))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_3_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93748992))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97943360))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_feed_forward2_linear1_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_219_cast_fp16 = silu(x = linear_35_cast_fp16)[name = tensor<string, []>("input_219_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_3_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97959808))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102154176))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_3_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_3_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97951616))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102145984))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_36_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_3_feed_forward2_linear2_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
             tensor<fp16, []> var_850_to_fp16 = const()[name = tensor<string, []>("op_850_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_851_cast_fp16 = mul(x = linear_36_cast_fp16, y = var_850_to_fp16)[name = tensor<string, []>("op_851_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_225_cast_fp16 = add(x = input_213_cast_fp16, y = var_851_cast_fp16)[name = tensor<string, []>("input_225_cast_fp16")];
             tensor<int32, [1]> input_227_axes_0 = const()[name = tensor<string, []>("input_227_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102156288)))];
-            tensor<fp16, [1024]> model_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102158400)))];
+            tensor<fp16, [1024]> model_layers_3_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102148096)))];
+            tensor<fp16, [1024]> model_layers_3_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_3_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102150208)))];
             tensor<fp16, [1, 126, 1024]> input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = model_layers_3_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_3_norm_out_weight_to_fp16, x = input_225_cast_fp16)[name = tensor<string, []>("input_227_cast_fp16")];
             tensor<int32, [1]> input_229_axes_0 = const()[name = tensor<string, []>("input_229_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102160512)))];
-            tensor<fp16, [1024]> model_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102162624)))];
+            tensor<fp16, [1024]> model_layers_4_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102152320)))];
+            tensor<fp16, [1024]> model_layers_4_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102154432)))];
             tensor<fp16, [1, 126, 1024]> input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, beta = model_layers_4_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_4_norm_feed_forward1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_4_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102164736))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106359104))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_4_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102156544))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106350912))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_feed_forward1_linear1_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_233_cast_fp16 = silu(x = linear_37_cast_fp16)[name = tensor<string, []>("input_233_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_4_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106367360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110561728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_4_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(106359168))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110553536))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_4_feed_forward1_linear2_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
             tensor<fp16, []> var_879_to_fp16 = const()[name = tensor<string, []>("op_879_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_880_cast_fp16 = mul(x = linear_38_cast_fp16, y = var_879_to_fp16)[name = tensor<string, []>("op_880_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_239_cast_fp16 = add(x = input_227_cast_fp16, y = var_880_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")];
             tensor<int32, [1]> query_9_axes_0 = const()[name = tensor<string, []>("query_9_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110563840)))];
-            tensor<fp16, [1024]> model_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110565952)))];
+            tensor<fp16, [1024]> model_layers_4_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110555648)))];
+            tensor<fp16, [1024]> model_layers_4_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110557760)))];
             tensor<fp16, [1, 126, 1024]> query_9_cast_fp16 = layer_norm(axes = query_9_axes_0, beta = model_layers_4_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_4_norm_self_att_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110568064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111616704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110559872))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111608512))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_39_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_4_self_attn_linear_q_weight_to_fp16_quantized, x = query_9_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
             tensor<int32, [4]> var_896 = const()[name = tensor<string, []>("op_896"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_25_cast_fp16 = reshape(shape = var_896, x = linear_39_cast_fp16)[name = tensor<string, []>("q_25_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111618816))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112667456))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111610624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112659264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_40_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_4_self_attn_linear_k_weight_to_fp16_quantized, x = query_9_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
             tensor<int32, [4]> var_900 = const()[name = tensor<string, []>("op_900"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_17_cast_fp16 = reshape(shape = var_900, x = linear_40_cast_fp16)[name = tensor<string, []>("k_17_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112669568))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113718208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112661376))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113710016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_41_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_4_self_attn_linear_v_weight_to_fp16_quantized, x = query_9_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
             tensor<int32, [4]> var_904 = const()[name = tensor<string, []>("op_904"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_9_cast_fp16 = reshape(shape = var_904, x = linear_41_cast_fp16)[name = tensor<string, []>("v_9_cast_fp16")];
             tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<string, []>("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_4_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_4_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113720320)))];
-            tensor<fp16, [1, 126, 8, 128]> var_916_cast_fp16 = add(x = q_25_cast_fp16, y = model_layers_4_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_916_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_4_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_4_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113722432)))];
-            tensor<fp16, [1, 126, 8, 128]> var_918_cast_fp16 = add(x = q_25_cast_fp16, y = model_layers_4_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_918_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_4_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113712128))), scale = tensor<fp16, [8]>([0x1.084p-7, 0x1.948p-8, 0x1.0dcp-7, 0x1.3fp-7, 0x1.fep-7, 0x1.2fp-7, 0x1.a84p-8, 0x1.748p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_916_cast_fp16 = add(x = q_25_cast_fp16, y = model_layers_4_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_916_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_4_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113713216))), scale = tensor<fp16, [8]>([0x1.3dp-8, 0x1.554p-9, 0x1.13p-7, 0x1.1b4p-8, 0x1.a68p-11, 0x1.174p-9, 0x1.2dcp-8, 0x1.6f8p-10]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_918_cast_fp16 = add(x = q_25_cast_fp16, y = model_layers_4_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_918_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_95_transpose_x_0 = const()[name = tensor<string, []>("x_95_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_95_transpose_y_0 = const()[name = tensor<string, []>("x_95_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_920_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_920_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113724544))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113981632))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_920_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_920_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113714304))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113971392))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_9_cast_fp16 = transpose(perm = q_with_bias_v_9_perm_0, x = var_918_cast_fp16)[name = tensor<string, []>("transpose_283")];
             tensor<fp16, [1, 8, 126, 251]> x_95_cast_fp16 = matmul(transpose_x = x_95_transpose_x_0, transpose_y = x_95_transpose_y_0, x = q_with_bias_v_9_cast_fp16, y = op_920_to_fp16_quantized)[name = tensor<string, []>("x_95_cast_fp16")];
             tensor<int32, [8]> x_97_pad_0 = const()[name = tensor<string, []>("x_97_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -732,12 +732,12 @@ program(1.0)
             tensor<int32, [3]> var_953 = const()[name = tensor<string, []>("op_953"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_952_cast_fp16 = transpose(perm = var_952_perm_0, x = x_101_cast_fp16)[name = tensor<string, []>("transpose_279")];
             tensor<fp16, [1, 126, 1024]> input_243_cast_fp16 = reshape(shape = var_953, x = var_952_cast_fp16)[name = tensor<string, []>("input_243_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113982208))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115030848))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113971968))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115020608))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_43_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_4_self_attn_linear_out_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_247_cast_fp16 = add(x = input_239_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("input_247_cast_fp16")];
             tensor<int32, [1]> x_105_axes_0 = const()[name = tensor<string, []>("x_105_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115032960)))];
-            tensor<fp16, [1024]> model_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115035072)))];
+            tensor<fp16, [1024]> model_layers_4_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115022720)))];
+            tensor<fp16, [1024]> model_layers_4_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115024832)))];
             tensor<fp16, [1, 126, 1024]> x_105_cast_fp16 = layer_norm(axes = x_105_axes_0, beta = model_layers_4_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_4_norm_conv_weight_to_fp16, x = input_247_cast_fp16)[name = tensor<string, []>("x_105_cast_fp16")];
             tensor<int32, [3]> input_249_perm_0 = const()[name = tensor<string, []>("input_249_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_251_pad_type_0 = const()[name = tensor<string, []>("input_251_pad_type_0"), val = tensor<string, []>("valid")];
@@ -745,7 +745,7 @@ program(1.0)
             tensor<int32, [2]> input_251_pad_0 = const()[name = tensor<string, []>("input_251_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_251_dilations_0 = const()[name = tensor<string, []>("input_251_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_251_groups_0 = const()[name = tensor<string, []>("input_251_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115037184))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117134400))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115026944))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117124160))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_249_cast_fp16 = transpose(perm = input_249_perm_0, x = x_105_cast_fp16)[name = tensor<string, []>("transpose_278")];
             tensor<fp16, [1, 2048, 126]> input_251_cast_fp16 = conv(dilations = input_251_dilations_0, groups = input_251_groups_0, pad = input_251_pad_0, pad_type = input_251_pad_type_0, strides = input_251_strides_0, weight = model_layers_4_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")];
             tensor<int32, []> x_107_split_num_splits_0 = const()[name = tensor<string, []>("x_107_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -763,8 +763,8 @@ program(1.0)
             tensor<int32, [1]> input_257_strides_0 = const()[name = tensor<string, []>("input_257_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_257_pad_0 = const()[name = tensor<string, []>("input_257_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_257_dilations_0 = const()[name = tensor<string, []>("input_257_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_256_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_256_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117138560))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117147840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_257_to_fp16 = const()[name = tensor<string, []>("const_257_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117149952)))];
+            tensor<fp16, [1024, 1, 9]> const_256_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_256_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117128320))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117137600))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_257_to_fp16 = const()[name = tensor<string, []>("const_257_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117139712)))];
             tensor<fp16, [1, 1024, 126]> input_259_cast_fp16 = conv(bias = const_257_to_fp16, dilations = input_257_dilations_0, groups = input_257_groups_0, pad = input_257_pad_0, pad_type = input_257_pad_type_0, strides = input_257_strides_0, weight = const_256_to_fp16_quantized, x = input_255_cast_fp16)[name = tensor<string, []>("input_259_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_261_cast_fp16 = silu(x = input_259_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")];
             tensor<string, []> x_109_pad_type_0 = const()[name = tensor<string, []>("x_109_pad_type_0"), val = tensor<string, []>("valid")];
@@ -772,64 +772,64 @@ program(1.0)
             tensor<int32, [2]> x_109_pad_0 = const()[name = tensor<string, []>("x_109_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_109_dilations_0 = const()[name = tensor<string, []>("x_109_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_109_groups_0 = const()[name = tensor<string, []>("x_109_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117152064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118200704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117141824))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118190464))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_109_cast_fp16 = conv(dilations = x_109_dilations_0, groups = x_109_groups_0, pad = x_109_pad_0, pad_type = x_109_pad_type_0, strides = x_109_strides_0, weight = model_layers_4_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = tensor<string, []>("x_109_cast_fp16")];
             tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_263_cast_fp16 = transpose(perm = input_263_perm_0, x = x_109_cast_fp16)[name = tensor<string, []>("transpose_277")];
             tensor<fp16, [1, 126, 1024]> input_265_cast_fp16 = add(x = input_247_cast_fp16, y = input_263_cast_fp16)[name = tensor<string, []>("input_265_cast_fp16")];
             tensor<int32, [1]> input_267_axes_0 = const()[name = tensor<string, []>("input_267_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118202816)))];
-            tensor<fp16, [1024]> model_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118204928)))];
+            tensor<fp16, [1024]> model_layers_4_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118192576)))];
+            tensor<fp16, [1024]> model_layers_4_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118194688)))];
             tensor<fp16, [1, 126, 1024]> input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = model_layers_4_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_4_norm_feed_forward2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor<string, []>("input_267_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_4_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118207040))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122401408))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_4_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(118196800))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122391168))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_feed_forward2_linear1_weight_to_fp16_quantized, x = input_267_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_271_cast_fp16 = silu(x = linear_44_cast_fp16)[name = tensor<string, []>("input_271_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_4_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122409664))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126604032))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_4_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_4_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122399424))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126593792))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_45_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_4_feed_forward2_linear2_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
             tensor<fp16, []> var_1013_to_fp16 = const()[name = tensor<string, []>("op_1013_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1014_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_1013_to_fp16)[name = tensor<string, []>("op_1014_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_277_cast_fp16 = add(x = input_265_cast_fp16, y = var_1014_cast_fp16)[name = tensor<string, []>("input_277_cast_fp16")];
             tensor<int32, [1]> input_279_axes_0 = const()[name = tensor<string, []>("input_279_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126606144)))];
-            tensor<fp16, [1024]> model_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126608256)))];
+            tensor<fp16, [1024]> model_layers_4_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126595904)))];
+            tensor<fp16, [1024]> model_layers_4_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_4_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126598016)))];
             tensor<fp16, [1, 126, 1024]> input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = model_layers_4_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_4_norm_out_weight_to_fp16, x = input_277_cast_fp16)[name = tensor<string, []>("input_279_cast_fp16")];
             tensor<int32, [1]> input_281_axes_0 = const()[name = tensor<string, []>("input_281_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126610368)))];
-            tensor<fp16, [1024]> model_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126612480)))];
+            tensor<fp16, [1024]> model_layers_5_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126600128)))];
+            tensor<fp16, [1024]> model_layers_5_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126602240)))];
             tensor<fp16, [1, 126, 1024]> input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = model_layers_5_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_5_norm_feed_forward1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor<string, []>("input_281_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_5_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126614592))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130808960))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_5_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126604352))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130798720))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_feed_forward1_linear1_weight_to_fp16_quantized, x = input_281_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_285_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<string, []>("input_285_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_5_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130817216))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135011584))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_5_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130806976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135001344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_47_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_5_feed_forward1_linear2_weight_to_fp16_quantized, x = input_285_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
             tensor<fp16, []> var_1042_to_fp16 = const()[name = tensor<string, []>("op_1042_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1043_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_1042_to_fp16)[name = tensor<string, []>("op_1043_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_291_cast_fp16 = add(x = input_279_cast_fp16, y = var_1043_cast_fp16)[name = tensor<string, []>("input_291_cast_fp16")];
             tensor<int32, [1]> query_11_axes_0 = const()[name = tensor<string, []>("query_11_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135013696)))];
-            tensor<fp16, [1024]> model_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135015808)))];
+            tensor<fp16, [1024]> model_layers_5_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135003456)))];
+            tensor<fp16, [1024]> model_layers_5_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135005568)))];
             tensor<fp16, [1, 126, 1024]> query_11_cast_fp16 = layer_norm(axes = query_11_axes_0, beta = model_layers_5_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_5_norm_self_att_weight_to_fp16, x = input_291_cast_fp16)[name = tensor<string, []>("query_11_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135017920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136066560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135007680))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136056320))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_48_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_5_self_attn_linear_q_weight_to_fp16_quantized, x = query_11_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
             tensor<int32, [4]> var_1059 = const()[name = tensor<string, []>("op_1059"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_31_cast_fp16 = reshape(shape = var_1059, x = linear_48_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136068672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137117312))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136058432))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137107072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_49_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_5_self_attn_linear_k_weight_to_fp16_quantized, x = query_11_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
             tensor<int32, [4]> var_1063 = const()[name = tensor<string, []>("op_1063"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_21_cast_fp16 = reshape(shape = var_1063, x = linear_49_cast_fp16)[name = tensor<string, []>("k_21_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137119424))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138168064))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137109184))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138157824))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_5_self_attn_linear_v_weight_to_fp16_quantized, x = query_11_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
             tensor<int32, [4]> var_1067 = const()[name = tensor<string, []>("op_1067"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_11_cast_fp16 = reshape(shape = var_1067, x = linear_50_cast_fp16)[name = tensor<string, []>("v_11_cast_fp16")];
             tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_5_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_5_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138170176)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1079_cast_fp16 = add(x = q_31_cast_fp16, y = model_layers_5_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1079_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_5_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_5_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138172288)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1081_cast_fp16 = add(x = q_31_cast_fp16, y = model_layers_5_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1081_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_5_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138159936))), scale = tensor<fp16, [8]>([0x1.118p-7, 0x1.8ap-8, 0x1.59p-7, 0x1.4dp-7, 0x1.d14p-8, 0x1.3c4p-7, 0x1.758p-7, 0x1.6cp-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1079_cast_fp16 = add(x = q_31_cast_fp16, y = model_layers_5_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_1079_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_5_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138161024))), scale = tensor<fp16, [8]>([0x1.8a4p-9, 0x1.dap-8, 0x1.228p-8, 0x1.3dcp-9, 0x1.f9cp-8, 0x1.1c8p-8, 0x1.e3p-8, 0x1.948p-9]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1081_cast_fp16 = add(x = q_31_cast_fp16, y = model_layers_5_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_1081_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_1083_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1083_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138174400))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138431488))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_1083_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1083_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138162112))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138419200))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_11_cast_fp16 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1081_cast_fp16)[name = tensor<string, []>("transpose_276")];
             tensor<fp16, [1, 8, 126, 251]> x_117_cast_fp16 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = q_with_bias_v_11_cast_fp16, y = op_1083_to_fp16_quantized)[name = tensor<string, []>("x_117_cast_fp16")];
             tensor<int32, [8]> x_119_pad_0 = const()[name = tensor<string, []>("x_119_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -869,12 +869,12 @@ program(1.0)
             tensor<int32, [3]> var_1116 = const()[name = tensor<string, []>("op_1116"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_1115_cast_fp16 = transpose(perm = var_1115_perm_0, x = x_123_cast_fp16)[name = tensor<string, []>("transpose_272")];
             tensor<fp16, [1, 126, 1024]> input_295_cast_fp16 = reshape(shape = var_1116, x = var_1115_cast_fp16)[name = tensor<string, []>("input_295_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138432064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139480704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138419776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139468416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_52_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_5_self_attn_linear_out_weight_to_fp16_quantized, x = input_295_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_299_cast_fp16 = add(x = input_291_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input_299_cast_fp16")];
             tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139482816)))];
-            tensor<fp16, [1024]> model_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139484928)))];
+            tensor<fp16, [1024]> model_layers_5_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139470528)))];
+            tensor<fp16, [1024]> model_layers_5_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139472640)))];
             tensor<fp16, [1, 126, 1024]> x_127_cast_fp16 = layer_norm(axes = x_127_axes_0, beta = model_layers_5_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_5_norm_conv_weight_to_fp16, x = input_299_cast_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
             tensor<int32, [3]> input_301_perm_0 = const()[name = tensor<string, []>("input_301_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_303_pad_type_0 = const()[name = tensor<string, []>("input_303_pad_type_0"), val = tensor<string, []>("valid")];
@@ -882,7 +882,7 @@ program(1.0)
             tensor<int32, [2]> input_303_pad_0 = const()[name = tensor<string, []>("input_303_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_303_dilations_0 = const()[name = tensor<string, []>("input_303_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_303_groups_0 = const()[name = tensor<string, []>("input_303_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139487040))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141584256))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139474752))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141571968))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_301_cast_fp16 = transpose(perm = input_301_perm_0, x = x_127_cast_fp16)[name = tensor<string, []>("transpose_271")];
             tensor<fp16, [1, 2048, 126]> input_303_cast_fp16 = conv(dilations = input_303_dilations_0, groups = input_303_groups_0, pad = input_303_pad_0, pad_type = input_303_pad_type_0, strides = input_303_strides_0, weight = model_layers_5_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_301_cast_fp16)[name = tensor<string, []>("input_303_cast_fp16")];
             tensor<int32, []> x_129_split_num_splits_0 = const()[name = tensor<string, []>("x_129_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -900,8 +900,8 @@ program(1.0)
             tensor<int32, [1]> input_309_strides_0 = const()[name = tensor<string, []>("input_309_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_309_pad_0 = const()[name = tensor<string, []>("input_309_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_309_dilations_0 = const()[name = tensor<string, []>("input_309_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_258_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_258_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141588416))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141597696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_259_to_fp16 = const()[name = tensor<string, []>("const_259_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141599808)))];
+            tensor<fp16, [1024, 1, 9]> const_258_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_258_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141576128))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141585408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_259_to_fp16 = const()[name = tensor<string, []>("const_259_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141587520)))];
             tensor<fp16, [1, 1024, 126]> input_311_cast_fp16 = conv(bias = const_259_to_fp16, dilations = input_309_dilations_0, groups = input_309_groups_0, pad = input_309_pad_0, pad_type = input_309_pad_type_0, strides = input_309_strides_0, weight = const_258_to_fp16_quantized, x = input_307_cast_fp16)[name = tensor<string, []>("input_311_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_313_cast_fp16 = silu(x = input_311_cast_fp16)[name = tensor<string, []>("input_313_cast_fp16")];
             tensor<string, []> x_131_pad_type_0 = const()[name = tensor<string, []>("x_131_pad_type_0"), val = tensor<string, []>("valid")];
@@ -909,64 +909,64 @@ program(1.0)
             tensor<int32, [2]> x_131_pad_0 = const()[name = tensor<string, []>("x_131_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_131_dilations_0 = const()[name = tensor<string, []>("x_131_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_131_groups_0 = const()[name = tensor<string, []>("x_131_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141601920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142650560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141589632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142638272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_131_cast_fp16 = conv(dilations = x_131_dilations_0, groups = x_131_groups_0, pad = x_131_pad_0, pad_type = x_131_pad_type_0, strides = x_131_strides_0, weight = model_layers_5_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_313_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")];
             tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_315_cast_fp16 = transpose(perm = input_315_perm_0, x = x_131_cast_fp16)[name = tensor<string, []>("transpose_270")];
             tensor<fp16, [1, 126, 1024]> input_317_cast_fp16 = add(x = input_299_cast_fp16, y = input_315_cast_fp16)[name = tensor<string, []>("input_317_cast_fp16")];
             tensor<int32, [1]> input_319_axes_0 = const()[name = tensor<string, []>("input_319_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142652672)))];
-            tensor<fp16, [1024]> model_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142654784)))];
+            tensor<fp16, [1024]> model_layers_5_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142640384)))];
+            tensor<fp16, [1024]> model_layers_5_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142642496)))];
             tensor<fp16, [1, 126, 1024]> input_319_cast_fp16 = layer_norm(axes = input_319_axes_0, beta = model_layers_5_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_5_norm_feed_forward2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor<string, []>("input_319_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_5_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142656896))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146851264))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_5_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(142644608))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146838976))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_feed_forward2_linear1_weight_to_fp16_quantized, x = input_319_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_323_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<string, []>("input_323_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_5_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146859520))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151053888))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_5_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_5_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146847232))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151041600))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_5_feed_forward2_linear2_weight_to_fp16_quantized, x = input_323_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
             tensor<fp16, []> var_1176_to_fp16 = const()[name = tensor<string, []>("op_1176_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1177_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_1176_to_fp16)[name = tensor<string, []>("op_1177_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_329_cast_fp16 = add(x = input_317_cast_fp16, y = var_1177_cast_fp16)[name = tensor<string, []>("input_329_cast_fp16")];
             tensor<int32, [1]> input_331_axes_0 = const()[name = tensor<string, []>("input_331_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151056000)))];
-            tensor<fp16, [1024]> model_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151058112)))];
+            tensor<fp16, [1024]> model_layers_5_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151043712)))];
+            tensor<fp16, [1024]> model_layers_5_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_5_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151045824)))];
             tensor<fp16, [1, 126, 1024]> input_331_cast_fp16 = layer_norm(axes = input_331_axes_0, beta = model_layers_5_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_5_norm_out_weight_to_fp16, x = input_329_cast_fp16)[name = tensor<string, []>("input_331_cast_fp16")];
             tensor<int32, [1]> input_333_axes_0 = const()[name = tensor<string, []>("input_333_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151060224)))];
-            tensor<fp16, [1024]> model_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151062336)))];
+            tensor<fp16, [1024]> model_layers_6_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151047936)))];
+            tensor<fp16, [1024]> model_layers_6_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151050048)))];
             tensor<fp16, [1, 126, 1024]> input_333_cast_fp16 = layer_norm(axes = input_333_axes_0, beta = model_layers_6_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_6_norm_feed_forward1_weight_to_fp16, x = input_331_cast_fp16)[name = tensor<string, []>("input_333_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_6_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151064448))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155258816))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_6_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151052160))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155246528))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_feed_forward1_linear1_weight_to_fp16_quantized, x = input_333_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_337_cast_fp16 = silu(x = linear_55_cast_fp16)[name = tensor<string, []>("input_337_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_6_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155267072))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159461440))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_6_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155254784))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159449152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_56_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_6_feed_forward1_linear2_weight_to_fp16_quantized, x = input_337_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
             tensor<fp16, []> var_1205_to_fp16 = const()[name = tensor<string, []>("op_1205_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1206_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_1205_to_fp16)[name = tensor<string, []>("op_1206_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_343_cast_fp16 = add(x = input_331_cast_fp16, y = var_1206_cast_fp16)[name = tensor<string, []>("input_343_cast_fp16")];
             tensor<int32, [1]> query_13_axes_0 = const()[name = tensor<string, []>("query_13_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159463552)))];
-            tensor<fp16, [1024]> model_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159465664)))];
+            tensor<fp16, [1024]> model_layers_6_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159451264)))];
+            tensor<fp16, [1024]> model_layers_6_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159453376)))];
             tensor<fp16, [1, 126, 1024]> query_13_cast_fp16 = layer_norm(axes = query_13_axes_0, beta = model_layers_6_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_6_norm_self_att_weight_to_fp16, x = input_343_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159467776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160516416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159455488))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160504128))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_57_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_6_self_attn_linear_q_weight_to_fp16_quantized, x = query_13_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
             tensor<int32, [4]> var_1222 = const()[name = tensor<string, []>("op_1222"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_37_cast_fp16 = reshape(shape = var_1222, x = linear_57_cast_fp16)[name = tensor<string, []>("q_37_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160518528))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161567168))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160506240))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161554880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_6_self_attn_linear_k_weight_to_fp16_quantized, x = query_13_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
             tensor<int32, [4]> var_1226 = const()[name = tensor<string, []>("op_1226"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_25_cast_fp16 = reshape(shape = var_1226, x = linear_58_cast_fp16)[name = tensor<string, []>("k_25_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161569280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162617920))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161556992))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162605632))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_59_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_6_self_attn_linear_v_weight_to_fp16_quantized, x = query_13_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
             tensor<int32, [4]> var_1230 = const()[name = tensor<string, []>("op_1230"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_13_cast_fp16 = reshape(shape = var_1230, x = linear_59_cast_fp16)[name = tensor<string, []>("v_13_cast_fp16")];
             tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<string, []>("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_6_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_6_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162620032)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1242_cast_fp16 = add(x = q_37_cast_fp16, y = model_layers_6_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1242_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_6_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_6_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162622144)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1244_cast_fp16 = add(x = q_37_cast_fp16, y = model_layers_6_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1244_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_6_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162607744))), scale = tensor<fp16, [8]>([0x1.a08p-9, 0x1.75p-7, 0x1.01cp-8, 0x1.cp-8, 0x1.26p-7, 0x1.07cp-7, 0x1.dep-8, 0x1.86p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1242_cast_fp16 = add(x = q_37_cast_fp16, y = model_layers_6_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_1242_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_6_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162608832))), scale = tensor<fp16, [8]>([0x1.adcp-10, 0x1.978p-9, 0x1.d5p-8, 0x1.238p-8, 0x1.72cp-8, 0x1.7ecp-9, 0x1.24cp-7, 0x1.34cp-9]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1244_cast_fp16 = add(x = q_37_cast_fp16, y = model_layers_6_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_1244_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_139_transpose_x_0 = const()[name = tensor<string, []>("x_139_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_139_transpose_y_0 = const()[name = tensor<string, []>("x_139_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_1246_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1246_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162624256))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162881344))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_1246_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1246_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162609920))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162867008))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_13_cast_fp16 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1244_cast_fp16)[name = tensor<string, []>("transpose_269")];
             tensor<fp16, [1, 8, 126, 251]> x_139_cast_fp16 = matmul(transpose_x = x_139_transpose_x_0, transpose_y = x_139_transpose_y_0, x = q_with_bias_v_13_cast_fp16, y = op_1246_to_fp16_quantized)[name = tensor<string, []>("x_139_cast_fp16")];
             tensor<int32, [8]> x_141_pad_0 = const()[name = tensor<string, []>("x_141_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1006,12 +1006,12 @@ program(1.0)
             tensor<int32, [3]> var_1279 = const()[name = tensor<string, []>("op_1279"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_1278_cast_fp16 = transpose(perm = var_1278_perm_0, x = x_145_cast_fp16)[name = tensor<string, []>("transpose_265")];
             tensor<fp16, [1, 126, 1024]> input_347_cast_fp16 = reshape(shape = var_1279, x = var_1278_cast_fp16)[name = tensor<string, []>("input_347_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162881920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163930560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162867584))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163916224))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_61_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_6_self_attn_linear_out_weight_to_fp16_quantized, x = input_347_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_351_cast_fp16 = add(x = input_343_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("input_351_cast_fp16")];
             tensor<int32, [1]> x_149_axes_0 = const()[name = tensor<string, []>("x_149_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163932672)))];
-            tensor<fp16, [1024]> model_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163934784)))];
+            tensor<fp16, [1024]> model_layers_6_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163918336)))];
+            tensor<fp16, [1024]> model_layers_6_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163920448)))];
             tensor<fp16, [1, 126, 1024]> x_149_cast_fp16 = layer_norm(axes = x_149_axes_0, beta = model_layers_6_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_6_norm_conv_weight_to_fp16, x = input_351_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
             tensor<int32, [3]> input_353_perm_0 = const()[name = tensor<string, []>("input_353_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_355_pad_type_0 = const()[name = tensor<string, []>("input_355_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1019,7 +1019,7 @@ program(1.0)
             tensor<int32, [2]> input_355_pad_0 = const()[name = tensor<string, []>("input_355_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_355_dilations_0 = const()[name = tensor<string, []>("input_355_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_355_groups_0 = const()[name = tensor<string, []>("input_355_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163936896))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166034112))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163922560))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166019776))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_353_cast_fp16 = transpose(perm = input_353_perm_0, x = x_149_cast_fp16)[name = tensor<string, []>("transpose_264")];
             tensor<fp16, [1, 2048, 126]> input_355_cast_fp16 = conv(dilations = input_355_dilations_0, groups = input_355_groups_0, pad = input_355_pad_0, pad_type = input_355_pad_type_0, strides = input_355_strides_0, weight = model_layers_6_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_353_cast_fp16)[name = tensor<string, []>("input_355_cast_fp16")];
             tensor<int32, []> x_151_split_num_splits_0 = const()[name = tensor<string, []>("x_151_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1037,8 +1037,8 @@ program(1.0)
             tensor<int32, [1]> input_361_strides_0 = const()[name = tensor<string, []>("input_361_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_361_pad_0 = const()[name = tensor<string, []>("input_361_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_361_dilations_0 = const()[name = tensor<string, []>("input_361_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_260_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_260_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166038272))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166047552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_261_to_fp16 = const()[name = tensor<string, []>("const_261_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166049664)))];
+            tensor<fp16, [1024, 1, 9]> const_260_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_260_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166023936))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166033216))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_261_to_fp16 = const()[name = tensor<string, []>("const_261_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166035328)))];
             tensor<fp16, [1, 1024, 126]> input_363_cast_fp16 = conv(bias = const_261_to_fp16, dilations = input_361_dilations_0, groups = input_361_groups_0, pad = input_361_pad_0, pad_type = input_361_pad_type_0, strides = input_361_strides_0, weight = const_260_to_fp16_quantized, x = input_359_cast_fp16)[name = tensor<string, []>("input_363_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_365_cast_fp16 = silu(x = input_363_cast_fp16)[name = tensor<string, []>("input_365_cast_fp16")];
             tensor<string, []> x_153_pad_type_0 = const()[name = tensor<string, []>("x_153_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1046,64 +1046,64 @@ program(1.0)
             tensor<int32, [2]> x_153_pad_0 = const()[name = tensor<string, []>("x_153_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_153_dilations_0 = const()[name = tensor<string, []>("x_153_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_153_groups_0 = const()[name = tensor<string, []>("x_153_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166051776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167100416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166037440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167086080))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_153_cast_fp16 = conv(dilations = x_153_dilations_0, groups = x_153_groups_0, pad = x_153_pad_0, pad_type = x_153_pad_type_0, strides = x_153_strides_0, weight = model_layers_6_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_365_cast_fp16)[name = tensor<string, []>("x_153_cast_fp16")];
             tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_367_cast_fp16 = transpose(perm = input_367_perm_0, x = x_153_cast_fp16)[name = tensor<string, []>("transpose_263")];
             tensor<fp16, [1, 126, 1024]> input_369_cast_fp16 = add(x = input_351_cast_fp16, y = input_367_cast_fp16)[name = tensor<string, []>("input_369_cast_fp16")];
             tensor<int32, [1]> input_371_axes_0 = const()[name = tensor<string, []>("input_371_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167102528)))];
-            tensor<fp16, [1024]> model_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167104640)))];
+            tensor<fp16, [1024]> model_layers_6_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167088192)))];
+            tensor<fp16, [1024]> model_layers_6_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167090304)))];
             tensor<fp16, [1, 126, 1024]> input_371_cast_fp16 = layer_norm(axes = input_371_axes_0, beta = model_layers_6_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_6_norm_feed_forward2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor<string, []>("input_371_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_6_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167106752))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171301120))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_6_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167092416))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171286784))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_feed_forward2_linear1_weight_to_fp16_quantized, x = input_371_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_375_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor<string, []>("input_375_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_6_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171309376))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175503744))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_6_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_6_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171295040))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175489408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_63_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_6_feed_forward2_linear2_weight_to_fp16_quantized, x = input_375_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
             tensor<fp16, []> var_1339_to_fp16 = const()[name = tensor<string, []>("op_1339_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1340_cast_fp16 = mul(x = linear_63_cast_fp16, y = var_1339_to_fp16)[name = tensor<string, []>("op_1340_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_381_cast_fp16 = add(x = input_369_cast_fp16, y = var_1340_cast_fp16)[name = tensor<string, []>("input_381_cast_fp16")];
             tensor<int32, [1]> input_383_axes_0 = const()[name = tensor<string, []>("input_383_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175505856)))];
-            tensor<fp16, [1024]> model_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175507968)))];
+            tensor<fp16, [1024]> model_layers_6_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175491520)))];
+            tensor<fp16, [1024]> model_layers_6_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_6_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175493632)))];
             tensor<fp16, [1, 126, 1024]> input_383_cast_fp16 = layer_norm(axes = input_383_axes_0, beta = model_layers_6_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_6_norm_out_weight_to_fp16, x = input_381_cast_fp16)[name = tensor<string, []>("input_383_cast_fp16")];
             tensor<int32, [1]> input_385_axes_0 = const()[name = tensor<string, []>("input_385_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175510080)))];
-            tensor<fp16, [1024]> model_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175512192)))];
+            tensor<fp16, [1024]> model_layers_7_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175495744)))];
+            tensor<fp16, [1024]> model_layers_7_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175497856)))];
             tensor<fp16, [1, 126, 1024]> input_385_cast_fp16 = layer_norm(axes = input_385_axes_0, beta = model_layers_7_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_7_norm_feed_forward1_weight_to_fp16, x = input_383_cast_fp16)[name = tensor<string, []>("input_385_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_7_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175514304))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179708672))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_7_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(175499968))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179694336))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_feed_forward1_linear1_weight_to_fp16_quantized, x = input_385_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_389_cast_fp16 = silu(x = linear_64_cast_fp16)[name = tensor<string, []>("input_389_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_7_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179716928))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183911296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_7_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(179702592))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183896960))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_65_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_7_feed_forward1_linear2_weight_to_fp16_quantized, x = input_389_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
             tensor<fp16, []> var_1368_to_fp16 = const()[name = tensor<string, []>("op_1368_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1369_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_1368_to_fp16)[name = tensor<string, []>("op_1369_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_395_cast_fp16 = add(x = input_383_cast_fp16, y = var_1369_cast_fp16)[name = tensor<string, []>("input_395_cast_fp16")];
             tensor<int32, [1]> query_15_axes_0 = const()[name = tensor<string, []>("query_15_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183913408)))];
-            tensor<fp16, [1024]> model_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183915520)))];
+            tensor<fp16, [1024]> model_layers_7_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183899072)))];
+            tensor<fp16, [1024]> model_layers_7_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183901184)))];
             tensor<fp16, [1, 126, 1024]> query_15_cast_fp16 = layer_norm(axes = query_15_axes_0, beta = model_layers_7_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_7_norm_self_att_weight_to_fp16, x = input_395_cast_fp16)[name = tensor<string, []>("query_15_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183917632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184966272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183903296))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184951936))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_7_self_attn_linear_q_weight_to_fp16_quantized, x = query_15_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
             tensor<int32, [4]> var_1385 = const()[name = tensor<string, []>("op_1385"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_43_cast_fp16 = reshape(shape = var_1385, x = linear_66_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184968384))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186017024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184954048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186002688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_67_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_7_self_attn_linear_k_weight_to_fp16_quantized, x = query_15_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
             tensor<int32, [4]> var_1389 = const()[name = tensor<string, []>("op_1389"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_29_cast_fp16 = reshape(shape = var_1389, x = linear_67_cast_fp16)[name = tensor<string, []>("k_29_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186019136))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187067776))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186004800))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187053440))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_68_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_7_self_attn_linear_v_weight_to_fp16_quantized, x = query_15_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
             tensor<int32, [4]> var_1393 = const()[name = tensor<string, []>("op_1393"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_15_cast_fp16 = reshape(shape = var_1393, x = linear_68_cast_fp16)[name = tensor<string, []>("v_15_cast_fp16")];
             tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_7_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_7_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187069888)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1405_cast_fp16 = add(x = q_43_cast_fp16, y = model_layers_7_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1405_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_7_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_7_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187072000)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1407_cast_fp16 = add(x = q_43_cast_fp16, y = model_layers_7_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1407_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_7_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187055552))), scale = tensor<fp16, [8]>([0x1.b94p-8, 0x1.224p-7, 0x1.1b4p-8, 0x1.1dp-7, 0x1.8d4p-8, 0x1.bbp-8, 0x1.614p-8, 0x1.01cp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1405_cast_fp16 = add(x = q_43_cast_fp16, y = model_layers_7_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_1405_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_7_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187056640))), scale = tensor<fp16, [8]>([0x1.5acp-9, 0x1.9dcp-11, 0x1.a34p-8, 0x1.d98p-9, 0x1.edp-9, 0x1.904p-8, 0x1.1ccp-8, 0x1.358p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1407_cast_fp16 = add(x = q_43_cast_fp16, y = model_layers_7_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_1407_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_161_transpose_x_0 = const()[name = tensor<string, []>("x_161_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_161_transpose_y_0 = const()[name = tensor<string, []>("x_161_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_1409_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1409_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187074112))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187331200))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_1409_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1409_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187057728))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187314816))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_15_cast_fp16 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1407_cast_fp16)[name = tensor<string, []>("transpose_262")];
             tensor<fp16, [1, 8, 126, 251]> x_161_cast_fp16 = matmul(transpose_x = x_161_transpose_x_0, transpose_y = x_161_transpose_y_0, x = q_with_bias_v_15_cast_fp16, y = op_1409_to_fp16_quantized)[name = tensor<string, []>("x_161_cast_fp16")];
             tensor<int32, [8]> x_163_pad_0 = const()[name = tensor<string, []>("x_163_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1143,12 +1143,12 @@ program(1.0)
             tensor<int32, [3]> var_1442 = const()[name = tensor<string, []>("op_1442"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_1441_cast_fp16 = transpose(perm = var_1441_perm_0, x = x_167_cast_fp16)[name = tensor<string, []>("transpose_258")];
             tensor<fp16, [1, 126, 1024]> input_399_cast_fp16 = reshape(shape = var_1442, x = var_1441_cast_fp16)[name = tensor<string, []>("input_399_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187331776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188380416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187315392))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188364032))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_7_self_attn_linear_out_weight_to_fp16_quantized, x = input_399_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_403_cast_fp16 = add(x = input_395_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input_403_cast_fp16")];
             tensor<int32, [1]> x_171_axes_0 = const()[name = tensor<string, []>("x_171_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188382528)))];
-            tensor<fp16, [1024]> model_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188384640)))];
+            tensor<fp16, [1024]> model_layers_7_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188366144)))];
+            tensor<fp16, [1024]> model_layers_7_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188368256)))];
             tensor<fp16, [1, 126, 1024]> x_171_cast_fp16 = layer_norm(axes = x_171_axes_0, beta = model_layers_7_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_7_norm_conv_weight_to_fp16, x = input_403_cast_fp16)[name = tensor<string, []>("x_171_cast_fp16")];
             tensor<int32, [3]> input_405_perm_0 = const()[name = tensor<string, []>("input_405_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_407_pad_type_0 = const()[name = tensor<string, []>("input_407_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1156,7 +1156,7 @@ program(1.0)
             tensor<int32, [2]> input_407_pad_0 = const()[name = tensor<string, []>("input_407_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_407_dilations_0 = const()[name = tensor<string, []>("input_407_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_407_groups_0 = const()[name = tensor<string, []>("input_407_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188386752))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190483968))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(188370368))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190467584))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_405_cast_fp16 = transpose(perm = input_405_perm_0, x = x_171_cast_fp16)[name = tensor<string, []>("transpose_257")];
             tensor<fp16, [1, 2048, 126]> input_407_cast_fp16 = conv(dilations = input_407_dilations_0, groups = input_407_groups_0, pad = input_407_pad_0, pad_type = input_407_pad_type_0, strides = input_407_strides_0, weight = model_layers_7_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_405_cast_fp16)[name = tensor<string, []>("input_407_cast_fp16")];
             tensor<int32, []> x_173_split_num_splits_0 = const()[name = tensor<string, []>("x_173_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1174,8 +1174,8 @@ program(1.0)
             tensor<int32, [1]> input_413_strides_0 = const()[name = tensor<string, []>("input_413_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_413_pad_0 = const()[name = tensor<string, []>("input_413_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_413_dilations_0 = const()[name = tensor<string, []>("input_413_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_262_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_262_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190488128))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190497408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_263_to_fp16 = const()[name = tensor<string, []>("const_263_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190499520)))];
+            tensor<fp16, [1024, 1, 9]> const_262_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_262_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190471744))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190481024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_263_to_fp16 = const()[name = tensor<string, []>("const_263_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190483136)))];
             tensor<fp16, [1, 1024, 126]> input_415_cast_fp16 = conv(bias = const_263_to_fp16, dilations = input_413_dilations_0, groups = input_413_groups_0, pad = input_413_pad_0, pad_type = input_413_pad_type_0, strides = input_413_strides_0, weight = const_262_to_fp16_quantized, x = input_411_cast_fp16)[name = tensor<string, []>("input_415_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_417_cast_fp16 = silu(x = input_415_cast_fp16)[name = tensor<string, []>("input_417_cast_fp16")];
             tensor<string, []> x_175_pad_type_0 = const()[name = tensor<string, []>("x_175_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1183,64 +1183,64 @@ program(1.0)
             tensor<int32, [2]> x_175_pad_0 = const()[name = tensor<string, []>("x_175_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_175_dilations_0 = const()[name = tensor<string, []>("x_175_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_175_groups_0 = const()[name = tensor<string, []>("x_175_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190501632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191550272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190485248))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191533888))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_175_cast_fp16 = conv(dilations = x_175_dilations_0, groups = x_175_groups_0, pad = x_175_pad_0, pad_type = x_175_pad_type_0, strides = x_175_strides_0, weight = model_layers_7_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_417_cast_fp16)[name = tensor<string, []>("x_175_cast_fp16")];
             tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_419_cast_fp16 = transpose(perm = input_419_perm_0, x = x_175_cast_fp16)[name = tensor<string, []>("transpose_256")];
             tensor<fp16, [1, 126, 1024]> input_421_cast_fp16 = add(x = input_403_cast_fp16, y = input_419_cast_fp16)[name = tensor<string, []>("input_421_cast_fp16")];
             tensor<int32, [1]> input_423_axes_0 = const()[name = tensor<string, []>("input_423_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191552384)))];
-            tensor<fp16, [1024]> model_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191554496)))];
+            tensor<fp16, [1024]> model_layers_7_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191536000)))];
+            tensor<fp16, [1024]> model_layers_7_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191538112)))];
             tensor<fp16, [1, 126, 1024]> input_423_cast_fp16 = layer_norm(axes = input_423_axes_0, beta = model_layers_7_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_7_norm_feed_forward2_weight_to_fp16, x = input_421_cast_fp16)[name = tensor<string, []>("input_423_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_7_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191556608))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195750976))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_7_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(191540224))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195734592))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_feed_forward2_linear1_weight_to_fp16_quantized, x = input_423_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_427_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor<string, []>("input_427_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_7_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195759232))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199953600))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_7_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_7_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(195742848))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199937216))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_72_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_7_feed_forward2_linear2_weight_to_fp16_quantized, x = input_427_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
             tensor<fp16, []> var_1502_to_fp16 = const()[name = tensor<string, []>("op_1502_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1503_cast_fp16 = mul(x = linear_72_cast_fp16, y = var_1502_to_fp16)[name = tensor<string, []>("op_1503_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_433_cast_fp16 = add(x = input_421_cast_fp16, y = var_1503_cast_fp16)[name = tensor<string, []>("input_433_cast_fp16")];
             tensor<int32, [1]> input_435_axes_0 = const()[name = tensor<string, []>("input_435_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199955712)))];
-            tensor<fp16, [1024]> model_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199957824)))];
+            tensor<fp16, [1024]> model_layers_7_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199939328)))];
+            tensor<fp16, [1024]> model_layers_7_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_7_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199941440)))];
             tensor<fp16, [1, 126, 1024]> input_435_cast_fp16 = layer_norm(axes = input_435_axes_0, beta = model_layers_7_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_7_norm_out_weight_to_fp16, x = input_433_cast_fp16)[name = tensor<string, []>("input_435_cast_fp16")];
             tensor<int32, [1]> input_437_axes_0 = const()[name = tensor<string, []>("input_437_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199959936)))];
-            tensor<fp16, [1024]> model_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199962048)))];
+            tensor<fp16, [1024]> model_layers_8_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199943552)))];
+            tensor<fp16, [1024]> model_layers_8_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199945664)))];
             tensor<fp16, [1, 126, 1024]> input_437_cast_fp16 = layer_norm(axes = input_437_axes_0, beta = model_layers_8_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_8_norm_feed_forward1_weight_to_fp16, x = input_435_cast_fp16)[name = tensor<string, []>("input_437_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_8_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199964160))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204158528))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_8_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(199947776))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204142144))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_feed_forward1_linear1_weight_to_fp16_quantized, x = input_437_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_441_cast_fp16 = silu(x = linear_73_cast_fp16)[name = tensor<string, []>("input_441_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_8_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204166784))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208361152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_8_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(204150400))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208344768))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_8_feed_forward1_linear2_weight_to_fp16_quantized, x = input_441_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
             tensor<fp16, []> var_1531_to_fp16 = const()[name = tensor<string, []>("op_1531_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1532_cast_fp16 = mul(x = linear_74_cast_fp16, y = var_1531_to_fp16)[name = tensor<string, []>("op_1532_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_447_cast_fp16 = add(x = input_435_cast_fp16, y = var_1532_cast_fp16)[name = tensor<string, []>("input_447_cast_fp16")];
             tensor<int32, [1]> query_17_axes_0 = const()[name = tensor<string, []>("query_17_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208363264)))];
-            tensor<fp16, [1024]> model_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208365376)))];
+            tensor<fp16, [1024]> model_layers_8_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208346880)))];
+            tensor<fp16, [1024]> model_layers_8_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208348992)))];
             tensor<fp16, [1, 126, 1024]> query_17_cast_fp16 = layer_norm(axes = query_17_axes_0, beta = model_layers_8_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_8_norm_self_att_weight_to_fp16, x = input_447_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208367488))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209416128))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208351104))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209399744))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_75_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_8_self_attn_linear_q_weight_to_fp16_quantized, x = query_17_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
             tensor<int32, [4]> var_1548 = const()[name = tensor<string, []>("op_1548"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_49_cast_fp16 = reshape(shape = var_1548, x = linear_75_cast_fp16)[name = tensor<string, []>("q_49_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209418240))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210466880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209401856))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210450496))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_76_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_8_self_attn_linear_k_weight_to_fp16_quantized, x = query_17_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
             tensor<int32, [4]> var_1552 = const()[name = tensor<string, []>("op_1552"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_33_cast_fp16 = reshape(shape = var_1552, x = linear_76_cast_fp16)[name = tensor<string, []>("k_33_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210468992))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211517632))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210452608))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211501248))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_77_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_8_self_attn_linear_v_weight_to_fp16_quantized, x = query_17_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
             tensor<int32, [4]> var_1556 = const()[name = tensor<string, []>("op_1556"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_17_cast_fp16 = reshape(shape = var_1556, x = linear_77_cast_fp16)[name = tensor<string, []>("v_17_cast_fp16")];
             tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<string, []>("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_8_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_8_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211519744)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1568_cast_fp16 = add(x = q_49_cast_fp16, y = model_layers_8_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1568_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_8_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_8_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211521856)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1570_cast_fp16 = add(x = q_49_cast_fp16, y = model_layers_8_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1570_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_8_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211503360))), scale = tensor<fp16, [8]>([0x1.d1p-8, 0x1.848p-8, 0x1.8dcp-8, 0x1.bbcp-7, 0x1.4b4p-8, 0x1.0ccp-7, 0x1.234p-7, 0x1.34p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1568_cast_fp16 = add(x = q_49_cast_fp16, y = model_layers_8_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_1568_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_8_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211504448))), scale = tensor<fp16, [8]>([0x1.608p-8, 0x1.538p-8, 0x1.6b8p-8, 0x1.f94p-9, 0x1.21cp-7, 0x1.c98p-9, 0x1.d08p-8, 0x1.638p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1570_cast_fp16 = add(x = q_49_cast_fp16, y = model_layers_8_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_1570_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_183_transpose_x_0 = const()[name = tensor<string, []>("x_183_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_183_transpose_y_0 = const()[name = tensor<string, []>("x_183_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_1572_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1572_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211523968))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211781056))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_1572_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1572_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211505536))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211762624))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_17_cast_fp16 = transpose(perm = q_with_bias_v_17_perm_0, x = var_1570_cast_fp16)[name = tensor<string, []>("transpose_255")];
             tensor<fp16, [1, 8, 126, 251]> x_183_cast_fp16 = matmul(transpose_x = x_183_transpose_x_0, transpose_y = x_183_transpose_y_0, x = q_with_bias_v_17_cast_fp16, y = op_1572_to_fp16_quantized)[name = tensor<string, []>("x_183_cast_fp16")];
             tensor<int32, [8]> x_185_pad_0 = const()[name = tensor<string, []>("x_185_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1280,12 +1280,12 @@ program(1.0)
             tensor<int32, [3]> var_1605 = const()[name = tensor<string, []>("op_1605"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_1604_cast_fp16 = transpose(perm = var_1604_perm_0, x = x_189_cast_fp16)[name = tensor<string, []>("transpose_251")];
             tensor<fp16, [1, 126, 1024]> input_451_cast_fp16 = reshape(shape = var_1605, x = var_1604_cast_fp16)[name = tensor<string, []>("input_451_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211781632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212830272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211763200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212811840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_79_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_8_self_attn_linear_out_weight_to_fp16_quantized, x = input_451_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_455_cast_fp16 = add(x = input_447_cast_fp16, y = linear_79_cast_fp16)[name = tensor<string, []>("input_455_cast_fp16")];
             tensor<int32, [1]> x_193_axes_0 = const()[name = tensor<string, []>("x_193_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212832384)))];
-            tensor<fp16, [1024]> model_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212834496)))];
+            tensor<fp16, [1024]> model_layers_8_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212813952)))];
+            tensor<fp16, [1024]> model_layers_8_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212816064)))];
             tensor<fp16, [1, 126, 1024]> x_193_cast_fp16 = layer_norm(axes = x_193_axes_0, beta = model_layers_8_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_8_norm_conv_weight_to_fp16, x = input_455_cast_fp16)[name = tensor<string, []>("x_193_cast_fp16")];
             tensor<int32, [3]> input_457_perm_0 = const()[name = tensor<string, []>("input_457_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_459_pad_type_0 = const()[name = tensor<string, []>("input_459_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1293,7 +1293,7 @@ program(1.0)
             tensor<int32, [2]> input_459_pad_0 = const()[name = tensor<string, []>("input_459_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_459_dilations_0 = const()[name = tensor<string, []>("input_459_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_459_groups_0 = const()[name = tensor<string, []>("input_459_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212836608))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214933824))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(212818176))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214915392))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_457_cast_fp16 = transpose(perm = input_457_perm_0, x = x_193_cast_fp16)[name = tensor<string, []>("transpose_250")];
             tensor<fp16, [1, 2048, 126]> input_459_cast_fp16 = conv(dilations = input_459_dilations_0, groups = input_459_groups_0, pad = input_459_pad_0, pad_type = input_459_pad_type_0, strides = input_459_strides_0, weight = model_layers_8_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_457_cast_fp16)[name = tensor<string, []>("input_459_cast_fp16")];
             tensor<int32, []> x_195_split_num_splits_0 = const()[name = tensor<string, []>("x_195_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1311,8 +1311,8 @@ program(1.0)
             tensor<int32, [1]> input_465_strides_0 = const()[name = tensor<string, []>("input_465_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_465_pad_0 = const()[name = tensor<string, []>("input_465_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_465_dilations_0 = const()[name = tensor<string, []>("input_465_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_264_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_264_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214937984))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214947264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_265_to_fp16 = const()[name = tensor<string, []>("const_265_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214949376)))];
+            tensor<fp16, [1024, 1, 9]> const_264_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_264_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214919552))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214928832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_265_to_fp16 = const()[name = tensor<string, []>("const_265_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214930944)))];
             tensor<fp16, [1, 1024, 126]> input_467_cast_fp16 = conv(bias = const_265_to_fp16, dilations = input_465_dilations_0, groups = input_465_groups_0, pad = input_465_pad_0, pad_type = input_465_pad_type_0, strides = input_465_strides_0, weight = const_264_to_fp16_quantized, x = input_463_cast_fp16)[name = tensor<string, []>("input_467_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_469_cast_fp16 = silu(x = input_467_cast_fp16)[name = tensor<string, []>("input_469_cast_fp16")];
             tensor<string, []> x_197_pad_type_0 = const()[name = tensor<string, []>("x_197_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1320,64 +1320,64 @@ program(1.0)
             tensor<int32, [2]> x_197_pad_0 = const()[name = tensor<string, []>("x_197_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_197_dilations_0 = const()[name = tensor<string, []>("x_197_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_197_groups_0 = const()[name = tensor<string, []>("x_197_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214951488))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216000128))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214933056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215981696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_197_cast_fp16 = conv(dilations = x_197_dilations_0, groups = x_197_groups_0, pad = x_197_pad_0, pad_type = x_197_pad_type_0, strides = x_197_strides_0, weight = model_layers_8_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_469_cast_fp16)[name = tensor<string, []>("x_197_cast_fp16")];
             tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_471_cast_fp16 = transpose(perm = input_471_perm_0, x = x_197_cast_fp16)[name = tensor<string, []>("transpose_249")];
             tensor<fp16, [1, 126, 1024]> input_473_cast_fp16 = add(x = input_455_cast_fp16, y = input_471_cast_fp16)[name = tensor<string, []>("input_473_cast_fp16")];
             tensor<int32, [1]> input_475_axes_0 = const()[name = tensor<string, []>("input_475_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216002240)))];
-            tensor<fp16, [1024]> model_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216004352)))];
+            tensor<fp16, [1024]> model_layers_8_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215983808)))];
+            tensor<fp16, [1024]> model_layers_8_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215985920)))];
             tensor<fp16, [1, 126, 1024]> input_475_cast_fp16 = layer_norm(axes = input_475_axes_0, beta = model_layers_8_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_8_norm_feed_forward2_weight_to_fp16, x = input_473_cast_fp16)[name = tensor<string, []>("input_475_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_8_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(216006464))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220200832))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_8_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(215988032))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220182400))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_feed_forward2_linear1_weight_to_fp16_quantized, x = input_475_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_479_cast_fp16 = silu(x = linear_80_cast_fp16)[name = tensor<string, []>("input_479_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_8_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220209088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224403456))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_8_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_8_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(220190656))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224385024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_81_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_8_feed_forward2_linear2_weight_to_fp16_quantized, x = input_479_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
             tensor<fp16, []> var_1665_to_fp16 = const()[name = tensor<string, []>("op_1665_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1666_cast_fp16 = mul(x = linear_81_cast_fp16, y = var_1665_to_fp16)[name = tensor<string, []>("op_1666_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_485_cast_fp16 = add(x = input_473_cast_fp16, y = var_1666_cast_fp16)[name = tensor<string, []>("input_485_cast_fp16")];
             tensor<int32, [1]> input_487_axes_0 = const()[name = tensor<string, []>("input_487_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224405568)))];
-            tensor<fp16, [1024]> model_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224407680)))];
+            tensor<fp16, [1024]> model_layers_8_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224387136)))];
+            tensor<fp16, [1024]> model_layers_8_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_8_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224389248)))];
             tensor<fp16, [1, 126, 1024]> input_487_cast_fp16 = layer_norm(axes = input_487_axes_0, beta = model_layers_8_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_8_norm_out_weight_to_fp16, x = input_485_cast_fp16)[name = tensor<string, []>("input_487_cast_fp16")];
             tensor<int32, [1]> input_489_axes_0 = const()[name = tensor<string, []>("input_489_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224409792)))];
-            tensor<fp16, [1024]> model_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224411904)))];
+            tensor<fp16, [1024]> model_layers_9_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224391360)))];
+            tensor<fp16, [1024]> model_layers_9_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224393472)))];
             tensor<fp16, [1, 126, 1024]> input_489_cast_fp16 = layer_norm(axes = input_489_axes_0, beta = model_layers_9_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_9_norm_feed_forward1_weight_to_fp16, x = input_487_cast_fp16)[name = tensor<string, []>("input_489_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_9_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224414016))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228608384))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_9_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224395584))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228589952))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_82_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_feed_forward1_linear1_weight_to_fp16_quantized, x = input_489_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_493_cast_fp16 = silu(x = linear_82_cast_fp16)[name = tensor<string, []>("input_493_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_9_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228616640))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232811008))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_9_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228598208))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232792576))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_83_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_9_feed_forward1_linear2_weight_to_fp16_quantized, x = input_493_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
             tensor<fp16, []> var_1694_to_fp16 = const()[name = tensor<string, []>("op_1694_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1695_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_1694_to_fp16)[name = tensor<string, []>("op_1695_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_499_cast_fp16 = add(x = input_487_cast_fp16, y = var_1695_cast_fp16)[name = tensor<string, []>("input_499_cast_fp16")];
             tensor<int32, [1]> query_19_axes_0 = const()[name = tensor<string, []>("query_19_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232813120)))];
-            tensor<fp16, [1024]> model_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232815232)))];
+            tensor<fp16, [1024]> model_layers_9_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232794688)))];
+            tensor<fp16, [1024]> model_layers_9_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232796800)))];
             tensor<fp16, [1, 126, 1024]> query_19_cast_fp16 = layer_norm(axes = query_19_axes_0, beta = model_layers_9_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_9_norm_self_att_weight_to_fp16, x = input_499_cast_fp16)[name = tensor<string, []>("query_19_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232817344))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233865984))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232798912))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233847552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_84_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_9_self_attn_linear_q_weight_to_fp16_quantized, x = query_19_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
             tensor<int32, [4]> var_1711 = const()[name = tensor<string, []>("op_1711"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_55_cast_fp16 = reshape(shape = var_1711, x = linear_84_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233868096))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234916736))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233849664))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234898304))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_85_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_9_self_attn_linear_k_weight_to_fp16_quantized, x = query_19_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
             tensor<int32, [4]> var_1715 = const()[name = tensor<string, []>("op_1715"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_37_cast_fp16 = reshape(shape = var_1715, x = linear_85_cast_fp16)[name = tensor<string, []>("k_37_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234918848))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235967488))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234900416))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235949056))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_9_self_attn_linear_v_weight_to_fp16_quantized, x = query_19_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
             tensor<int32, [4]> var_1719 = const()[name = tensor<string, []>("op_1719"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_19_cast_fp16 = reshape(shape = var_1719, x = linear_86_cast_fp16)[name = tensor<string, []>("v_19_cast_fp16")];
             tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_9_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_9_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235969600)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1731_cast_fp16 = add(x = q_55_cast_fp16, y = model_layers_9_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1731_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_9_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_9_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235971712)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1733_cast_fp16 = add(x = q_55_cast_fp16, y = model_layers_9_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1733_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_9_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235951168))), scale = tensor<fp16, [8]>([0x1.15p-7, 0x1.4d8p-8, 0x1.eb8p-8, 0x1.3f4p-6, 0x1.bdcp-8, 0x1.e5cp-8, 0x1.968p-8, 0x1.034p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1731_cast_fp16 = add(x = q_55_cast_fp16, y = model_layers_9_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_1731_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_9_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235952256))), scale = tensor<fp16, [8]>([0x1.46p-8, 0x1.5c4p-7, 0x1.fa4p-8, 0x1.0a4p-8, 0x1.a3p-9, 0x1.364p-7, 0x1.694p-7, 0x1.f8p-10]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1733_cast_fp16 = add(x = q_55_cast_fp16, y = model_layers_9_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_1733_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_205_transpose_x_0 = const()[name = tensor<string, []>("x_205_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_205_transpose_y_0 = const()[name = tensor<string, []>("x_205_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_1735_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1735_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235973824))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236230912))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_1735_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1735_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235953344))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236210432))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_19_cast_fp16 = transpose(perm = q_with_bias_v_19_perm_0, x = var_1733_cast_fp16)[name = tensor<string, []>("transpose_248")];
             tensor<fp16, [1, 8, 126, 251]> x_205_cast_fp16 = matmul(transpose_x = x_205_transpose_x_0, transpose_y = x_205_transpose_y_0, x = q_with_bias_v_19_cast_fp16, y = op_1735_to_fp16_quantized)[name = tensor<string, []>("x_205_cast_fp16")];
             tensor<int32, [8]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1417,12 +1417,12 @@ program(1.0)
             tensor<int32, [3]> var_1768 = const()[name = tensor<string, []>("op_1768"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = x_211_cast_fp16)[name = tensor<string, []>("transpose_244")];
             tensor<fp16, [1, 126, 1024]> input_503_cast_fp16 = reshape(shape = var_1768, x = var_1767_cast_fp16)[name = tensor<string, []>("input_503_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236231488))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237280128))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(236211008))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237259648))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_88_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_9_self_attn_linear_out_weight_to_fp16_quantized, x = input_503_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_507_cast_fp16 = add(x = input_499_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input_507_cast_fp16")];
             tensor<int32, [1]> x_215_axes_0 = const()[name = tensor<string, []>("x_215_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237282240)))];
-            tensor<fp16, [1024]> model_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237284352)))];
+            tensor<fp16, [1024]> model_layers_9_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237261760)))];
+            tensor<fp16, [1024]> model_layers_9_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237263872)))];
             tensor<fp16, [1, 126, 1024]> x_215_cast_fp16 = layer_norm(axes = x_215_axes_0, beta = model_layers_9_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_9_norm_conv_weight_to_fp16, x = input_507_cast_fp16)[name = tensor<string, []>("x_215_cast_fp16")];
             tensor<int32, [3]> input_509_perm_0 = const()[name = tensor<string, []>("input_509_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_511_pad_type_0 = const()[name = tensor<string, []>("input_511_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1430,7 +1430,7 @@ program(1.0)
             tensor<int32, [2]> input_511_pad_0 = const()[name = tensor<string, []>("input_511_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_511_dilations_0 = const()[name = tensor<string, []>("input_511_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_511_groups_0 = const()[name = tensor<string, []>("input_511_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237286464))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239383680))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237265984))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239363200))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_509_cast_fp16 = transpose(perm = input_509_perm_0, x = x_215_cast_fp16)[name = tensor<string, []>("transpose_243")];
             tensor<fp16, [1, 2048, 126]> input_511_cast_fp16 = conv(dilations = input_511_dilations_0, groups = input_511_groups_0, pad = input_511_pad_0, pad_type = input_511_pad_type_0, strides = input_511_strides_0, weight = model_layers_9_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_509_cast_fp16)[name = tensor<string, []>("input_511_cast_fp16")];
             tensor<int32, []> x_217_split_num_splits_0 = const()[name = tensor<string, []>("x_217_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1448,8 +1448,8 @@ program(1.0)
             tensor<int32, [1]> input_517_strides_0 = const()[name = tensor<string, []>("input_517_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_517_pad_0 = const()[name = tensor<string, []>("input_517_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_517_dilations_0 = const()[name = tensor<string, []>("input_517_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_266_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_266_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239387840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239397120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_267_to_fp16 = const()[name = tensor<string, []>("const_267_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239399232)))];
+            tensor<fp16, [1024, 1, 9]> const_266_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_266_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239367360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239376640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_267_to_fp16 = const()[name = tensor<string, []>("const_267_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239378752)))];
             tensor<fp16, [1, 1024, 126]> input_519_cast_fp16 = conv(bias = const_267_to_fp16, dilations = input_517_dilations_0, groups = input_517_groups_0, pad = input_517_pad_0, pad_type = input_517_pad_type_0, strides = input_517_strides_0, weight = const_266_to_fp16_quantized, x = input_515_cast_fp16)[name = tensor<string, []>("input_519_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_521_cast_fp16 = silu(x = input_519_cast_fp16)[name = tensor<string, []>("input_521_cast_fp16")];
             tensor<string, []> x_219_pad_type_0 = const()[name = tensor<string, []>("x_219_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1457,64 +1457,64 @@ program(1.0)
             tensor<int32, [2]> x_219_pad_0 = const()[name = tensor<string, []>("x_219_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_219_dilations_0 = const()[name = tensor<string, []>("x_219_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_219_groups_0 = const()[name = tensor<string, []>("x_219_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239401344))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240449984))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(239380864))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240429504))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_219_cast_fp16 = conv(dilations = x_219_dilations_0, groups = x_219_groups_0, pad = x_219_pad_0, pad_type = x_219_pad_type_0, strides = x_219_strides_0, weight = model_layers_9_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_521_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
             tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_523_cast_fp16 = transpose(perm = input_523_perm_0, x = x_219_cast_fp16)[name = tensor<string, []>("transpose_242")];
             tensor<fp16, [1, 126, 1024]> input_525_cast_fp16 = add(x = input_507_cast_fp16, y = input_523_cast_fp16)[name = tensor<string, []>("input_525_cast_fp16")];
             tensor<int32, [1]> input_527_axes_0 = const()[name = tensor<string, []>("input_527_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240452096)))];
-            tensor<fp16, [1024]> model_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240454208)))];
+            tensor<fp16, [1024]> model_layers_9_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240431616)))];
+            tensor<fp16, [1024]> model_layers_9_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240433728)))];
             tensor<fp16, [1, 126, 1024]> input_527_cast_fp16 = layer_norm(axes = input_527_axes_0, beta = model_layers_9_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_9_norm_feed_forward2_weight_to_fp16, x = input_525_cast_fp16)[name = tensor<string, []>("input_527_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_9_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240456320))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244650688))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_9_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240435840))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244630208))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_feed_forward2_linear1_weight_to_fp16_quantized, x = input_527_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_531_cast_fp16 = silu(x = linear_89_cast_fp16)[name = tensor<string, []>("input_531_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_9_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244658944))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248853312))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_9_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_9_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244638464))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248832832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_9_feed_forward2_linear2_weight_to_fp16_quantized, x = input_531_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
             tensor<fp16, []> var_1828_to_fp16 = const()[name = tensor<string, []>("op_1828_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1829_cast_fp16 = mul(x = linear_90_cast_fp16, y = var_1828_to_fp16)[name = tensor<string, []>("op_1829_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_537_cast_fp16 = add(x = input_525_cast_fp16, y = var_1829_cast_fp16)[name = tensor<string, []>("input_537_cast_fp16")];
             tensor<int32, [1]> input_539_axes_0 = const()[name = tensor<string, []>("input_539_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248855424)))];
-            tensor<fp16, [1024]> model_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248857536)))];
+            tensor<fp16, [1024]> model_layers_9_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248834944)))];
+            tensor<fp16, [1024]> model_layers_9_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_9_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248837056)))];
             tensor<fp16, [1, 126, 1024]> input_539_cast_fp16 = layer_norm(axes = input_539_axes_0, beta = model_layers_9_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_9_norm_out_weight_to_fp16, x = input_537_cast_fp16)[name = tensor<string, []>("input_539_cast_fp16")];
             tensor<int32, [1]> input_541_axes_0 = const()[name = tensor<string, []>("input_541_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248859648)))];
-            tensor<fp16, [1024]> model_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248861760)))];
+            tensor<fp16, [1024]> model_layers_10_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248839168)))];
+            tensor<fp16, [1024]> model_layers_10_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248841280)))];
             tensor<fp16, [1, 126, 1024]> input_541_cast_fp16 = layer_norm(axes = input_541_axes_0, beta = model_layers_10_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_10_norm_feed_forward1_weight_to_fp16, x = input_539_cast_fp16)[name = tensor<string, []>("input_541_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_10_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248863872))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253058240))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_10_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(248843392))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253037760))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_feed_forward1_linear1_weight_to_fp16_quantized, x = input_541_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_545_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor<string, []>("input_545_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_10_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253066496))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257260864))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_10_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253046016))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257240384))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_92_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_10_feed_forward1_linear2_weight_to_fp16_quantized, x = input_545_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
             tensor<fp16, []> var_1857_to_fp16 = const()[name = tensor<string, []>("op_1857_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1858_cast_fp16 = mul(x = linear_92_cast_fp16, y = var_1857_to_fp16)[name = tensor<string, []>("op_1858_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_551_cast_fp16 = add(x = input_539_cast_fp16, y = var_1858_cast_fp16)[name = tensor<string, []>("input_551_cast_fp16")];
             tensor<int32, [1]> query_21_axes_0 = const()[name = tensor<string, []>("query_21_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257262976)))];
-            tensor<fp16, [1024]> model_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257265088)))];
+            tensor<fp16, [1024]> model_layers_10_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257242496)))];
+            tensor<fp16, [1024]> model_layers_10_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257244608)))];
             tensor<fp16, [1, 126, 1024]> query_21_cast_fp16 = layer_norm(axes = query_21_axes_0, beta = model_layers_10_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_10_norm_self_att_weight_to_fp16, x = input_551_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257267200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258315840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257246720))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258295360))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_93_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_10_self_attn_linear_q_weight_to_fp16_quantized, x = query_21_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
             tensor<int32, [4]> var_1874 = const()[name = tensor<string, []>("op_1874"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_61_cast_fp16 = reshape(shape = var_1874, x = linear_93_cast_fp16)[name = tensor<string, []>("q_61_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258317952))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259366592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258297472))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259346112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_94_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_10_self_attn_linear_k_weight_to_fp16_quantized, x = query_21_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
             tensor<int32, [4]> var_1878 = const()[name = tensor<string, []>("op_1878"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_41_cast_fp16 = reshape(shape = var_1878, x = linear_94_cast_fp16)[name = tensor<string, []>("k_41_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259368704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260417344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259348224))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260396864))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_95_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_10_self_attn_linear_v_weight_to_fp16_quantized, x = query_21_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
             tensor<int32, [4]> var_1882 = const()[name = tensor<string, []>("op_1882"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_21_cast_fp16 = reshape(shape = var_1882, x = linear_95_cast_fp16)[name = tensor<string, []>("v_21_cast_fp16")];
             tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<string, []>("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_10_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_10_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260419456)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1894_cast_fp16 = add(x = q_61_cast_fp16, y = model_layers_10_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_1894_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_10_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_10_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260421568)))];
-            tensor<fp16, [1, 126, 8, 128]> var_1896_cast_fp16 = add(x = q_61_cast_fp16, y = model_layers_10_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_1896_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_10_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260398976))), scale = tensor<fp16, [8]>([0x1.988p-8, 0x1.19p-7, 0x1.6b4p-7, 0x1.414p-8, 0x1.278p-8, 0x1.1d8p-7, 0x1.f08p-8, 0x1.488p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1894_cast_fp16 = add(x = q_61_cast_fp16, y = model_layers_10_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_1894_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_10_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260400064))), scale = tensor<fp16, [8]>([0x1.46p-7, 0x1.bccp-8, 0x1.158p-9, 0x1.e9p-9, 0x1.23p-8, 0x1.f2cp-9, 0x1.4bcp-7, 0x1.034p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_1896_cast_fp16 = add(x = q_61_cast_fp16, y = model_layers_10_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_1896_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_227_transpose_x_0 = const()[name = tensor<string, []>("x_227_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_227_transpose_y_0 = const()[name = tensor<string, []>("x_227_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_1898_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1898_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260423680))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260680768))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_1898_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1898_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260401152))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260658240))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_21_cast_fp16 = transpose(perm = q_with_bias_v_21_perm_0, x = var_1896_cast_fp16)[name = tensor<string, []>("transpose_241")];
             tensor<fp16, [1, 8, 126, 251]> x_227_cast_fp16 = matmul(transpose_x = x_227_transpose_x_0, transpose_y = x_227_transpose_y_0, x = q_with_bias_v_21_cast_fp16, y = op_1898_to_fp16_quantized)[name = tensor<string, []>("x_227_cast_fp16")];
             tensor<int32, [8]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1554,12 +1554,12 @@ program(1.0)
             tensor<int32, [3]> var_1931 = const()[name = tensor<string, []>("op_1931"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_1930_cast_fp16 = transpose(perm = var_1930_perm_0, x = x_233_cast_fp16)[name = tensor<string, []>("transpose_237")];
             tensor<fp16, [1, 126, 1024]> input_555_cast_fp16 = reshape(shape = var_1931, x = var_1930_cast_fp16)[name = tensor<string, []>("input_555_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260681344))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261729984))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260658816))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261707456))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_97_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_10_self_attn_linear_out_weight_to_fp16_quantized, x = input_555_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_559_cast_fp16 = add(x = input_551_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("input_559_cast_fp16")];
             tensor<int32, [1]> x_237_axes_0 = const()[name = tensor<string, []>("x_237_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261732096)))];
-            tensor<fp16, [1024]> model_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261734208)))];
+            tensor<fp16, [1024]> model_layers_10_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261709568)))];
+            tensor<fp16, [1024]> model_layers_10_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261711680)))];
             tensor<fp16, [1, 126, 1024]> x_237_cast_fp16 = layer_norm(axes = x_237_axes_0, beta = model_layers_10_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_10_norm_conv_weight_to_fp16, x = input_559_cast_fp16)[name = tensor<string, []>("x_237_cast_fp16")];
             tensor<int32, [3]> input_561_perm_0 = const()[name = tensor<string, []>("input_561_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_563_pad_type_0 = const()[name = tensor<string, []>("input_563_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1567,7 +1567,7 @@ program(1.0)
             tensor<int32, [2]> input_563_pad_0 = const()[name = tensor<string, []>("input_563_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_563_dilations_0 = const()[name = tensor<string, []>("input_563_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_563_groups_0 = const()[name = tensor<string, []>("input_563_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261736320))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263833536))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(261713792))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263811008))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_561_cast_fp16 = transpose(perm = input_561_perm_0, x = x_237_cast_fp16)[name = tensor<string, []>("transpose_236")];
             tensor<fp16, [1, 2048, 126]> input_563_cast_fp16 = conv(dilations = input_563_dilations_0, groups = input_563_groups_0, pad = input_563_pad_0, pad_type = input_563_pad_type_0, strides = input_563_strides_0, weight = model_layers_10_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_561_cast_fp16)[name = tensor<string, []>("input_563_cast_fp16")];
             tensor<int32, []> x_239_split_num_splits_0 = const()[name = tensor<string, []>("x_239_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1585,8 +1585,8 @@ program(1.0)
             tensor<int32, [1]> input_569_strides_0 = const()[name = tensor<string, []>("input_569_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_569_pad_0 = const()[name = tensor<string, []>("input_569_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_569_dilations_0 = const()[name = tensor<string, []>("input_569_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_268_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_268_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263837696))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263846976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_269_to_fp16 = const()[name = tensor<string, []>("const_269_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263849088)))];
+            tensor<fp16, [1024, 1, 9]> const_268_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_268_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263815168))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263824448))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_269_to_fp16 = const()[name = tensor<string, []>("const_269_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263826560)))];
             tensor<fp16, [1, 1024, 126]> input_571_cast_fp16 = conv(bias = const_269_to_fp16, dilations = input_569_dilations_0, groups = input_569_groups_0, pad = input_569_pad_0, pad_type = input_569_pad_type_0, strides = input_569_strides_0, weight = const_268_to_fp16_quantized, x = input_567_cast_fp16)[name = tensor<string, []>("input_571_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_573_cast_fp16 = silu(x = input_571_cast_fp16)[name = tensor<string, []>("input_573_cast_fp16")];
             tensor<string, []> x_241_pad_type_0 = const()[name = tensor<string, []>("x_241_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1594,64 +1594,64 @@ program(1.0)
             tensor<int32, [2]> x_241_pad_0 = const()[name = tensor<string, []>("x_241_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_241_dilations_0 = const()[name = tensor<string, []>("x_241_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_241_groups_0 = const()[name = tensor<string, []>("x_241_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263851200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264899840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263828672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264877312))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_241_cast_fp16 = conv(dilations = x_241_dilations_0, groups = x_241_groups_0, pad = x_241_pad_0, pad_type = x_241_pad_type_0, strides = x_241_strides_0, weight = model_layers_10_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_573_cast_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
             tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_575_cast_fp16 = transpose(perm = input_575_perm_0, x = x_241_cast_fp16)[name = tensor<string, []>("transpose_235")];
             tensor<fp16, [1, 126, 1024]> input_577_cast_fp16 = add(x = input_559_cast_fp16, y = input_575_cast_fp16)[name = tensor<string, []>("input_577_cast_fp16")];
             tensor<int32, [1]> input_579_axes_0 = const()[name = tensor<string, []>("input_579_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264901952)))];
-            tensor<fp16, [1024]> model_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264904064)))];
+            tensor<fp16, [1024]> model_layers_10_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264879424)))];
+            tensor<fp16, [1024]> model_layers_10_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264881536)))];
             tensor<fp16, [1, 126, 1024]> input_579_cast_fp16 = layer_norm(axes = input_579_axes_0, beta = model_layers_10_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_10_norm_feed_forward2_weight_to_fp16, x = input_577_cast_fp16)[name = tensor<string, []>("input_579_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_10_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264906176))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269100544))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_10_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264883648))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269078016))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_98_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_feed_forward2_linear1_weight_to_fp16_quantized, x = input_579_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_583_cast_fp16 = silu(x = linear_98_cast_fp16)[name = tensor<string, []>("input_583_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_10_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269108800))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273303168))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_10_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_10_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269086272))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273280640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_99_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_10_feed_forward2_linear2_weight_to_fp16_quantized, x = input_583_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
             tensor<fp16, []> var_1991_to_fp16 = const()[name = tensor<string, []>("op_1991_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_1992_cast_fp16 = mul(x = linear_99_cast_fp16, y = var_1991_to_fp16)[name = tensor<string, []>("op_1992_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_589_cast_fp16 = add(x = input_577_cast_fp16, y = var_1992_cast_fp16)[name = tensor<string, []>("input_589_cast_fp16")];
             tensor<int32, [1]> input_591_axes_0 = const()[name = tensor<string, []>("input_591_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273305280)))];
-            tensor<fp16, [1024]> model_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273307392)))];
+            tensor<fp16, [1024]> model_layers_10_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273282752)))];
+            tensor<fp16, [1024]> model_layers_10_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_10_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273284864)))];
             tensor<fp16, [1, 126, 1024]> input_591_cast_fp16 = layer_norm(axes = input_591_axes_0, beta = model_layers_10_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_10_norm_out_weight_to_fp16, x = input_589_cast_fp16)[name = tensor<string, []>("input_591_cast_fp16")];
             tensor<int32, [1]> input_593_axes_0 = const()[name = tensor<string, []>("input_593_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273309504)))];
-            tensor<fp16, [1024]> model_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273311616)))];
+            tensor<fp16, [1024]> model_layers_11_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273286976)))];
+            tensor<fp16, [1024]> model_layers_11_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273289088)))];
             tensor<fp16, [1, 126, 1024]> input_593_cast_fp16 = layer_norm(axes = input_593_axes_0, beta = model_layers_11_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_11_norm_feed_forward1_weight_to_fp16, x = input_591_cast_fp16)[name = tensor<string, []>("input_593_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_11_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273313728))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277508096))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_11_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(273291200))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277485568))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_feed_forward1_linear1_weight_to_fp16_quantized, x = input_593_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_597_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor<string, []>("input_597_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_11_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277516352))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281710720))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_11_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277493824))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281688192))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_101_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_11_feed_forward1_linear2_weight_to_fp16_quantized, x = input_597_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
             tensor<fp16, []> var_2020_to_fp16 = const()[name = tensor<string, []>("op_2020_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2021_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_2020_to_fp16)[name = tensor<string, []>("op_2021_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_603_cast_fp16 = add(x = input_591_cast_fp16, y = var_2021_cast_fp16)[name = tensor<string, []>("input_603_cast_fp16")];
             tensor<int32, [1]> query_23_axes_0 = const()[name = tensor<string, []>("query_23_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281712832)))];
-            tensor<fp16, [1024]> model_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281714944)))];
+            tensor<fp16, [1024]> model_layers_11_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281690304)))];
+            tensor<fp16, [1024]> model_layers_11_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281692416)))];
             tensor<fp16, [1, 126, 1024]> query_23_cast_fp16 = layer_norm(axes = query_23_axes_0, beta = model_layers_11_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_11_norm_self_att_weight_to_fp16, x = input_603_cast_fp16)[name = tensor<string, []>("query_23_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281717056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282765696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281694528))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282743168))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_102_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_11_self_attn_linear_q_weight_to_fp16_quantized, x = query_23_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
             tensor<int32, [4]> var_2037 = const()[name = tensor<string, []>("op_2037"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_67_cast_fp16 = reshape(shape = var_2037, x = linear_102_cast_fp16)[name = tensor<string, []>("q_67_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282767808))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283816448))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282745280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283793920))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_103_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_11_self_attn_linear_k_weight_to_fp16_quantized, x = query_23_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
             tensor<int32, [4]> var_2041 = const()[name = tensor<string, []>("op_2041"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_45_cast_fp16 = reshape(shape = var_2041, x = linear_103_cast_fp16)[name = tensor<string, []>("k_45_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283818560))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284867200))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283796032))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284844672))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_104_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_11_self_attn_linear_v_weight_to_fp16_quantized, x = query_23_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
             tensor<int32, [4]> var_2045 = const()[name = tensor<string, []>("op_2045"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_23_cast_fp16 = reshape(shape = var_2045, x = linear_104_cast_fp16)[name = tensor<string, []>("v_23_cast_fp16")];
             tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_11_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_11_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284869312)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2057_cast_fp16 = add(x = q_67_cast_fp16, y = model_layers_11_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2057_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_11_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_11_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284871424)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2059_cast_fp16 = add(x = q_67_cast_fp16, y = model_layers_11_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2059_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_11_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284846784))), scale = tensor<fp16, [8]>([0x1.d9cp-8, 0x1.a74p-8, 0x1.1e8p-7, 0x1.e1cp-8, 0x1.eb8p-8, 0x1.e38p-8, 0x1.f54p-8, 0x1.f64p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2057_cast_fp16 = add(x = q_67_cast_fp16, y = model_layers_11_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_2057_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_11_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284847872))), scale = tensor<fp16, [8]>([0x1.b58p-8, 0x1.32p-7, 0x1.338p-9, 0x1.dccp-8, 0x1.3a8p-9, 0x1.664p-8, 0x1.2a4p-7, 0x1.27p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2059_cast_fp16 = add(x = q_67_cast_fp16, y = model_layers_11_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_2059_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_249_transpose_x_0 = const()[name = tensor<string, []>("x_249_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_249_transpose_y_0 = const()[name = tensor<string, []>("x_249_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_2061_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2061_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284873536))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285130624))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_2061_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2061_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284848960))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285106048))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_23_cast_fp16 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2059_cast_fp16)[name = tensor<string, []>("transpose_234")];
             tensor<fp16, [1, 8, 126, 251]> x_249_cast_fp16 = matmul(transpose_x = x_249_transpose_x_0, transpose_y = x_249_transpose_y_0, x = q_with_bias_v_23_cast_fp16, y = op_2061_to_fp16_quantized)[name = tensor<string, []>("x_249_cast_fp16")];
             tensor<int32, [8]> x_251_pad_0 = const()[name = tensor<string, []>("x_251_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1691,12 +1691,12 @@ program(1.0)
             tensor<int32, [3]> var_2094 = const()[name = tensor<string, []>("op_2094"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_2093_cast_fp16 = transpose(perm = var_2093_perm_0, x = x_255_cast_fp16)[name = tensor<string, []>("transpose_230")];
             tensor<fp16, [1, 126, 1024]> input_607_cast_fp16 = reshape(shape = var_2094, x = var_2093_cast_fp16)[name = tensor<string, []>("input_607_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285131200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286179840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285106624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286155264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_106_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_11_self_attn_linear_out_weight_to_fp16_quantized, x = input_607_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_611_cast_fp16 = add(x = input_603_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input_611_cast_fp16")];
             tensor<int32, [1]> x_259_axes_0 = const()[name = tensor<string, []>("x_259_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286181952)))];
-            tensor<fp16, [1024]> model_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286184064)))];
+            tensor<fp16, [1024]> model_layers_11_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286157376)))];
+            tensor<fp16, [1024]> model_layers_11_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286159488)))];
             tensor<fp16, [1, 126, 1024]> x_259_cast_fp16 = layer_norm(axes = x_259_axes_0, beta = model_layers_11_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_11_norm_conv_weight_to_fp16, x = input_611_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
             tensor<int32, [3]> input_613_perm_0 = const()[name = tensor<string, []>("input_613_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_615_pad_type_0 = const()[name = tensor<string, []>("input_615_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1704,7 +1704,7 @@ program(1.0)
             tensor<int32, [2]> input_615_pad_0 = const()[name = tensor<string, []>("input_615_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_615_dilations_0 = const()[name = tensor<string, []>("input_615_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_615_groups_0 = const()[name = tensor<string, []>("input_615_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286186176))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288283392))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286161600))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288258816))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_613_cast_fp16 = transpose(perm = input_613_perm_0, x = x_259_cast_fp16)[name = tensor<string, []>("transpose_229")];
             tensor<fp16, [1, 2048, 126]> input_615_cast_fp16 = conv(dilations = input_615_dilations_0, groups = input_615_groups_0, pad = input_615_pad_0, pad_type = input_615_pad_type_0, strides = input_615_strides_0, weight = model_layers_11_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_613_cast_fp16)[name = tensor<string, []>("input_615_cast_fp16")];
             tensor<int32, []> x_261_split_num_splits_0 = const()[name = tensor<string, []>("x_261_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1722,8 +1722,8 @@ program(1.0)
             tensor<int32, [1]> input_621_strides_0 = const()[name = tensor<string, []>("input_621_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_621_pad_0 = const()[name = tensor<string, []>("input_621_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_621_dilations_0 = const()[name = tensor<string, []>("input_621_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_270_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_270_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288287552))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288296832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_271_to_fp16 = const()[name = tensor<string, []>("const_271_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288298944)))];
+            tensor<fp16, [1024, 1, 9]> const_270_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_270_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288262976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288272256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_271_to_fp16 = const()[name = tensor<string, []>("const_271_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288274368)))];
             tensor<fp16, [1, 1024, 126]> input_623_cast_fp16 = conv(bias = const_271_to_fp16, dilations = input_621_dilations_0, groups = input_621_groups_0, pad = input_621_pad_0, pad_type = input_621_pad_type_0, strides = input_621_strides_0, weight = const_270_to_fp16_quantized, x = input_619_cast_fp16)[name = tensor<string, []>("input_623_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_625_cast_fp16 = silu(x = input_623_cast_fp16)[name = tensor<string, []>("input_625_cast_fp16")];
             tensor<string, []> x_263_pad_type_0 = const()[name = tensor<string, []>("x_263_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1731,64 +1731,64 @@ program(1.0)
             tensor<int32, [2]> x_263_pad_0 = const()[name = tensor<string, []>("x_263_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_263_dilations_0 = const()[name = tensor<string, []>("x_263_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_263_groups_0 = const()[name = tensor<string, []>("x_263_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288301056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289349696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(288276480))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289325120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_263_cast_fp16 = conv(dilations = x_263_dilations_0, groups = x_263_groups_0, pad = x_263_pad_0, pad_type = x_263_pad_type_0, strides = x_263_strides_0, weight = model_layers_11_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_625_cast_fp16)[name = tensor<string, []>("x_263_cast_fp16")];
             tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_627_cast_fp16 = transpose(perm = input_627_perm_0, x = x_263_cast_fp16)[name = tensor<string, []>("transpose_228")];
             tensor<fp16, [1, 126, 1024]> input_629_cast_fp16 = add(x = input_611_cast_fp16, y = input_627_cast_fp16)[name = tensor<string, []>("input_629_cast_fp16")];
             tensor<int32, [1]> input_631_axes_0 = const()[name = tensor<string, []>("input_631_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289351808)))];
-            tensor<fp16, [1024]> model_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289353920)))];
+            tensor<fp16, [1024]> model_layers_11_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289327232)))];
+            tensor<fp16, [1024]> model_layers_11_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289329344)))];
             tensor<fp16, [1, 126, 1024]> input_631_cast_fp16 = layer_norm(axes = input_631_axes_0, beta = model_layers_11_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_11_norm_feed_forward2_weight_to_fp16, x = input_629_cast_fp16)[name = tensor<string, []>("input_631_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_11_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289356032))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293550400))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_11_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289331456))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293525824))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_feed_forward2_linear1_weight_to_fp16_quantized, x = input_631_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_635_cast_fp16 = silu(x = linear_107_cast_fp16)[name = tensor<string, []>("input_635_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_11_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293558656))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297753024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_11_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_11_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293534080))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297728448))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_108_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_11_feed_forward2_linear2_weight_to_fp16_quantized, x = input_635_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
             tensor<fp16, []> var_2154_to_fp16 = const()[name = tensor<string, []>("op_2154_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2155_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_2154_to_fp16)[name = tensor<string, []>("op_2155_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_641_cast_fp16 = add(x = input_629_cast_fp16, y = var_2155_cast_fp16)[name = tensor<string, []>("input_641_cast_fp16")];
             tensor<int32, [1]> input_643_axes_0 = const()[name = tensor<string, []>("input_643_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297755136)))];
-            tensor<fp16, [1024]> model_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297757248)))];
+            tensor<fp16, [1024]> model_layers_11_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297730560)))];
+            tensor<fp16, [1024]> model_layers_11_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_11_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297732672)))];
             tensor<fp16, [1, 126, 1024]> input_643_cast_fp16 = layer_norm(axes = input_643_axes_0, beta = model_layers_11_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_11_norm_out_weight_to_fp16, x = input_641_cast_fp16)[name = tensor<string, []>("input_643_cast_fp16")];
             tensor<int32, [1]> input_645_axes_0 = const()[name = tensor<string, []>("input_645_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297759360)))];
-            tensor<fp16, [1024]> model_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297761472)))];
+            tensor<fp16, [1024]> model_layers_12_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297734784)))];
+            tensor<fp16, [1024]> model_layers_12_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297736896)))];
             tensor<fp16, [1, 126, 1024]> input_645_cast_fp16 = layer_norm(axes = input_645_axes_0, beta = model_layers_12_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_12_norm_feed_forward1_weight_to_fp16, x = input_643_cast_fp16)[name = tensor<string, []>("input_645_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_12_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297763584))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301957952))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_12_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297739008))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301933376))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_109_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_feed_forward1_linear1_weight_to_fp16_quantized, x = input_645_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_649_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<string, []>("input_649_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_12_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301966208))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306160576))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_12_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301941632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306136000))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_110_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_12_feed_forward1_linear2_weight_to_fp16_quantized, x = input_649_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")];
             tensor<fp16, []> var_2183_to_fp16 = const()[name = tensor<string, []>("op_2183_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2184_cast_fp16 = mul(x = linear_110_cast_fp16, y = var_2183_to_fp16)[name = tensor<string, []>("op_2184_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_655_cast_fp16 = add(x = input_643_cast_fp16, y = var_2184_cast_fp16)[name = tensor<string, []>("input_655_cast_fp16")];
             tensor<int32, [1]> query_25_axes_0 = const()[name = tensor<string, []>("query_25_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306162688)))];
-            tensor<fp16, [1024]> model_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306164800)))];
+            tensor<fp16, [1024]> model_layers_12_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306138112)))];
+            tensor<fp16, [1024]> model_layers_12_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306140224)))];
             tensor<fp16, [1, 126, 1024]> query_25_cast_fp16 = layer_norm(axes = query_25_axes_0, beta = model_layers_12_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_12_norm_self_att_weight_to_fp16, x = input_655_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306166912))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307215552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306142336))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307190976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_111_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_12_self_attn_linear_q_weight_to_fp16_quantized, x = query_25_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
             tensor<int32, [4]> var_2200 = const()[name = tensor<string, []>("op_2200"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_73_cast_fp16 = reshape(shape = var_2200, x = linear_111_cast_fp16)[name = tensor<string, []>("q_73_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307217664))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308266304))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307193088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308241728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_112_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_12_self_attn_linear_k_weight_to_fp16_quantized, x = query_25_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
             tensor<int32, [4]> var_2204 = const()[name = tensor<string, []>("op_2204"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_49_cast_fp16 = reshape(shape = var_2204, x = linear_112_cast_fp16)[name = tensor<string, []>("k_49_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308268416))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309317056))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308243840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309292480))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_113_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_12_self_attn_linear_v_weight_to_fp16_quantized, x = query_25_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
             tensor<int32, [4]> var_2208 = const()[name = tensor<string, []>("op_2208"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_25_cast_fp16 = reshape(shape = var_2208, x = linear_113_cast_fp16)[name = tensor<string, []>("v_25_cast_fp16")];
             tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<string, []>("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_12_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_12_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309319168)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2220_cast_fp16 = add(x = q_73_cast_fp16, y = model_layers_12_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2220_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_12_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_12_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309321280)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2222_cast_fp16 = add(x = q_73_cast_fp16, y = model_layers_12_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2222_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_12_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309294592))), scale = tensor<fp16, [8]>([0x1.998p-8, 0x1.cdcp-8, 0x1.c98p-8, 0x1.16p-7, 0x1.084p-7, 0x1.17p-7, 0x1.2c8p-7, 0x1.2dcp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2220_cast_fp16 = add(x = q_73_cast_fp16, y = model_layers_12_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_2220_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_12_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309295680))), scale = tensor<fp16, [8]>([0x1.64p-9, 0x1.5b8p-7, 0x1.65p-8, 0x1.82p-8, 0x1.d54p-8, 0x1.138p-7, 0x1.40cp-7, 0x1.284p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2222_cast_fp16 = add(x = q_73_cast_fp16, y = model_layers_12_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_2222_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_271_transpose_x_0 = const()[name = tensor<string, []>("x_271_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_271_transpose_y_0 = const()[name = tensor<string, []>("x_271_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_2224_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2224_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309323392))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309580480))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_2224_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2224_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309296768))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309553856))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_25_cast_fp16 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2222_cast_fp16)[name = tensor<string, []>("transpose_227")];
             tensor<fp16, [1, 8, 126, 251]> x_271_cast_fp16 = matmul(transpose_x = x_271_transpose_x_0, transpose_y = x_271_transpose_y_0, x = q_with_bias_v_25_cast_fp16, y = op_2224_to_fp16_quantized)[name = tensor<string, []>("x_271_cast_fp16")];
             tensor<int32, [8]> x_273_pad_0 = const()[name = tensor<string, []>("x_273_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1828,12 +1828,12 @@ program(1.0)
             tensor<int32, [3]> var_2257 = const()[name = tensor<string, []>("op_2257"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_2256_cast_fp16 = transpose(perm = var_2256_perm_0, x = x_277_cast_fp16)[name = tensor<string, []>("transpose_223")];
             tensor<fp16, [1, 126, 1024]> input_659_cast_fp16 = reshape(shape = var_2257, x = var_2256_cast_fp16)[name = tensor<string, []>("input_659_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309581056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310629696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(309554432))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310603072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_115_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_12_self_attn_linear_out_weight_to_fp16_quantized, x = input_659_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_663_cast_fp16 = add(x = input_655_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("input_663_cast_fp16")];
             tensor<int32, [1]> x_281_axes_0 = const()[name = tensor<string, []>("x_281_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310631808)))];
-            tensor<fp16, [1024]> model_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310633920)))];
+            tensor<fp16, [1024]> model_layers_12_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310605184)))];
+            tensor<fp16, [1024]> model_layers_12_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310607296)))];
             tensor<fp16, [1, 126, 1024]> x_281_cast_fp16 = layer_norm(axes = x_281_axes_0, beta = model_layers_12_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_12_norm_conv_weight_to_fp16, x = input_663_cast_fp16)[name = tensor<string, []>("x_281_cast_fp16")];
             tensor<int32, [3]> input_665_perm_0 = const()[name = tensor<string, []>("input_665_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_667_pad_type_0 = const()[name = tensor<string, []>("input_667_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1841,7 +1841,7 @@ program(1.0)
             tensor<int32, [2]> input_667_pad_0 = const()[name = tensor<string, []>("input_667_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_667_dilations_0 = const()[name = tensor<string, []>("input_667_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_667_groups_0 = const()[name = tensor<string, []>("input_667_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310636032))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312733248))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310609408))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312706624))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_665_cast_fp16 = transpose(perm = input_665_perm_0, x = x_281_cast_fp16)[name = tensor<string, []>("transpose_222")];
             tensor<fp16, [1, 2048, 126]> input_667_cast_fp16 = conv(dilations = input_667_dilations_0, groups = input_667_groups_0, pad = input_667_pad_0, pad_type = input_667_pad_type_0, strides = input_667_strides_0, weight = model_layers_12_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_665_cast_fp16)[name = tensor<string, []>("input_667_cast_fp16")];
             tensor<int32, []> x_283_split_num_splits_0 = const()[name = tensor<string, []>("x_283_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1859,8 +1859,8 @@ program(1.0)
             tensor<int32, [1]> input_673_strides_0 = const()[name = tensor<string, []>("input_673_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_673_pad_0 = const()[name = tensor<string, []>("input_673_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_673_dilations_0 = const()[name = tensor<string, []>("input_673_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_272_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_272_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312737408))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312746688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_273_to_fp16 = const()[name = tensor<string, []>("const_273_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312748800)))];
+            tensor<fp16, [1024, 1, 9]> const_272_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_272_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312710784))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312720064))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_273_to_fp16 = const()[name = tensor<string, []>("const_273_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312722176)))];
             tensor<fp16, [1, 1024, 126]> input_675_cast_fp16 = conv(bias = const_273_to_fp16, dilations = input_673_dilations_0, groups = input_673_groups_0, pad = input_673_pad_0, pad_type = input_673_pad_type_0, strides = input_673_strides_0, weight = const_272_to_fp16_quantized, x = input_671_cast_fp16)[name = tensor<string, []>("input_675_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_677_cast_fp16 = silu(x = input_675_cast_fp16)[name = tensor<string, []>("input_677_cast_fp16")];
             tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1868,64 +1868,64 @@ program(1.0)
             tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312750912))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313799552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(312724288))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313772928))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_285_cast_fp16 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = model_layers_12_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_677_cast_fp16)[name = tensor<string, []>("x_285_cast_fp16")];
             tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_679_cast_fp16 = transpose(perm = input_679_perm_0, x = x_285_cast_fp16)[name = tensor<string, []>("transpose_221")];
             tensor<fp16, [1, 126, 1024]> input_681_cast_fp16 = add(x = input_663_cast_fp16, y = input_679_cast_fp16)[name = tensor<string, []>("input_681_cast_fp16")];
             tensor<int32, [1]> input_683_axes_0 = const()[name = tensor<string, []>("input_683_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313801664)))];
-            tensor<fp16, [1024]> model_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313803776)))];
+            tensor<fp16, [1024]> model_layers_12_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313775040)))];
+            tensor<fp16, [1024]> model_layers_12_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313777152)))];
             tensor<fp16, [1, 126, 1024]> input_683_cast_fp16 = layer_norm(axes = input_683_axes_0, beta = model_layers_12_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_12_norm_feed_forward2_weight_to_fp16, x = input_681_cast_fp16)[name = tensor<string, []>("input_683_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_12_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313805888))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318000256))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_12_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313779264))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317973632))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_116_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_feed_forward2_linear1_weight_to_fp16_quantized, x = input_683_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_687_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<string, []>("input_687_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_12_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318008512))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322202880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_12_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_12_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317981888))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322176256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_117_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_12_feed_forward2_linear2_weight_to_fp16_quantized, x = input_687_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
             tensor<fp16, []> var_2317_to_fp16 = const()[name = tensor<string, []>("op_2317_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2318_cast_fp16 = mul(x = linear_117_cast_fp16, y = var_2317_to_fp16)[name = tensor<string, []>("op_2318_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_693_cast_fp16 = add(x = input_681_cast_fp16, y = var_2318_cast_fp16)[name = tensor<string, []>("input_693_cast_fp16")];
             tensor<int32, [1]> input_695_axes_0 = const()[name = tensor<string, []>("input_695_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322204992)))];
-            tensor<fp16, [1024]> model_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322207104)))];
+            tensor<fp16, [1024]> model_layers_12_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322178368)))];
+            tensor<fp16, [1024]> model_layers_12_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_12_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322180480)))];
             tensor<fp16, [1, 126, 1024]> input_695_cast_fp16 = layer_norm(axes = input_695_axes_0, beta = model_layers_12_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_12_norm_out_weight_to_fp16, x = input_693_cast_fp16)[name = tensor<string, []>("input_695_cast_fp16")];
             tensor<int32, [1]> input_697_axes_0 = const()[name = tensor<string, []>("input_697_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322209216)))];
-            tensor<fp16, [1024]> model_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322211328)))];
+            tensor<fp16, [1024]> model_layers_13_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322182592)))];
+            tensor<fp16, [1024]> model_layers_13_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322184704)))];
             tensor<fp16, [1, 126, 1024]> input_697_cast_fp16 = layer_norm(axes = input_697_axes_0, beta = model_layers_13_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_13_norm_feed_forward1_weight_to_fp16, x = input_695_cast_fp16)[name = tensor<string, []>("input_697_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_13_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322213440))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326407808))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_13_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322186816))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326381184))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_feed_forward1_linear1_weight_to_fp16_quantized, x = input_697_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_701_cast_fp16 = silu(x = linear_118_cast_fp16)[name = tensor<string, []>("input_701_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_13_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326416064))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330610432))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_13_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326389440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330583808))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_119_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_13_feed_forward1_linear2_weight_to_fp16_quantized, x = input_701_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")];
             tensor<fp16, []> var_2346_to_fp16 = const()[name = tensor<string, []>("op_2346_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2347_cast_fp16 = mul(x = linear_119_cast_fp16, y = var_2346_to_fp16)[name = tensor<string, []>("op_2347_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_707_cast_fp16 = add(x = input_695_cast_fp16, y = var_2347_cast_fp16)[name = tensor<string, []>("input_707_cast_fp16")];
             tensor<int32, [1]> query_27_axes_0 = const()[name = tensor<string, []>("query_27_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330612544)))];
-            tensor<fp16, [1024]> model_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330614656)))];
+            tensor<fp16, [1024]> model_layers_13_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330585920)))];
+            tensor<fp16, [1024]> model_layers_13_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330588032)))];
             tensor<fp16, [1, 126, 1024]> query_27_cast_fp16 = layer_norm(axes = query_27_axes_0, beta = model_layers_13_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_13_norm_self_att_weight_to_fp16, x = input_707_cast_fp16)[name = tensor<string, []>("query_27_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330616768))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331665408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330590144))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331638784))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_120_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_13_self_attn_linear_q_weight_to_fp16_quantized, x = query_27_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")];
             tensor<int32, [4]> var_2363 = const()[name = tensor<string, []>("op_2363"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_79_cast_fp16 = reshape(shape = var_2363, x = linear_120_cast_fp16)[name = tensor<string, []>("q_79_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331667520))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332716160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331640896))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332689536))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_121_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_13_self_attn_linear_k_weight_to_fp16_quantized, x = query_27_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
             tensor<int32, [4]> var_2367 = const()[name = tensor<string, []>("op_2367"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_53_cast_fp16 = reshape(shape = var_2367, x = linear_121_cast_fp16)[name = tensor<string, []>("k_53_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332718272))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333766912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332691648))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333740288))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_122_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_13_self_attn_linear_v_weight_to_fp16_quantized, x = query_27_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
             tensor<int32, [4]> var_2371 = const()[name = tensor<string, []>("op_2371"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_27_cast_fp16 = reshape(shape = var_2371, x = linear_122_cast_fp16)[name = tensor<string, []>("v_27_cast_fp16")];
             tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_13_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_13_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333769024)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2383_cast_fp16 = add(x = q_79_cast_fp16, y = model_layers_13_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2383_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_13_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_13_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333771136)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2385_cast_fp16 = add(x = q_79_cast_fp16, y = model_layers_13_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2385_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_13_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333742400))), scale = tensor<fp16, [8]>([0x1.9fcp-7, 0x1.388p-7, 0x1.7a8p-8, 0x1.208p-6, 0x1.95p-7, 0x1.474p-7, 0x1.11cp-7, 0x1.514p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2383_cast_fp16 = add(x = q_79_cast_fp16, y = model_layers_13_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_2383_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_13_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333743488))), scale = tensor<fp16, [8]>([0x1.244p-7, 0x1.c98p-9, 0x1.fa8p-8, 0x1.accp-8, 0x1.d04p-8, 0x1.37cp-7, 0x1.f0cp-9, 0x1.07cp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2385_cast_fp16 = add(x = q_79_cast_fp16, y = model_layers_13_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_2385_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_2387_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2387_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333773248))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334030336))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_2387_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2387_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333744576))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334001664))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_27_cast_fp16 = transpose(perm = q_with_bias_v_27_perm_0, x = var_2385_cast_fp16)[name = tensor<string, []>("transpose_220")];
             tensor<fp16, [1, 8, 126, 251]> x_293_cast_fp16 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_27_cast_fp16, y = op_2387_to_fp16_quantized)[name = tensor<string, []>("x_293_cast_fp16")];
             tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -1965,12 +1965,12 @@ program(1.0)
             tensor<int32, [3]> var_2420 = const()[name = tensor<string, []>("op_2420"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_2419_cast_fp16 = transpose(perm = var_2419_perm_0, x = x_299_cast_fp16)[name = tensor<string, []>("transpose_216")];
             tensor<fp16, [1, 126, 1024]> input_711_cast_fp16 = reshape(shape = var_2420, x = var_2419_cast_fp16)[name = tensor<string, []>("input_711_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334030912))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335079552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(334002240))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335050880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_124_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_13_self_attn_linear_out_weight_to_fp16_quantized, x = input_711_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_715_cast_fp16 = add(x = input_707_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("input_715_cast_fp16")];
             tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335081664)))];
-            tensor<fp16, [1024]> model_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335083776)))];
+            tensor<fp16, [1024]> model_layers_13_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335052992)))];
+            tensor<fp16, [1024]> model_layers_13_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335055104)))];
             tensor<fp16, [1, 126, 1024]> x_303_cast_fp16 = layer_norm(axes = x_303_axes_0, beta = model_layers_13_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_13_norm_conv_weight_to_fp16, x = input_715_cast_fp16)[name = tensor<string, []>("x_303_cast_fp16")];
             tensor<int32, [3]> input_717_perm_0 = const()[name = tensor<string, []>("input_717_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_719_pad_type_0 = const()[name = tensor<string, []>("input_719_pad_type_0"), val = tensor<string, []>("valid")];
@@ -1978,7 +1978,7 @@ program(1.0)
             tensor<int32, [2]> input_719_pad_0 = const()[name = tensor<string, []>("input_719_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_719_dilations_0 = const()[name = tensor<string, []>("input_719_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_719_groups_0 = const()[name = tensor<string, []>("input_719_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335085888))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337183104))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335057216))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337154432))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_717_cast_fp16 = transpose(perm = input_717_perm_0, x = x_303_cast_fp16)[name = tensor<string, []>("transpose_215")];
             tensor<fp16, [1, 2048, 126]> input_719_cast_fp16 = conv(dilations = input_719_dilations_0, groups = input_719_groups_0, pad = input_719_pad_0, pad_type = input_719_pad_type_0, strides = input_719_strides_0, weight = model_layers_13_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_717_cast_fp16)[name = tensor<string, []>("input_719_cast_fp16")];
             tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -1996,8 +1996,8 @@ program(1.0)
             tensor<int32, [1]> input_725_strides_0 = const()[name = tensor<string, []>("input_725_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_725_pad_0 = const()[name = tensor<string, []>("input_725_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_725_dilations_0 = const()[name = tensor<string, []>("input_725_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_274_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_274_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337187264))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337196544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_275_to_fp16 = const()[name = tensor<string, []>("const_275_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337198656)))];
+            tensor<fp16, [1024, 1, 9]> const_274_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_274_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337158592))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337167872))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_275_to_fp16 = const()[name = tensor<string, []>("const_275_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337169984)))];
             tensor<fp16, [1, 1024, 126]> input_727_cast_fp16 = conv(bias = const_275_to_fp16, dilations = input_725_dilations_0, groups = input_725_groups_0, pad = input_725_pad_0, pad_type = input_725_pad_type_0, strides = input_725_strides_0, weight = const_274_to_fp16_quantized, x = input_723_cast_fp16)[name = tensor<string, []>("input_727_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_729_cast_fp16 = silu(x = input_727_cast_fp16)[name = tensor<string, []>("input_729_cast_fp16")];
             tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2005,64 +2005,64 @@ program(1.0)
             tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<string, []>("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337200768))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338249408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(337172096))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338220736))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_307_cast_fp16 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = model_layers_13_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_729_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
             tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_731_cast_fp16 = transpose(perm = input_731_perm_0, x = x_307_cast_fp16)[name = tensor<string, []>("transpose_214")];
             tensor<fp16, [1, 126, 1024]> input_733_cast_fp16 = add(x = input_715_cast_fp16, y = input_731_cast_fp16)[name = tensor<string, []>("input_733_cast_fp16")];
             tensor<int32, [1]> input_735_axes_0 = const()[name = tensor<string, []>("input_735_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338251520)))];
-            tensor<fp16, [1024]> model_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338253632)))];
+            tensor<fp16, [1024]> model_layers_13_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338222848)))];
+            tensor<fp16, [1024]> model_layers_13_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338224960)))];
             tensor<fp16, [1, 126, 1024]> input_735_cast_fp16 = layer_norm(axes = input_735_axes_0, beta = model_layers_13_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_13_norm_feed_forward2_weight_to_fp16, x = input_733_cast_fp16)[name = tensor<string, []>("input_735_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_13_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338255744))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342450112))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_13_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338227072))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342421440))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_feed_forward2_linear1_weight_to_fp16_quantized, x = input_735_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_739_cast_fp16 = silu(x = linear_125_cast_fp16)[name = tensor<string, []>("input_739_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_13_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342458368))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346652736))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_13_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_13_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342429696))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346624064))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_126_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_13_feed_forward2_linear2_weight_to_fp16_quantized, x = input_739_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
             tensor<fp16, []> var_2480_to_fp16 = const()[name = tensor<string, []>("op_2480_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2481_cast_fp16 = mul(x = linear_126_cast_fp16, y = var_2480_to_fp16)[name = tensor<string, []>("op_2481_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_745_cast_fp16 = add(x = input_733_cast_fp16, y = var_2481_cast_fp16)[name = tensor<string, []>("input_745_cast_fp16")];
             tensor<int32, [1]> input_747_axes_0 = const()[name = tensor<string, []>("input_747_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346654848)))];
-            tensor<fp16, [1024]> model_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346656960)))];
+            tensor<fp16, [1024]> model_layers_13_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346626176)))];
+            tensor<fp16, [1024]> model_layers_13_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_13_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346628288)))];
             tensor<fp16, [1, 126, 1024]> input_747_cast_fp16 = layer_norm(axes = input_747_axes_0, beta = model_layers_13_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_13_norm_out_weight_to_fp16, x = input_745_cast_fp16)[name = tensor<string, []>("input_747_cast_fp16")];
             tensor<int32, [1]> input_749_axes_0 = const()[name = tensor<string, []>("input_749_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346659072)))];
-            tensor<fp16, [1024]> model_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346661184)))];
+            tensor<fp16, [1024]> model_layers_14_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346630400)))];
+            tensor<fp16, [1024]> model_layers_14_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346632512)))];
             tensor<fp16, [1, 126, 1024]> input_749_cast_fp16 = layer_norm(axes = input_749_axes_0, beta = model_layers_14_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_14_norm_feed_forward1_weight_to_fp16, x = input_747_cast_fp16)[name = tensor<string, []>("input_749_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_14_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346663296))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350857664))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_14_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(346634624))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350828992))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_feed_forward1_linear1_weight_to_fp16_quantized, x = input_749_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_753_cast_fp16 = silu(x = linear_127_cast_fp16)[name = tensor<string, []>("input_753_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_14_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350865920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355060288))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_14_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(350837248))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355031616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_128_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_14_feed_forward1_linear2_weight_to_fp16_quantized, x = input_753_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
             tensor<fp16, []> var_2509_to_fp16 = const()[name = tensor<string, []>("op_2509_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2510_cast_fp16 = mul(x = linear_128_cast_fp16, y = var_2509_to_fp16)[name = tensor<string, []>("op_2510_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_759_cast_fp16 = add(x = input_747_cast_fp16, y = var_2510_cast_fp16)[name = tensor<string, []>("input_759_cast_fp16")];
             tensor<int32, [1]> query_29_axes_0 = const()[name = tensor<string, []>("query_29_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355062400)))];
-            tensor<fp16, [1024]> model_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355064512)))];
+            tensor<fp16, [1024]> model_layers_14_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355033728)))];
+            tensor<fp16, [1024]> model_layers_14_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355035840)))];
             tensor<fp16, [1, 126, 1024]> query_29_cast_fp16 = layer_norm(axes = query_29_axes_0, beta = model_layers_14_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_14_norm_self_att_weight_to_fp16, x = input_759_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355066624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356115264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355037952))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356086592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_129_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_14_self_attn_linear_q_weight_to_fp16_quantized, x = query_29_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")];
             tensor<int32, [4]> var_2526 = const()[name = tensor<string, []>("op_2526"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_85_cast_fp16 = reshape(shape = var_2526, x = linear_129_cast_fp16)[name = tensor<string, []>("q_85_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356117376))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357166016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356088704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357137344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_130_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_14_self_attn_linear_k_weight_to_fp16_quantized, x = query_29_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")];
             tensor<int32, [4]> var_2530 = const()[name = tensor<string, []>("op_2530"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_57_cast_fp16 = reshape(shape = var_2530, x = linear_130_cast_fp16)[name = tensor<string, []>("k_57_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357168128))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358216768))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357139456))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358188096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_131_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_14_self_attn_linear_v_weight_to_fp16_quantized, x = query_29_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
             tensor<int32, [4]> var_2534 = const()[name = tensor<string, []>("op_2534"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_29_cast_fp16 = reshape(shape = var_2534, x = linear_131_cast_fp16)[name = tensor<string, []>("v_29_cast_fp16")];
             tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<string, []>("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_14_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_14_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358218880)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2546_cast_fp16 = add(x = q_85_cast_fp16, y = model_layers_14_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2546_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_14_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_14_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358220992)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2548_cast_fp16 = add(x = q_85_cast_fp16, y = model_layers_14_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2548_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_14_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358190208))), scale = tensor<fp16, [8]>([0x1.37p-7, 0x1.7c8p-7, 0x1.024p-6, 0x1.61p-7, 0x1.d9cp-8, 0x1.8c4p-7, 0x1.158p-7, 0x1.9d8p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2546_cast_fp16 = add(x = q_85_cast_fp16, y = model_layers_14_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_2546_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_14_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358191296))), scale = tensor<fp16, [8]>([0x1.958p-8, 0x1.624p-9, 0x1.388p-8, 0x1.554p-7, 0x1.24p-7, 0x1.93p-8, 0x1.d6cp-8, 0x1.998p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2548_cast_fp16 = add(x = q_85_cast_fp16, y = model_layers_14_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_2548_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_315_transpose_x_0 = const()[name = tensor<string, []>("x_315_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_315_transpose_y_0 = const()[name = tensor<string, []>("x_315_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_2550_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2550_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358223104))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358480192))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_2550_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2550_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358192384))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358449472))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_29_cast_fp16 = transpose(perm = q_with_bias_v_29_perm_0, x = var_2548_cast_fp16)[name = tensor<string, []>("transpose_213")];
             tensor<fp16, [1, 8, 126, 251]> x_315_cast_fp16 = matmul(transpose_x = x_315_transpose_x_0, transpose_y = x_315_transpose_y_0, x = q_with_bias_v_29_cast_fp16, y = op_2550_to_fp16_quantized)[name = tensor<string, []>("x_315_cast_fp16")];
             tensor<int32, [8]> x_317_pad_0 = const()[name = tensor<string, []>("x_317_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -2102,12 +2102,12 @@ program(1.0)
             tensor<int32, [3]> var_2583 = const()[name = tensor<string, []>("op_2583"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_2582_cast_fp16 = transpose(perm = var_2582_perm_0, x = x_321_cast_fp16)[name = tensor<string, []>("transpose_209")];
             tensor<fp16, [1, 126, 1024]> input_763_cast_fp16 = reshape(shape = var_2583, x = var_2582_cast_fp16)[name = tensor<string, []>("input_763_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358480768))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359529408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358450048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359498688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_133_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_14_self_attn_linear_out_weight_to_fp16_quantized, x = input_763_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_767_cast_fp16 = add(x = input_759_cast_fp16, y = linear_133_cast_fp16)[name = tensor<string, []>("input_767_cast_fp16")];
             tensor<int32, [1]> x_325_axes_0 = const()[name = tensor<string, []>("x_325_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359531520)))];
-            tensor<fp16, [1024]> model_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359533632)))];
+            tensor<fp16, [1024]> model_layers_14_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359500800)))];
+            tensor<fp16, [1024]> model_layers_14_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359502912)))];
             tensor<fp16, [1, 126, 1024]> x_325_cast_fp16 = layer_norm(axes = x_325_axes_0, beta = model_layers_14_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_14_norm_conv_weight_to_fp16, x = input_767_cast_fp16)[name = tensor<string, []>("x_325_cast_fp16")];
             tensor<int32, [3]> input_769_perm_0 = const()[name = tensor<string, []>("input_769_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_771_pad_type_0 = const()[name = tensor<string, []>("input_771_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2115,7 +2115,7 @@ program(1.0)
             tensor<int32, [2]> input_771_pad_0 = const()[name = tensor<string, []>("input_771_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_771_dilations_0 = const()[name = tensor<string, []>("input_771_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_771_groups_0 = const()[name = tensor<string, []>("input_771_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359535744))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361632960))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359505024))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361602240))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_769_cast_fp16 = transpose(perm = input_769_perm_0, x = x_325_cast_fp16)[name = tensor<string, []>("transpose_208")];
             tensor<fp16, [1, 2048, 126]> input_771_cast_fp16 = conv(dilations = input_771_dilations_0, groups = input_771_groups_0, pad = input_771_pad_0, pad_type = input_771_pad_type_0, strides = input_771_strides_0, weight = model_layers_14_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_769_cast_fp16)[name = tensor<string, []>("input_771_cast_fp16")];
             tensor<int32, []> x_327_split_num_splits_0 = const()[name = tensor<string, []>("x_327_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -2133,8 +2133,8 @@ program(1.0)
             tensor<int32, [1]> input_777_strides_0 = const()[name = tensor<string, []>("input_777_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_777_pad_0 = const()[name = tensor<string, []>("input_777_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_777_dilations_0 = const()[name = tensor<string, []>("input_777_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_276_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_276_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361637120))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361646400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_277_to_fp16 = const()[name = tensor<string, []>("const_277_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361648512)))];
+            tensor<fp16, [1024, 1, 9]> const_276_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_276_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361606400))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361615680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_277_to_fp16 = const()[name = tensor<string, []>("const_277_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361617792)))];
             tensor<fp16, [1, 1024, 126]> input_779_cast_fp16 = conv(bias = const_277_to_fp16, dilations = input_777_dilations_0, groups = input_777_groups_0, pad = input_777_pad_0, pad_type = input_777_pad_type_0, strides = input_777_strides_0, weight = const_276_to_fp16_quantized, x = input_775_cast_fp16)[name = tensor<string, []>("input_779_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_781_cast_fp16 = silu(x = input_779_cast_fp16)[name = tensor<string, []>("input_781_cast_fp16")];
             tensor<string, []> x_329_pad_type_0 = const()[name = tensor<string, []>("x_329_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2142,64 +2142,64 @@ program(1.0)
             tensor<int32, [2]> x_329_pad_0 = const()[name = tensor<string, []>("x_329_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_329_dilations_0 = const()[name = tensor<string, []>("x_329_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_329_groups_0 = const()[name = tensor<string, []>("x_329_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361650624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362699264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361619904))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362668544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_329_cast_fp16 = conv(dilations = x_329_dilations_0, groups = x_329_groups_0, pad = x_329_pad_0, pad_type = x_329_pad_type_0, strides = x_329_strides_0, weight = model_layers_14_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_781_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
             tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_783_cast_fp16 = transpose(perm = input_783_perm_0, x = x_329_cast_fp16)[name = tensor<string, []>("transpose_207")];
             tensor<fp16, [1, 126, 1024]> input_785_cast_fp16 = add(x = input_767_cast_fp16, y = input_783_cast_fp16)[name = tensor<string, []>("input_785_cast_fp16")];
             tensor<int32, [1]> input_787_axes_0 = const()[name = tensor<string, []>("input_787_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362701376)))];
-            tensor<fp16, [1024]> model_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362703488)))];
+            tensor<fp16, [1024]> model_layers_14_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362670656)))];
+            tensor<fp16, [1024]> model_layers_14_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362672768)))];
             tensor<fp16, [1, 126, 1024]> input_787_cast_fp16 = layer_norm(axes = input_787_axes_0, beta = model_layers_14_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_14_norm_feed_forward2_weight_to_fp16, x = input_785_cast_fp16)[name = tensor<string, []>("input_787_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_14_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362705600))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366899968))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_14_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(362674880))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366869248))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_feed_forward2_linear1_weight_to_fp16_quantized, x = input_787_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_791_cast_fp16 = silu(x = linear_134_cast_fp16)[name = tensor<string, []>("input_791_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_14_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366908224))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371102592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_14_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_14_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366877504))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371071872))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_135_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_14_feed_forward2_linear2_weight_to_fp16_quantized, x = input_791_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
             tensor<fp16, []> var_2643_to_fp16 = const()[name = tensor<string, []>("op_2643_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2644_cast_fp16 = mul(x = linear_135_cast_fp16, y = var_2643_to_fp16)[name = tensor<string, []>("op_2644_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_797_cast_fp16 = add(x = input_785_cast_fp16, y = var_2644_cast_fp16)[name = tensor<string, []>("input_797_cast_fp16")];
             tensor<int32, [1]> input_799_axes_0 = const()[name = tensor<string, []>("input_799_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371104704)))];
-            tensor<fp16, [1024]> model_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371106816)))];
+            tensor<fp16, [1024]> model_layers_14_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371073984)))];
+            tensor<fp16, [1024]> model_layers_14_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_14_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371076096)))];
             tensor<fp16, [1, 126, 1024]> input_799_cast_fp16 = layer_norm(axes = input_799_axes_0, beta = model_layers_14_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_14_norm_out_weight_to_fp16, x = input_797_cast_fp16)[name = tensor<string, []>("input_799_cast_fp16")];
             tensor<int32, [1]> input_801_axes_0 = const()[name = tensor<string, []>("input_801_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371108928)))];
-            tensor<fp16, [1024]> model_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371111040)))];
+            tensor<fp16, [1024]> model_layers_15_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371078208)))];
+            tensor<fp16, [1024]> model_layers_15_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371080320)))];
             tensor<fp16, [1, 126, 1024]> input_801_cast_fp16 = layer_norm(axes = input_801_axes_0, beta = model_layers_15_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_15_norm_feed_forward1_weight_to_fp16, x = input_799_cast_fp16)[name = tensor<string, []>("input_801_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_15_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371113152))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375307520))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_15_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(371082432))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375276800))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_feed_forward1_linear1_weight_to_fp16_quantized, x = input_801_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_805_cast_fp16 = silu(x = linear_136_cast_fp16)[name = tensor<string, []>("input_805_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_15_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375315776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379510144))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_15_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375285056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379479424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_137_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_15_feed_forward1_linear2_weight_to_fp16_quantized, x = input_805_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
             tensor<fp16, []> var_2672_to_fp16 = const()[name = tensor<string, []>("op_2672_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2673_cast_fp16 = mul(x = linear_137_cast_fp16, y = var_2672_to_fp16)[name = tensor<string, []>("op_2673_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_811_cast_fp16 = add(x = input_799_cast_fp16, y = var_2673_cast_fp16)[name = tensor<string, []>("input_811_cast_fp16")];
             tensor<int32, [1]> query_31_axes_0 = const()[name = tensor<string, []>("query_31_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379512256)))];
-            tensor<fp16, [1024]> model_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379514368)))];
+            tensor<fp16, [1024]> model_layers_15_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379481536)))];
+            tensor<fp16, [1024]> model_layers_15_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379483648)))];
             tensor<fp16, [1, 126, 1024]> query_31_cast_fp16 = layer_norm(axes = query_31_axes_0, beta = model_layers_15_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_15_norm_self_att_weight_to_fp16, x = input_811_cast_fp16)[name = tensor<string, []>("query_31_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379516480))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380565120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379485760))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380534400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_138_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_15_self_attn_linear_q_weight_to_fp16_quantized, x = query_31_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
             tensor<int32, [4]> var_2689 = const()[name = tensor<string, []>("op_2689"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_91_cast_fp16 = reshape(shape = var_2689, x = linear_138_cast_fp16)[name = tensor<string, []>("q_91_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380567232))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381615872))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380536512))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381585152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_139_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_15_self_attn_linear_k_weight_to_fp16_quantized, x = query_31_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")];
             tensor<int32, [4]> var_2693 = const()[name = tensor<string, []>("op_2693"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_61_cast_fp16 = reshape(shape = var_2693, x = linear_139_cast_fp16)[name = tensor<string, []>("k_61_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381617984))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382666624))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381587264))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382635904))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_140_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_15_self_attn_linear_v_weight_to_fp16_quantized, x = query_31_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")];
             tensor<int32, [4]> var_2697 = const()[name = tensor<string, []>("op_2697"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_31_cast_fp16 = reshape(shape = var_2697, x = linear_140_cast_fp16)[name = tensor<string, []>("v_31_cast_fp16")];
             tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<string, []>("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_15_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_15_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382668736)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2709_cast_fp16 = add(x = q_91_cast_fp16, y = model_layers_15_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2709_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_15_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_15_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382670848)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2711_cast_fp16 = add(x = q_91_cast_fp16, y = model_layers_15_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2711_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_15_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382638016))), scale = tensor<fp16, [8]>([0x1.32p-8, 0x1.724p-8, 0x1.874p-8, 0x1.ad4p-8, 0x1.298p-8, 0x1.968p-8, 0x1.9a4p-8, 0x1.c4cp-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2709_cast_fp16 = add(x = q_91_cast_fp16, y = model_layers_15_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_2709_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_15_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382639104))), scale = tensor<fp16, [8]>([0x1.d04p-9, 0x1.c94p-8, 0x1.ca4p-8, 0x1.1fcp-7, 0x1.6dcp-8, 0x1.2fp-8, 0x1.a74p-8, 0x1.0a8p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2711_cast_fp16 = add(x = q_91_cast_fp16, y = model_layers_15_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_2711_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_337_transpose_x_0 = const()[name = tensor<string, []>("x_337_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_337_transpose_y_0 = const()[name = tensor<string, []>("x_337_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_2713_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2713_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382672960))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382930048))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_2713_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2713_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382640192))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382897280))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_31_cast_fp16 = transpose(perm = q_with_bias_v_31_perm_0, x = var_2711_cast_fp16)[name = tensor<string, []>("transpose_206")];
             tensor<fp16, [1, 8, 126, 251]> x_337_cast_fp16 = matmul(transpose_x = x_337_transpose_x_0, transpose_y = x_337_transpose_y_0, x = q_with_bias_v_31_cast_fp16, y = op_2713_to_fp16_quantized)[name = tensor<string, []>("x_337_cast_fp16")];
             tensor<int32, [8]> x_339_pad_0 = const()[name = tensor<string, []>("x_339_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -2239,12 +2239,12 @@ program(1.0)
             tensor<int32, [3]> var_2746 = const()[name = tensor<string, []>("op_2746"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_2745_cast_fp16 = transpose(perm = var_2745_perm_0, x = x_343_cast_fp16)[name = tensor<string, []>("transpose_202")];
             tensor<fp16, [1, 126, 1024]> input_815_cast_fp16 = reshape(shape = var_2746, x = var_2745_cast_fp16)[name = tensor<string, []>("input_815_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382930624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383979264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382897856))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383946496))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_142_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_15_self_attn_linear_out_weight_to_fp16_quantized, x = input_815_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_819_cast_fp16 = add(x = input_811_cast_fp16, y = linear_142_cast_fp16)[name = tensor<string, []>("input_819_cast_fp16")];
             tensor<int32, [1]> x_347_axes_0 = const()[name = tensor<string, []>("x_347_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383981376)))];
-            tensor<fp16, [1024]> model_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383983488)))];
+            tensor<fp16, [1024]> model_layers_15_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383948608)))];
+            tensor<fp16, [1024]> model_layers_15_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383950720)))];
             tensor<fp16, [1, 126, 1024]> x_347_cast_fp16 = layer_norm(axes = x_347_axes_0, beta = model_layers_15_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_15_norm_conv_weight_to_fp16, x = input_819_cast_fp16)[name = tensor<string, []>("x_347_cast_fp16")];
             tensor<int32, [3]> input_821_perm_0 = const()[name = tensor<string, []>("input_821_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_823_pad_type_0 = const()[name = tensor<string, []>("input_823_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2252,7 +2252,7 @@ program(1.0)
             tensor<int32, [2]> input_823_pad_0 = const()[name = tensor<string, []>("input_823_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_823_dilations_0 = const()[name = tensor<string, []>("input_823_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_823_groups_0 = const()[name = tensor<string, []>("input_823_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383985600))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386082816))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383952832))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386050048))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_821_cast_fp16 = transpose(perm = input_821_perm_0, x = x_347_cast_fp16)[name = tensor<string, []>("transpose_201")];
             tensor<fp16, [1, 2048, 126]> input_823_cast_fp16 = conv(dilations = input_823_dilations_0, groups = input_823_groups_0, pad = input_823_pad_0, pad_type = input_823_pad_type_0, strides = input_823_strides_0, weight = model_layers_15_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_821_cast_fp16)[name = tensor<string, []>("input_823_cast_fp16")];
             tensor<int32, []> x_349_split_num_splits_0 = const()[name = tensor<string, []>("x_349_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -2270,8 +2270,8 @@ program(1.0)
             tensor<int32, [1]> input_829_strides_0 = const()[name = tensor<string, []>("input_829_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_829_pad_0 = const()[name = tensor<string, []>("input_829_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_829_dilations_0 = const()[name = tensor<string, []>("input_829_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_278_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_278_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386086976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386096256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_279_to_fp16 = const()[name = tensor<string, []>("const_279_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386098368)))];
+            tensor<fp16, [1024, 1, 9]> const_278_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_278_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386054208))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386063488))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_279_to_fp16 = const()[name = tensor<string, []>("const_279_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386065600)))];
             tensor<fp16, [1, 1024, 126]> input_831_cast_fp16 = conv(bias = const_279_to_fp16, dilations = input_829_dilations_0, groups = input_829_groups_0, pad = input_829_pad_0, pad_type = input_829_pad_type_0, strides = input_829_strides_0, weight = const_278_to_fp16_quantized, x = input_827_cast_fp16)[name = tensor<string, []>("input_831_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_833_cast_fp16 = silu(x = input_831_cast_fp16)[name = tensor<string, []>("input_833_cast_fp16")];
             tensor<string, []> x_351_pad_type_0 = const()[name = tensor<string, []>("x_351_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2279,64 +2279,64 @@ program(1.0)
             tensor<int32, [2]> x_351_pad_0 = const()[name = tensor<string, []>("x_351_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_351_dilations_0 = const()[name = tensor<string, []>("x_351_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_351_groups_0 = const()[name = tensor<string, []>("x_351_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386100480))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387149120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(386067712))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387116352))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_351_cast_fp16 = conv(dilations = x_351_dilations_0, groups = x_351_groups_0, pad = x_351_pad_0, pad_type = x_351_pad_type_0, strides = x_351_strides_0, weight = model_layers_15_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_833_cast_fp16)[name = tensor<string, []>("x_351_cast_fp16")];
             tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_835_cast_fp16 = transpose(perm = input_835_perm_0, x = x_351_cast_fp16)[name = tensor<string, []>("transpose_200")];
             tensor<fp16, [1, 126, 1024]> input_837_cast_fp16 = add(x = input_819_cast_fp16, y = input_835_cast_fp16)[name = tensor<string, []>("input_837_cast_fp16")];
             tensor<int32, [1]> input_839_axes_0 = const()[name = tensor<string, []>("input_839_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387151232)))];
-            tensor<fp16, [1024]> model_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387153344)))];
+            tensor<fp16, [1024]> model_layers_15_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387118464)))];
+            tensor<fp16, [1024]> model_layers_15_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387120576)))];
             tensor<fp16, [1, 126, 1024]> input_839_cast_fp16 = layer_norm(axes = input_839_axes_0, beta = model_layers_15_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_15_norm_feed_forward2_weight_to_fp16, x = input_837_cast_fp16)[name = tensor<string, []>("input_839_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_15_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387155456))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391349824))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_15_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387122688))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391317056))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_feed_forward2_linear1_weight_to_fp16_quantized, x = input_839_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_843_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor<string, []>("input_843_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_15_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391358080))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395552448))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_15_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_15_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391325312))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395519680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_144_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_15_feed_forward2_linear2_weight_to_fp16_quantized, x = input_843_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
             tensor<fp16, []> var_2806_to_fp16 = const()[name = tensor<string, []>("op_2806_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2807_cast_fp16 = mul(x = linear_144_cast_fp16, y = var_2806_to_fp16)[name = tensor<string, []>("op_2807_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_849_cast_fp16 = add(x = input_837_cast_fp16, y = var_2807_cast_fp16)[name = tensor<string, []>("input_849_cast_fp16")];
             tensor<int32, [1]> input_851_axes_0 = const()[name = tensor<string, []>("input_851_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395554560)))];
-            tensor<fp16, [1024]> model_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395556672)))];
+            tensor<fp16, [1024]> model_layers_15_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395521792)))];
+            tensor<fp16, [1024]> model_layers_15_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_15_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395523904)))];
             tensor<fp16, [1, 126, 1024]> input_851_cast_fp16 = layer_norm(axes = input_851_axes_0, beta = model_layers_15_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_15_norm_out_weight_to_fp16, x = input_849_cast_fp16)[name = tensor<string, []>("input_851_cast_fp16")];
             tensor<int32, [1]> input_853_axes_0 = const()[name = tensor<string, []>("input_853_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395558784)))];
-            tensor<fp16, [1024]> model_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395560896)))];
+            tensor<fp16, [1024]> model_layers_16_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395526016)))];
+            tensor<fp16, [1024]> model_layers_16_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395528128)))];
             tensor<fp16, [1, 126, 1024]> input_853_cast_fp16 = layer_norm(axes = input_853_axes_0, beta = model_layers_16_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_16_norm_feed_forward1_weight_to_fp16, x = input_851_cast_fp16)[name = tensor<string, []>("input_853_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_16_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395563008))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399757376))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_16_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(395530240))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399724608))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_145_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_feed_forward1_linear1_weight_to_fp16_quantized, x = input_853_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_857_cast_fp16 = silu(x = linear_145_cast_fp16)[name = tensor<string, []>("input_857_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_16_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399765632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403960000))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_16_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399732864))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403927232))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_146_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_16_feed_forward1_linear2_weight_to_fp16_quantized, x = input_857_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
             tensor<fp16, []> var_2835_to_fp16 = const()[name = tensor<string, []>("op_2835_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2836_cast_fp16 = mul(x = linear_146_cast_fp16, y = var_2835_to_fp16)[name = tensor<string, []>("op_2836_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_863_cast_fp16 = add(x = input_851_cast_fp16, y = var_2836_cast_fp16)[name = tensor<string, []>("input_863_cast_fp16")];
             tensor<int32, [1]> query_33_axes_0 = const()[name = tensor<string, []>("query_33_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403962112)))];
-            tensor<fp16, [1024]> model_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403964224)))];
+            tensor<fp16, [1024]> model_layers_16_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403929344)))];
+            tensor<fp16, [1024]> model_layers_16_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403931456)))];
             tensor<fp16, [1, 126, 1024]> query_33_cast_fp16 = layer_norm(axes = query_33_axes_0, beta = model_layers_16_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_16_norm_self_att_weight_to_fp16, x = input_863_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403966336))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405014976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403933568))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404982208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_147_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_16_self_attn_linear_q_weight_to_fp16_quantized, x = query_33_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
             tensor<int32, [4]> var_2852 = const()[name = tensor<string, []>("op_2852"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_97_cast_fp16 = reshape(shape = var_2852, x = linear_147_cast_fp16)[name = tensor<string, []>("q_97_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405017088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406065728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404984320))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406032960))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_148_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_16_self_attn_linear_k_weight_to_fp16_quantized, x = query_33_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
             tensor<int32, [4]> var_2856 = const()[name = tensor<string, []>("op_2856"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_65_cast_fp16 = reshape(shape = var_2856, x = linear_148_cast_fp16)[name = tensor<string, []>("k_65_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406067840))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407116480))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406035072))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407083712))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_149_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_16_self_attn_linear_v_weight_to_fp16_quantized, x = query_33_cast_fp16)[name = tensor<string, []>("linear_149_cast_fp16")];
             tensor<int32, [4]> var_2860 = const()[name = tensor<string, []>("op_2860"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_33_cast_fp16 = reshape(shape = var_2860, x = linear_149_cast_fp16)[name = tensor<string, []>("v_33_cast_fp16")];
             tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<string, []>("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_16_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_16_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407118592)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2872_cast_fp16 = add(x = q_97_cast_fp16, y = model_layers_16_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_2872_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_16_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_16_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407120704)))];
-            tensor<fp16, [1, 126, 8, 128]> var_2874_cast_fp16 = add(x = q_97_cast_fp16, y = model_layers_16_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_2874_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_16_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407085824))), scale = tensor<fp16, [8]>([0x1.74cp-7, 0x1.5b4p-7, 0x1.0e8p-8, 0x1.f6p-8, 0x1.3a4p-7, 0x1.868p-7, 0x1.04p-7, 0x1.644p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2872_cast_fp16 = add(x = q_97_cast_fp16, y = model_layers_16_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_2872_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_16_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407086912))), scale = tensor<fp16, [8]>([0x1.184p-8, 0x1.8c8p-9, 0x1.5d8p-8, 0x1.2ccp-7, 0x1.ffcp-8, 0x1.384p-7, 0x1.6acp-8, 0x1.35cp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_2874_cast_fp16 = add(x = q_97_cast_fp16, y = model_layers_16_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_2874_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_359_transpose_x_0 = const()[name = tensor<string, []>("x_359_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_359_transpose_y_0 = const()[name = tensor<string, []>("x_359_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_2876_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2876_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407122816))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407379904))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_2876_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2876_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407088000))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407345088))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_33_cast_fp16 = transpose(perm = q_with_bias_v_33_perm_0, x = var_2874_cast_fp16)[name = tensor<string, []>("transpose_199")];
             tensor<fp16, [1, 8, 126, 251]> x_359_cast_fp16 = matmul(transpose_x = x_359_transpose_x_0, transpose_y = x_359_transpose_y_0, x = q_with_bias_v_33_cast_fp16, y = op_2876_to_fp16_quantized)[name = tensor<string, []>("x_359_cast_fp16")];
             tensor<int32, [8]> x_361_pad_0 = const()[name = tensor<string, []>("x_361_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -2376,12 +2376,12 @@ program(1.0)
             tensor<int32, [3]> var_2909 = const()[name = tensor<string, []>("op_2909"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_2908_cast_fp16 = transpose(perm = var_2908_perm_0, x = x_365_cast_fp16)[name = tensor<string, []>("transpose_195")];
             tensor<fp16, [1, 126, 1024]> input_867_cast_fp16 = reshape(shape = var_2909, x = var_2908_cast_fp16)[name = tensor<string, []>("input_867_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407380480))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408429120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407345664))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408394304))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_151_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_16_self_attn_linear_out_weight_to_fp16_quantized, x = input_867_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_871_cast_fp16 = add(x = input_863_cast_fp16, y = linear_151_cast_fp16)[name = tensor<string, []>("input_871_cast_fp16")];
             tensor<int32, [1]> x_369_axes_0 = const()[name = tensor<string, []>("x_369_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408431232)))];
-            tensor<fp16, [1024]> model_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408433344)))];
+            tensor<fp16, [1024]> model_layers_16_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408396416)))];
+            tensor<fp16, [1024]> model_layers_16_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408398528)))];
             tensor<fp16, [1, 126, 1024]> x_369_cast_fp16 = layer_norm(axes = x_369_axes_0, beta = model_layers_16_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_16_norm_conv_weight_to_fp16, x = input_871_cast_fp16)[name = tensor<string, []>("x_369_cast_fp16")];
             tensor<int32, [3]> input_873_perm_0 = const()[name = tensor<string, []>("input_873_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_875_pad_type_0 = const()[name = tensor<string, []>("input_875_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2389,7 +2389,7 @@ program(1.0)
             tensor<int32, [2]> input_875_pad_0 = const()[name = tensor<string, []>("input_875_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_875_dilations_0 = const()[name = tensor<string, []>("input_875_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_875_groups_0 = const()[name = tensor<string, []>("input_875_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408435456))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410532672))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408400640))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410497856))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_873_cast_fp16 = transpose(perm = input_873_perm_0, x = x_369_cast_fp16)[name = tensor<string, []>("transpose_194")];
             tensor<fp16, [1, 2048, 126]> input_875_cast_fp16 = conv(dilations = input_875_dilations_0, groups = input_875_groups_0, pad = input_875_pad_0, pad_type = input_875_pad_type_0, strides = input_875_strides_0, weight = model_layers_16_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_873_cast_fp16)[name = tensor<string, []>("input_875_cast_fp16")];
             tensor<int32, []> x_371_split_num_splits_0 = const()[name = tensor<string, []>("x_371_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -2407,8 +2407,8 @@ program(1.0)
             tensor<int32, [1]> input_881_strides_0 = const()[name = tensor<string, []>("input_881_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_881_pad_0 = const()[name = tensor<string, []>("input_881_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_881_dilations_0 = const()[name = tensor<string, []>("input_881_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_280_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_280_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410536832))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410546112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_281_to_fp16 = const()[name = tensor<string, []>("const_281_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410548224)))];
+            tensor<fp16, [1024, 1, 9]> const_280_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_280_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410502016))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410511296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_281_to_fp16 = const()[name = tensor<string, []>("const_281_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410513408)))];
             tensor<fp16, [1, 1024, 126]> input_883_cast_fp16 = conv(bias = const_281_to_fp16, dilations = input_881_dilations_0, groups = input_881_groups_0, pad = input_881_pad_0, pad_type = input_881_pad_type_0, strides = input_881_strides_0, weight = const_280_to_fp16_quantized, x = input_879_cast_fp16)[name = tensor<string, []>("input_883_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_885_cast_fp16 = silu(x = input_883_cast_fp16)[name = tensor<string, []>("input_885_cast_fp16")];
             tensor<string, []> x_373_pad_type_0 = const()[name = tensor<string, []>("x_373_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2416,64 +2416,64 @@ program(1.0)
             tensor<int32, [2]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_373_dilations_0 = const()[name = tensor<string, []>("x_373_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_373_groups_0 = const()[name = tensor<string, []>("x_373_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410550336))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411598976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(410515520))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411564160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_373_cast_fp16 = conv(dilations = x_373_dilations_0, groups = x_373_groups_0, pad = x_373_pad_0, pad_type = x_373_pad_type_0, strides = x_373_strides_0, weight = model_layers_16_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_885_cast_fp16)[name = tensor<string, []>("x_373_cast_fp16")];
             tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_887_cast_fp16 = transpose(perm = input_887_perm_0, x = x_373_cast_fp16)[name = tensor<string, []>("transpose_193")];
             tensor<fp16, [1, 126, 1024]> input_889_cast_fp16 = add(x = input_871_cast_fp16, y = input_887_cast_fp16)[name = tensor<string, []>("input_889_cast_fp16")];
             tensor<int32, [1]> input_891_axes_0 = const()[name = tensor<string, []>("input_891_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411601088)))];
-            tensor<fp16, [1024]> model_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411603200)))];
+            tensor<fp16, [1024]> model_layers_16_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411566272)))];
+            tensor<fp16, [1024]> model_layers_16_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411568384)))];
             tensor<fp16, [1, 126, 1024]> input_891_cast_fp16 = layer_norm(axes = input_891_axes_0, beta = model_layers_16_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_16_norm_feed_forward2_weight_to_fp16, x = input_889_cast_fp16)[name = tensor<string, []>("input_891_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_16_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411605312))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415799680))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_16_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411570496))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415764864))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_152_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_feed_forward2_linear1_weight_to_fp16_quantized, x = input_891_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_895_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor<string, []>("input_895_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_16_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415807936))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420002304))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_16_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_16_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415773120))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419967488))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_153_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_16_feed_forward2_linear2_weight_to_fp16_quantized, x = input_895_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
             tensor<fp16, []> var_2969_to_fp16 = const()[name = tensor<string, []>("op_2969_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2970_cast_fp16 = mul(x = linear_153_cast_fp16, y = var_2969_to_fp16)[name = tensor<string, []>("op_2970_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_901_cast_fp16 = add(x = input_889_cast_fp16, y = var_2970_cast_fp16)[name = tensor<string, []>("input_901_cast_fp16")];
             tensor<int32, [1]> input_903_axes_0 = const()[name = tensor<string, []>("input_903_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420004416)))];
-            tensor<fp16, [1024]> model_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420006528)))];
+            tensor<fp16, [1024]> model_layers_16_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419969600)))];
+            tensor<fp16, [1024]> model_layers_16_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_16_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419971712)))];
             tensor<fp16, [1, 126, 1024]> input_903_cast_fp16 = layer_norm(axes = input_903_axes_0, beta = model_layers_16_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_16_norm_out_weight_to_fp16, x = input_901_cast_fp16)[name = tensor<string, []>("input_903_cast_fp16")];
             tensor<int32, [1]> input_905_axes_0 = const()[name = tensor<string, []>("input_905_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420008640)))];
-            tensor<fp16, [1024]> model_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420010752)))];
+            tensor<fp16, [1024]> model_layers_17_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419973824)))];
+            tensor<fp16, [1024]> model_layers_17_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419975936)))];
             tensor<fp16, [1, 126, 1024]> input_905_cast_fp16 = layer_norm(axes = input_905_axes_0, beta = model_layers_17_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_17_norm_feed_forward1_weight_to_fp16, x = input_903_cast_fp16)[name = tensor<string, []>("input_905_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_17_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420012864))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424207232))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_17_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(419978048))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424172416))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_154_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_feed_forward1_linear1_weight_to_fp16_quantized, x = input_905_cast_fp16)[name = tensor<string, []>("linear_154_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_909_cast_fp16 = silu(x = linear_154_cast_fp16)[name = tensor<string, []>("input_909_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_17_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424215488))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428409856))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_17_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424180672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428375040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_155_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_17_feed_forward1_linear2_weight_to_fp16_quantized, x = input_909_cast_fp16)[name = tensor<string, []>("linear_155_cast_fp16")];
             tensor<fp16, []> var_2998_to_fp16 = const()[name = tensor<string, []>("op_2998_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_2999_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2998_to_fp16)[name = tensor<string, []>("op_2999_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_915_cast_fp16 = add(x = input_903_cast_fp16, y = var_2999_cast_fp16)[name = tensor<string, []>("input_915_cast_fp16")];
             tensor<int32, [1]> query_35_axes_0 = const()[name = tensor<string, []>("query_35_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_17_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428411968)))];
-            tensor<fp16, [1024]> model_layers_17_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428414080)))];
+            tensor<fp16, [1024]> model_layers_17_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428377152)))];
+            tensor<fp16, [1024]> model_layers_17_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428379264)))];
             tensor<fp16, [1, 126, 1024]> query_35_cast_fp16 = layer_norm(axes = query_35_axes_0, beta = model_layers_17_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_17_norm_self_att_weight_to_fp16, x = input_915_cast_fp16)[name = tensor<string, []>("query_35_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428416192))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429464832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428381376))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429430016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_156_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_17_self_attn_linear_q_weight_to_fp16_quantized, x = query_35_cast_fp16)[name = tensor<string, []>("linear_156_cast_fp16")];
             tensor<int32, [4]> var_3015 = const()[name = tensor<string, []>("op_3015"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_103_cast_fp16 = reshape(shape = var_3015, x = linear_156_cast_fp16)[name = tensor<string, []>("q_103_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429466944))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430515584))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429432128))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430480768))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_157_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_17_self_attn_linear_k_weight_to_fp16_quantized, x = query_35_cast_fp16)[name = tensor<string, []>("linear_157_cast_fp16")];
             tensor<int32, [4]> var_3019 = const()[name = tensor<string, []>("op_3019"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_69_cast_fp16 = reshape(shape = var_3019, x = linear_157_cast_fp16)[name = tensor<string, []>("k_69_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430517696))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431566336))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430482880))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431531520))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_158_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_17_self_attn_linear_v_weight_to_fp16_quantized, x = query_35_cast_fp16)[name = tensor<string, []>("linear_158_cast_fp16")];
             tensor<int32, [4]> var_3023 = const()[name = tensor<string, []>("op_3023"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_35_cast_fp16 = reshape(shape = var_3023, x = linear_158_cast_fp16)[name = tensor<string, []>("v_35_cast_fp16")];
             tensor<int32, [4]> value_35_perm_0 = const()[name = tensor<string, []>("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_17_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_17_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431568448)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3035_cast_fp16 = add(x = q_103_cast_fp16, y = model_layers_17_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3035_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_17_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_17_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431570560)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3037_cast_fp16 = add(x = q_103_cast_fp16, y = model_layers_17_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3037_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_17_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431533632))), scale = tensor<fp16, [8]>([0x1.cep-8, 0x1.444p-8, 0x1.424p-8, 0x1.54p-9, 0x1.4a8p-8, 0x1.6e4p-8, 0x1.034p-8, 0x1.62p-9]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3035_cast_fp16 = add(x = q_103_cast_fp16, y = model_layers_17_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_3035_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_17_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431534720))), scale = tensor<fp16, [8]>([0x1.448p-7, 0x1.37cp-7, 0x1.4f8p-8, 0x1.2b4p-7, 0x1.644p-7, 0x1.b48p-10, 0x1.898p-9, 0x1.864p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3037_cast_fp16 = add(x = q_103_cast_fp16, y = model_layers_17_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_3037_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_381_transpose_x_0 = const()[name = tensor<string, []>("x_381_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_381_transpose_y_0 = const()[name = tensor<string, []>("x_381_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_3039_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3039_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431572672))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431829760))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_3039_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3039_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431535808))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431792896))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_35_cast_fp16 = transpose(perm = q_with_bias_v_35_perm_0, x = var_3037_cast_fp16)[name = tensor<string, []>("transpose_192")];
             tensor<fp16, [1, 8, 126, 251]> x_381_cast_fp16 = matmul(transpose_x = x_381_transpose_x_0, transpose_y = x_381_transpose_y_0, x = q_with_bias_v_35_cast_fp16, y = op_3039_to_fp16_quantized)[name = tensor<string, []>("x_381_cast_fp16")];
             tensor<int32, [8]> x_383_pad_0 = const()[name = tensor<string, []>("x_383_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -2513,12 +2513,12 @@ program(1.0)
             tensor<int32, [3]> var_3072 = const()[name = tensor<string, []>("op_3072"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_3071_cast_fp16 = transpose(perm = var_3071_perm_0, x = x_387_cast_fp16)[name = tensor<string, []>("transpose_188")];
             tensor<fp16, [1, 126, 1024]> input_919_cast_fp16 = reshape(shape = var_3072, x = var_3071_cast_fp16)[name = tensor<string, []>("input_919_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431830336))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432878976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431793472))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432842112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_160_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_17_self_attn_linear_out_weight_to_fp16_quantized, x = input_919_cast_fp16)[name = tensor<string, []>("linear_160_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_923_cast_fp16 = add(x = input_915_cast_fp16, y = linear_160_cast_fp16)[name = tensor<string, []>("input_923_cast_fp16")];
             tensor<int32, [1]> x_391_axes_0 = const()[name = tensor<string, []>("x_391_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_17_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432881088)))];
-            tensor<fp16, [1024]> model_layers_17_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432883200)))];
+            tensor<fp16, [1024]> model_layers_17_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432844224)))];
+            tensor<fp16, [1024]> model_layers_17_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432846336)))];
             tensor<fp16, [1, 126, 1024]> x_391_cast_fp16 = layer_norm(axes = x_391_axes_0, beta = model_layers_17_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_17_norm_conv_weight_to_fp16, x = input_923_cast_fp16)[name = tensor<string, []>("x_391_cast_fp16")];
             tensor<int32, [3]> input_925_perm_0 = const()[name = tensor<string, []>("input_925_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_927_pad_type_0 = const()[name = tensor<string, []>("input_927_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2526,7 +2526,7 @@ program(1.0)
             tensor<int32, [2]> input_927_pad_0 = const()[name = tensor<string, []>("input_927_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_927_dilations_0 = const()[name = tensor<string, []>("input_927_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_927_groups_0 = const()[name = tensor<string, []>("input_927_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432885312))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434982528))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432848448))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434945664))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_925_cast_fp16 = transpose(perm = input_925_perm_0, x = x_391_cast_fp16)[name = tensor<string, []>("transpose_187")];
             tensor<fp16, [1, 2048, 126]> input_927_cast_fp16 = conv(dilations = input_927_dilations_0, groups = input_927_groups_0, pad = input_927_pad_0, pad_type = input_927_pad_type_0, strides = input_927_strides_0, weight = model_layers_17_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_925_cast_fp16)[name = tensor<string, []>("input_927_cast_fp16")];
             tensor<int32, []> x_393_split_num_splits_0 = const()[name = tensor<string, []>("x_393_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -2544,8 +2544,8 @@ program(1.0)
             tensor<int32, [1]> input_933_strides_0 = const()[name = tensor<string, []>("input_933_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_933_pad_0 = const()[name = tensor<string, []>("input_933_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_933_dilations_0 = const()[name = tensor<string, []>("input_933_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_282_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_282_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434986688))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434995968))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_283_to_fp16 = const()[name = tensor<string, []>("const_283_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434998080)))];
+            tensor<fp16, [1024, 1, 9]> const_282_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_282_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434949824))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434959104))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_283_to_fp16 = const()[name = tensor<string, []>("const_283_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434961216)))];
             tensor<fp16, [1, 1024, 126]> input_935_cast_fp16 = conv(bias = const_283_to_fp16, dilations = input_933_dilations_0, groups = input_933_groups_0, pad = input_933_pad_0, pad_type = input_933_pad_type_0, strides = input_933_strides_0, weight = const_282_to_fp16_quantized, x = input_931_cast_fp16)[name = tensor<string, []>("input_935_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_937_cast_fp16 = silu(x = input_935_cast_fp16)[name = tensor<string, []>("input_937_cast_fp16")];
             tensor<string, []> x_395_pad_type_0 = const()[name = tensor<string, []>("x_395_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2553,64 +2553,64 @@ program(1.0)
             tensor<int32, [2]> x_395_pad_0 = const()[name = tensor<string, []>("x_395_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_395_dilations_0 = const()[name = tensor<string, []>("x_395_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_395_groups_0 = const()[name = tensor<string, []>("x_395_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(435000192))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436048832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434963328))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436011968))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_395_cast_fp16 = conv(dilations = x_395_dilations_0, groups = x_395_groups_0, pad = x_395_pad_0, pad_type = x_395_pad_type_0, strides = x_395_strides_0, weight = model_layers_17_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_937_cast_fp16)[name = tensor<string, []>("x_395_cast_fp16")];
             tensor<int32, [3]> input_939_perm_0 = const()[name = tensor<string, []>("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_939_cast_fp16 = transpose(perm = input_939_perm_0, x = x_395_cast_fp16)[name = tensor<string, []>("transpose_186")];
             tensor<fp16, [1, 126, 1024]> input_941_cast_fp16 = add(x = input_923_cast_fp16, y = input_939_cast_fp16)[name = tensor<string, []>("input_941_cast_fp16")];
             tensor<int32, [1]> input_943_axes_0 = const()[name = tensor<string, []>("input_943_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436050944)))];
-            tensor<fp16, [1024]> model_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436053056)))];
+            tensor<fp16, [1024]> model_layers_17_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436014080)))];
+            tensor<fp16, [1024]> model_layers_17_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436016192)))];
             tensor<fp16, [1, 126, 1024]> input_943_cast_fp16 = layer_norm(axes = input_943_axes_0, beta = model_layers_17_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_17_norm_feed_forward2_weight_to_fp16, x = input_941_cast_fp16)[name = tensor<string, []>("input_943_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_17_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436055168))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440249536))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_17_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436018304))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440212672))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_161_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_feed_forward2_linear1_weight_to_fp16_quantized, x = input_943_cast_fp16)[name = tensor<string, []>("linear_161_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_947_cast_fp16 = silu(x = linear_161_cast_fp16)[name = tensor<string, []>("input_947_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_17_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440257792))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444452160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_17_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_17_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440220928))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444415296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_162_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_17_feed_forward2_linear2_weight_to_fp16_quantized, x = input_947_cast_fp16)[name = tensor<string, []>("linear_162_cast_fp16")];
             tensor<fp16, []> var_3132_to_fp16 = const()[name = tensor<string, []>("op_3132_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3133_cast_fp16 = mul(x = linear_162_cast_fp16, y = var_3132_to_fp16)[name = tensor<string, []>("op_3133_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_953_cast_fp16 = add(x = input_941_cast_fp16, y = var_3133_cast_fp16)[name = tensor<string, []>("input_953_cast_fp16")];
             tensor<int32, [1]> input_955_axes_0 = const()[name = tensor<string, []>("input_955_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_17_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444454272)))];
-            tensor<fp16, [1024]> model_layers_17_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444456384)))];
+            tensor<fp16, [1024]> model_layers_17_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444417408)))];
+            tensor<fp16, [1024]> model_layers_17_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_17_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444419520)))];
             tensor<fp16, [1, 126, 1024]> input_955_cast_fp16 = layer_norm(axes = input_955_axes_0, beta = model_layers_17_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_17_norm_out_weight_to_fp16, x = input_953_cast_fp16)[name = tensor<string, []>("input_955_cast_fp16")];
             tensor<int32, [1]> input_957_axes_0 = const()[name = tensor<string, []>("input_957_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444458496)))];
-            tensor<fp16, [1024]> model_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444460608)))];
+            tensor<fp16, [1024]> model_layers_18_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444421632)))];
+            tensor<fp16, [1024]> model_layers_18_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444423744)))];
             tensor<fp16, [1, 126, 1024]> input_957_cast_fp16 = layer_norm(axes = input_957_axes_0, beta = model_layers_18_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_18_norm_feed_forward1_weight_to_fp16, x = input_955_cast_fp16)[name = tensor<string, []>("input_957_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_18_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444462720))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448657088))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_18_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(444425856))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448620224))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_feed_forward1_linear1_weight_to_fp16_quantized, x = input_957_cast_fp16)[name = tensor<string, []>("linear_163_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_961_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor<string, []>("input_961_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_18_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448665344))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452859712))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_18_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448628480))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452822848))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_164_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_18_feed_forward1_linear2_weight_to_fp16_quantized, x = input_961_cast_fp16)[name = tensor<string, []>("linear_164_cast_fp16")];
             tensor<fp16, []> var_3161_to_fp16 = const()[name = tensor<string, []>("op_3161_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3162_cast_fp16 = mul(x = linear_164_cast_fp16, y = var_3161_to_fp16)[name = tensor<string, []>("op_3162_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_967_cast_fp16 = add(x = input_955_cast_fp16, y = var_3162_cast_fp16)[name = tensor<string, []>("input_967_cast_fp16")];
             tensor<int32, [1]> query_37_axes_0 = const()[name = tensor<string, []>("query_37_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_18_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452861824)))];
-            tensor<fp16, [1024]> model_layers_18_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452863936)))];
+            tensor<fp16, [1024]> model_layers_18_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452824960)))];
+            tensor<fp16, [1024]> model_layers_18_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452827072)))];
             tensor<fp16, [1, 126, 1024]> query_37_cast_fp16 = layer_norm(axes = query_37_axes_0, beta = model_layers_18_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_18_norm_self_att_weight_to_fp16, x = input_967_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452866048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453914688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452829184))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453877824))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_165_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_18_self_attn_linear_q_weight_to_fp16_quantized, x = query_37_cast_fp16)[name = tensor<string, []>("linear_165_cast_fp16")];
             tensor<int32, [4]> var_3178 = const()[name = tensor<string, []>("op_3178"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_109_cast_fp16 = reshape(shape = var_3178, x = linear_165_cast_fp16)[name = tensor<string, []>("q_109_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453916800))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454965440))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453879936))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454928576))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_166_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_18_self_attn_linear_k_weight_to_fp16_quantized, x = query_37_cast_fp16)[name = tensor<string, []>("linear_166_cast_fp16")];
             tensor<int32, [4]> var_3182 = const()[name = tensor<string, []>("op_3182"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_73_cast_fp16 = reshape(shape = var_3182, x = linear_166_cast_fp16)[name = tensor<string, []>("k_73_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454967552))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456016192))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454930688))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455979328))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_167_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_18_self_attn_linear_v_weight_to_fp16_quantized, x = query_37_cast_fp16)[name = tensor<string, []>("linear_167_cast_fp16")];
             tensor<int32, [4]> var_3186 = const()[name = tensor<string, []>("op_3186"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_37_cast_fp16 = reshape(shape = var_3186, x = linear_167_cast_fp16)[name = tensor<string, []>("v_37_cast_fp16")];
             tensor<int32, [4]> value_37_perm_0 = const()[name = tensor<string, []>("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_18_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_18_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456018304)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3198_cast_fp16 = add(x = q_109_cast_fp16, y = model_layers_18_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3198_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_18_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_18_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456020416)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3200_cast_fp16 = add(x = q_109_cast_fp16, y = model_layers_18_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3200_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_18_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455981440))), scale = tensor<fp16, [8]>([0x1.8ep-8, 0x1.86p-9, 0x1.d8cp-9, 0x1.808p-8, 0x1.76cp-8, 0x1.eb4p-9, 0x1.76p-8, 0x1.2bp-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3198_cast_fp16 = add(x = q_109_cast_fp16, y = model_layers_18_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_3198_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_18_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455982528))), scale = tensor<fp16, [8]>([0x1.a78p-9, 0x1.a44p-9, 0x1.3a8p-8, 0x1.194p-8, 0x1.7cp-9, 0x1.31p-7, 0x1.464p-7, 0x1.404p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3200_cast_fp16 = add(x = q_109_cast_fp16, y = model_layers_18_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_3200_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_3202_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3202_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456022528))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456279616))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_3202_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3202_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455983616))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456240704))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_37_cast_fp16 = transpose(perm = q_with_bias_v_37_perm_0, x = var_3200_cast_fp16)[name = tensor<string, []>("transpose_185")];
             tensor<fp16, [1, 8, 126, 251]> x_403_cast_fp16 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = q_with_bias_v_37_cast_fp16, y = op_3202_to_fp16_quantized)[name = tensor<string, []>("x_403_cast_fp16")];
             tensor<int32, [8]> x_405_pad_0 = const()[name = tensor<string, []>("x_405_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -2650,12 +2650,12 @@ program(1.0)
             tensor<int32, [3]> var_3235 = const()[name = tensor<string, []>("op_3235"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_3234_cast_fp16 = transpose(perm = var_3234_perm_0, x = x_409_cast_fp16)[name = tensor<string, []>("transpose_181")];
             tensor<fp16, [1, 126, 1024]> input_971_cast_fp16 = reshape(shape = var_3235, x = var_3234_cast_fp16)[name = tensor<string, []>("input_971_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456280192))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457328832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456241280))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457289920))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_169_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_18_self_attn_linear_out_weight_to_fp16_quantized, x = input_971_cast_fp16)[name = tensor<string, []>("linear_169_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_975_cast_fp16 = add(x = input_967_cast_fp16, y = linear_169_cast_fp16)[name = tensor<string, []>("input_975_cast_fp16")];
             tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_18_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457330944)))];
-            tensor<fp16, [1024]> model_layers_18_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457333056)))];
+            tensor<fp16, [1024]> model_layers_18_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457292032)))];
+            tensor<fp16, [1024]> model_layers_18_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457294144)))];
             tensor<fp16, [1, 126, 1024]> x_413_cast_fp16 = layer_norm(axes = x_413_axes_0, beta = model_layers_18_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_18_norm_conv_weight_to_fp16, x = input_975_cast_fp16)[name = tensor<string, []>("x_413_cast_fp16")];
             tensor<int32, [3]> input_977_perm_0 = const()[name = tensor<string, []>("input_977_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_979_pad_type_0 = const()[name = tensor<string, []>("input_979_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2663,7 +2663,7 @@ program(1.0)
             tensor<int32, [2]> input_979_pad_0 = const()[name = tensor<string, []>("input_979_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_979_dilations_0 = const()[name = tensor<string, []>("input_979_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_979_groups_0 = const()[name = tensor<string, []>("input_979_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457335168))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459432384))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457296256))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459393472))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_977_cast_fp16 = transpose(perm = input_977_perm_0, x = x_413_cast_fp16)[name = tensor<string, []>("transpose_180")];
             tensor<fp16, [1, 2048, 126]> input_979_cast_fp16 = conv(dilations = input_979_dilations_0, groups = input_979_groups_0, pad = input_979_pad_0, pad_type = input_979_pad_type_0, strides = input_979_strides_0, weight = model_layers_18_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_977_cast_fp16)[name = tensor<string, []>("input_979_cast_fp16")];
             tensor<int32, []> x_415_split_num_splits_0 = const()[name = tensor<string, []>("x_415_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -2681,8 +2681,8 @@ program(1.0)
             tensor<int32, [1]> input_985_strides_0 = const()[name = tensor<string, []>("input_985_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_985_pad_0 = const()[name = tensor<string, []>("input_985_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_985_dilations_0 = const()[name = tensor<string, []>("input_985_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_284_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_284_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459436544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459445824))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_285_to_fp16 = const()[name = tensor<string, []>("const_285_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459447936)))];
+            tensor<fp16, [1024, 1, 9]> const_284_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_284_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459397632))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459406912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_285_to_fp16 = const()[name = tensor<string, []>("const_285_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459409024)))];
             tensor<fp16, [1, 1024, 126]> input_987_cast_fp16 = conv(bias = const_285_to_fp16, dilations = input_985_dilations_0, groups = input_985_groups_0, pad = input_985_pad_0, pad_type = input_985_pad_type_0, strides = input_985_strides_0, weight = const_284_to_fp16_quantized, x = input_983_cast_fp16)[name = tensor<string, []>("input_987_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_989_cast_fp16 = silu(x = input_987_cast_fp16)[name = tensor<string, []>("input_989_cast_fp16")];
             tensor<string, []> x_417_pad_type_0 = const()[name = tensor<string, []>("x_417_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2690,64 +2690,64 @@ program(1.0)
             tensor<int32, [2]> x_417_pad_0 = const()[name = tensor<string, []>("x_417_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_417_dilations_0 = const()[name = tensor<string, []>("x_417_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_417_groups_0 = const()[name = tensor<string, []>("x_417_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459450048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460498688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459411136))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460459776))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_417_cast_fp16 = conv(dilations = x_417_dilations_0, groups = x_417_groups_0, pad = x_417_pad_0, pad_type = x_417_pad_type_0, strides = x_417_strides_0, weight = model_layers_18_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_989_cast_fp16)[name = tensor<string, []>("x_417_cast_fp16")];
             tensor<int32, [3]> input_991_perm_0 = const()[name = tensor<string, []>("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_991_cast_fp16 = transpose(perm = input_991_perm_0, x = x_417_cast_fp16)[name = tensor<string, []>("transpose_179")];
             tensor<fp16, [1, 126, 1024]> input_993_cast_fp16 = add(x = input_975_cast_fp16, y = input_991_cast_fp16)[name = tensor<string, []>("input_993_cast_fp16")];
             tensor<int32, [1]> input_995_axes_0 = const()[name = tensor<string, []>("input_995_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460500800)))];
-            tensor<fp16, [1024]> model_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460502912)))];
+            tensor<fp16, [1024]> model_layers_18_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460461888)))];
+            tensor<fp16, [1024]> model_layers_18_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460464000)))];
             tensor<fp16, [1, 126, 1024]> input_995_cast_fp16 = layer_norm(axes = input_995_axes_0, beta = model_layers_18_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_18_norm_feed_forward2_weight_to_fp16, x = input_993_cast_fp16)[name = tensor<string, []>("input_995_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460505024))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464699392))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_18_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460466112))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464660480))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_170_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_feed_forward2_linear1_weight_to_fp16_quantized, x = input_995_cast_fp16)[name = tensor<string, []>("linear_170_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_999_cast_fp16 = silu(x = linear_170_cast_fp16)[name = tensor<string, []>("input_999_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464707648))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468902016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_18_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_18_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464668736))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468863104))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_171_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_18_feed_forward2_linear2_weight_to_fp16_quantized, x = input_999_cast_fp16)[name = tensor<string, []>("linear_171_cast_fp16")];
             tensor<fp16, []> var_3295_to_fp16 = const()[name = tensor<string, []>("op_3295_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3296_cast_fp16 = mul(x = linear_171_cast_fp16, y = var_3295_to_fp16)[name = tensor<string, []>("op_3296_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1005_cast_fp16 = add(x = input_993_cast_fp16, y = var_3296_cast_fp16)[name = tensor<string, []>("input_1005_cast_fp16")];
             tensor<int32, [1]> input_1007_axes_0 = const()[name = tensor<string, []>("input_1007_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_18_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468904128)))];
-            tensor<fp16, [1024]> model_layers_18_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468906240)))];
+            tensor<fp16, [1024]> model_layers_18_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468865216)))];
+            tensor<fp16, [1024]> model_layers_18_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_18_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468867328)))];
             tensor<fp16, [1, 126, 1024]> input_1007_cast_fp16 = layer_norm(axes = input_1007_axes_0, beta = model_layers_18_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_18_norm_out_weight_to_fp16, x = input_1005_cast_fp16)[name = tensor<string, []>("input_1007_cast_fp16")];
             tensor<int32, [1]> input_1009_axes_0 = const()[name = tensor<string, []>("input_1009_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468908352)))];
-            tensor<fp16, [1024]> model_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468910464)))];
+            tensor<fp16, [1024]> model_layers_19_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468869440)))];
+            tensor<fp16, [1024]> model_layers_19_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468871552)))];
             tensor<fp16, [1, 126, 1024]> input_1009_cast_fp16 = layer_norm(axes = input_1009_axes_0, beta = model_layers_19_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_19_norm_feed_forward1_weight_to_fp16, x = input_1007_cast_fp16)[name = tensor<string, []>("input_1009_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468912576))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473106944))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_19_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468873664))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473068032))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_172_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1009_cast_fp16)[name = tensor<string, []>("linear_172_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1013_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor<string, []>("input_1013_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473115200))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477309568))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_19_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(473076288))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477270656))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_173_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_19_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1013_cast_fp16)[name = tensor<string, []>("linear_173_cast_fp16")];
             tensor<fp16, []> var_3324_to_fp16 = const()[name = tensor<string, []>("op_3324_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3325_cast_fp16 = mul(x = linear_173_cast_fp16, y = var_3324_to_fp16)[name = tensor<string, []>("op_3325_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1019_cast_fp16 = add(x = input_1007_cast_fp16, y = var_3325_cast_fp16)[name = tensor<string, []>("input_1019_cast_fp16")];
             tensor<int32, [1]> query_39_axes_0 = const()[name = tensor<string, []>("query_39_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_19_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477311680)))];
-            tensor<fp16, [1024]> model_layers_19_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477313792)))];
+            tensor<fp16, [1024]> model_layers_19_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477272768)))];
+            tensor<fp16, [1024]> model_layers_19_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477274880)))];
             tensor<fp16, [1, 126, 1024]> query_39_cast_fp16 = layer_norm(axes = query_39_axes_0, beta = model_layers_19_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_19_norm_self_att_weight_to_fp16, x = input_1019_cast_fp16)[name = tensor<string, []>("query_39_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477315904))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478364544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477276992))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478325632))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_174_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_19_self_attn_linear_q_weight_to_fp16_quantized, x = query_39_cast_fp16)[name = tensor<string, []>("linear_174_cast_fp16")];
             tensor<int32, [4]> var_3341 = const()[name = tensor<string, []>("op_3341"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_115_cast_fp16 = reshape(shape = var_3341, x = linear_174_cast_fp16)[name = tensor<string, []>("q_115_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478366656))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(479415296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478327744))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(479376384))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_175_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_19_self_attn_linear_k_weight_to_fp16_quantized, x = query_39_cast_fp16)[name = tensor<string, []>("linear_175_cast_fp16")];
             tensor<int32, [4]> var_3345 = const()[name = tensor<string, []>("op_3345"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_77_cast_fp16 = reshape(shape = var_3345, x = linear_175_cast_fp16)[name = tensor<string, []>("k_77_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(479417408))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480466048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(479378496))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480427136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_176_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_19_self_attn_linear_v_weight_to_fp16_quantized, x = query_39_cast_fp16)[name = tensor<string, []>("linear_176_cast_fp16")];
             tensor<int32, [4]> var_3349 = const()[name = tensor<string, []>("op_3349"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_39_cast_fp16 = reshape(shape = var_3349, x = linear_176_cast_fp16)[name = tensor<string, []>("v_39_cast_fp16")];
             tensor<int32, [4]> value_39_perm_0 = const()[name = tensor<string, []>("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_19_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_19_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480468160)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3361_cast_fp16 = add(x = q_115_cast_fp16, y = model_layers_19_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3361_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_19_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_19_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480470272)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3363_cast_fp16 = add(x = q_115_cast_fp16, y = model_layers_19_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3363_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_19_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480429248))), scale = tensor<fp16, [8]>([0x1.0f8p-7, 0x1.a58p-9, 0x1.acp-9, 0x1.9c8p-8, 0x1.1dp-8, 0x1.f9cp-9, 0x1.bbp-8, 0x1.954p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3361_cast_fp16 = add(x = q_115_cast_fp16, y = model_layers_19_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_3361_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_19_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480430336))), scale = tensor<fp16, [8]>([0x1.754p-9, 0x1.2f8p-7, 0x1.3c8p-7, 0x1.074p-7, 0x1.3ap-7, 0x1.8b4p-8, 0x1.d74p-8, 0x1.ec8p-9]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3363_cast_fp16 = add(x = q_115_cast_fp16, y = model_layers_19_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_3363_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_425_transpose_x_0 = const()[name = tensor<string, []>("x_425_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_425_transpose_y_0 = const()[name = tensor<string, []>("x_425_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_3365_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3365_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480472384))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480729472))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_3365_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3365_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480431424))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480688512))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_39_cast_fp16 = transpose(perm = q_with_bias_v_39_perm_0, x = var_3363_cast_fp16)[name = tensor<string, []>("transpose_178")];
             tensor<fp16, [1, 8, 126, 251]> x_425_cast_fp16 = matmul(transpose_x = x_425_transpose_x_0, transpose_y = x_425_transpose_y_0, x = q_with_bias_v_39_cast_fp16, y = op_3365_to_fp16_quantized)[name = tensor<string, []>("x_425_cast_fp16")];
             tensor<int32, [8]> x_427_pad_0 = const()[name = tensor<string, []>("x_427_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -2787,12 +2787,12 @@ program(1.0)
             tensor<int32, [3]> var_3398 = const()[name = tensor<string, []>("op_3398"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_3397_cast_fp16 = transpose(perm = var_3397_perm_0, x = x_431_cast_fp16)[name = tensor<string, []>("transpose_174")];
             tensor<fp16, [1, 126, 1024]> input_1023_cast_fp16 = reshape(shape = var_3398, x = var_3397_cast_fp16)[name = tensor<string, []>("input_1023_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480730048))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481778688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480689088))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481737728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_178_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_19_self_attn_linear_out_weight_to_fp16_quantized, x = input_1023_cast_fp16)[name = tensor<string, []>("linear_178_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1027_cast_fp16 = add(x = input_1019_cast_fp16, y = linear_178_cast_fp16)[name = tensor<string, []>("input_1027_cast_fp16")];
             tensor<int32, [1]> x_435_axes_0 = const()[name = tensor<string, []>("x_435_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_19_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481780800)))];
-            tensor<fp16, [1024]> model_layers_19_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481782912)))];
+            tensor<fp16, [1024]> model_layers_19_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481739840)))];
+            tensor<fp16, [1024]> model_layers_19_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481741952)))];
             tensor<fp16, [1, 126, 1024]> x_435_cast_fp16 = layer_norm(axes = x_435_axes_0, beta = model_layers_19_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_19_norm_conv_weight_to_fp16, x = input_1027_cast_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
             tensor<int32, [3]> input_1029_perm_0 = const()[name = tensor<string, []>("input_1029_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_1031_pad_type_0 = const()[name = tensor<string, []>("input_1031_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2800,7 +2800,7 @@ program(1.0)
             tensor<int32, [2]> input_1031_pad_0 = const()[name = tensor<string, []>("input_1031_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1031_dilations_0 = const()[name = tensor<string, []>("input_1031_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_1031_groups_0 = const()[name = tensor<string, []>("input_1031_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481785024))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483882240))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481744064))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483841280))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_1029_cast_fp16 = transpose(perm = input_1029_perm_0, x = x_435_cast_fp16)[name = tensor<string, []>("transpose_173")];
             tensor<fp16, [1, 2048, 126]> input_1031_cast_fp16 = conv(dilations = input_1031_dilations_0, groups = input_1031_groups_0, pad = input_1031_pad_0, pad_type = input_1031_pad_type_0, strides = input_1031_strides_0, weight = model_layers_19_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1029_cast_fp16)[name = tensor<string, []>("input_1031_cast_fp16")];
             tensor<int32, []> x_437_split_num_splits_0 = const()[name = tensor<string, []>("x_437_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -2818,8 +2818,8 @@ program(1.0)
             tensor<int32, [1]> input_1037_strides_0 = const()[name = tensor<string, []>("input_1037_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_1037_pad_0 = const()[name = tensor<string, []>("input_1037_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1037_dilations_0 = const()[name = tensor<string, []>("input_1037_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_286_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_286_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483886400))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483895680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_287_to_fp16 = const()[name = tensor<string, []>("const_287_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483897792)))];
+            tensor<fp16, [1024, 1, 9]> const_286_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_286_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483845440))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483854720))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_287_to_fp16 = const()[name = tensor<string, []>("const_287_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483856832)))];
             tensor<fp16, [1, 1024, 126]> input_1039_cast_fp16 = conv(bias = const_287_to_fp16, dilations = input_1037_dilations_0, groups = input_1037_groups_0, pad = input_1037_pad_0, pad_type = input_1037_pad_type_0, strides = input_1037_strides_0, weight = const_286_to_fp16_quantized, x = input_1035_cast_fp16)[name = tensor<string, []>("input_1039_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_1041_cast_fp16 = silu(x = input_1039_cast_fp16)[name = tensor<string, []>("input_1041_cast_fp16")];
             tensor<string, []> x_439_pad_type_0 = const()[name = tensor<string, []>("x_439_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2827,64 +2827,64 @@ program(1.0)
             tensor<int32, [2]> x_439_pad_0 = const()[name = tensor<string, []>("x_439_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_439_dilations_0 = const()[name = tensor<string, []>("x_439_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_439_groups_0 = const()[name = tensor<string, []>("x_439_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483899904))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484948544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(483858944))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484907584))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_439_cast_fp16 = conv(dilations = x_439_dilations_0, groups = x_439_groups_0, pad = x_439_pad_0, pad_type = x_439_pad_type_0, strides = x_439_strides_0, weight = model_layers_19_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1041_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
             tensor<int32, [3]> input_1043_perm_0 = const()[name = tensor<string, []>("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_1043_cast_fp16 = transpose(perm = input_1043_perm_0, x = x_439_cast_fp16)[name = tensor<string, []>("transpose_172")];
             tensor<fp16, [1, 126, 1024]> input_1045_cast_fp16 = add(x = input_1027_cast_fp16, y = input_1043_cast_fp16)[name = tensor<string, []>("input_1045_cast_fp16")];
             tensor<int32, [1]> input_1047_axes_0 = const()[name = tensor<string, []>("input_1047_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484950656)))];
-            tensor<fp16, [1024]> model_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484952768)))];
+            tensor<fp16, [1024]> model_layers_19_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484909696)))];
+            tensor<fp16, [1024]> model_layers_19_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484911808)))];
             tensor<fp16, [1, 126, 1024]> input_1047_cast_fp16 = layer_norm(axes = input_1047_axes_0, beta = model_layers_19_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_19_norm_feed_forward2_weight_to_fp16, x = input_1045_cast_fp16)[name = tensor<string, []>("input_1047_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484954880))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(489149248))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_19_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(484913920))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(489108288))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_179_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1047_cast_fp16)[name = tensor<string, []>("linear_179_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1051_cast_fp16 = silu(x = linear_179_cast_fp16)[name = tensor<string, []>("input_1051_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(489157504))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493351872))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_19_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_19_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(489116544))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493310912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_180_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_19_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1051_cast_fp16)[name = tensor<string, []>("linear_180_cast_fp16")];
             tensor<fp16, []> var_3458_to_fp16 = const()[name = tensor<string, []>("op_3458_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3459_cast_fp16 = mul(x = linear_180_cast_fp16, y = var_3458_to_fp16)[name = tensor<string, []>("op_3459_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1057_cast_fp16 = add(x = input_1045_cast_fp16, y = var_3459_cast_fp16)[name = tensor<string, []>("input_1057_cast_fp16")];
             tensor<int32, [1]> input_1059_axes_0 = const()[name = tensor<string, []>("input_1059_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_19_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493353984)))];
-            tensor<fp16, [1024]> model_layers_19_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493356096)))];
+            tensor<fp16, [1024]> model_layers_19_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493313024)))];
+            tensor<fp16, [1024]> model_layers_19_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_19_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493315136)))];
             tensor<fp16, [1, 126, 1024]> input_1059_cast_fp16 = layer_norm(axes = input_1059_axes_0, beta = model_layers_19_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_19_norm_out_weight_to_fp16, x = input_1057_cast_fp16)[name = tensor<string, []>("input_1059_cast_fp16")];
             tensor<int32, [1]> input_1061_axes_0 = const()[name = tensor<string, []>("input_1061_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493358208)))];
-            tensor<fp16, [1024]> model_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493360320)))];
+            tensor<fp16, [1024]> model_layers_20_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493317248)))];
+            tensor<fp16, [1024]> model_layers_20_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493319360)))];
             tensor<fp16, [1, 126, 1024]> input_1061_cast_fp16 = layer_norm(axes = input_1061_axes_0, beta = model_layers_20_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_20_norm_feed_forward1_weight_to_fp16, x = input_1059_cast_fp16)[name = tensor<string, []>("input_1061_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493362432))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497556800))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_20_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(493321472))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497515840))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1061_cast_fp16)[name = tensor<string, []>("linear_181_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1065_cast_fp16 = silu(x = linear_181_cast_fp16)[name = tensor<string, []>("input_1065_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497565056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501759424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_20_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497524096))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501718464))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_182_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_20_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1065_cast_fp16)[name = tensor<string, []>("linear_182_cast_fp16")];
             tensor<fp16, []> var_3487_to_fp16 = const()[name = tensor<string, []>("op_3487_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3488_cast_fp16 = mul(x = linear_182_cast_fp16, y = var_3487_to_fp16)[name = tensor<string, []>("op_3488_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1071_cast_fp16 = add(x = input_1059_cast_fp16, y = var_3488_cast_fp16)[name = tensor<string, []>("input_1071_cast_fp16")];
             tensor<int32, [1]> query_41_axes_0 = const()[name = tensor<string, []>("query_41_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_20_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501761536)))];
-            tensor<fp16, [1024]> model_layers_20_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501763648)))];
+            tensor<fp16, [1024]> model_layers_20_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501720576)))];
+            tensor<fp16, [1024]> model_layers_20_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501722688)))];
             tensor<fp16, [1, 126, 1024]> query_41_cast_fp16 = layer_norm(axes = query_41_axes_0, beta = model_layers_20_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_20_norm_self_att_weight_to_fp16, x = input_1071_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501765760))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502814400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501724800))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502773440))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_183_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_20_self_attn_linear_q_weight_to_fp16_quantized, x = query_41_cast_fp16)[name = tensor<string, []>("linear_183_cast_fp16")];
             tensor<int32, [4]> var_3504 = const()[name = tensor<string, []>("op_3504"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_121_cast_fp16 = reshape(shape = var_3504, x = linear_183_cast_fp16)[name = tensor<string, []>("q_121_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502816512))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503865152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502775552))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503824192))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_184_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_20_self_attn_linear_k_weight_to_fp16_quantized, x = query_41_cast_fp16)[name = tensor<string, []>("linear_184_cast_fp16")];
             tensor<int32, [4]> var_3508 = const()[name = tensor<string, []>("op_3508"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_81_cast_fp16 = reshape(shape = var_3508, x = linear_184_cast_fp16)[name = tensor<string, []>("k_81_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503867264))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504915904))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(503826304))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504874944))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_185_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_20_self_attn_linear_v_weight_to_fp16_quantized, x = query_41_cast_fp16)[name = tensor<string, []>("linear_185_cast_fp16")];
             tensor<int32, [4]> var_3512 = const()[name = tensor<string, []>("op_3512"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_41_cast_fp16 = reshape(shape = var_3512, x = linear_185_cast_fp16)[name = tensor<string, []>("v_41_cast_fp16")];
             tensor<int32, [4]> value_41_perm_0 = const()[name = tensor<string, []>("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_20_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_20_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504918016)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3524_cast_fp16 = add(x = q_121_cast_fp16, y = model_layers_20_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3524_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_20_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_20_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504920128)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3526_cast_fp16 = add(x = q_121_cast_fp16, y = model_layers_20_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3526_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_20_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504877056))), scale = tensor<fp16, [8]>([0x1.f6cp-9, 0x1.6f4p-8, 0x1.14p-8, 0x1.abp-8, 0x1.194p-7, 0x1.f5p-8, 0x1.074p-8, 0x1.33p-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3524_cast_fp16 = add(x = q_121_cast_fp16, y = model_layers_20_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_3524_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_20_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504878144))), scale = tensor<fp16, [8]>([0x1.5bp-7, 0x1.4bp-7, 0x1.1e8p-7, 0x1.1bp-8, 0x1.d98p-10, 0x1.41cp-7, 0x1.7b8p-8, 0x1.52cp-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3526_cast_fp16 = add(x = q_121_cast_fp16, y = model_layers_20_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_3526_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_447_transpose_x_0 = const()[name = tensor<string, []>("x_447_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_447_transpose_y_0 = const()[name = tensor<string, []>("x_447_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_3528_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3528_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504922240))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(505179328))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_3528_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3528_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504879232))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(505136320))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_41_cast_fp16 = transpose(perm = q_with_bias_v_41_perm_0, x = var_3526_cast_fp16)[name = tensor<string, []>("transpose_171")];
             tensor<fp16, [1, 8, 126, 251]> x_447_cast_fp16 = matmul(transpose_x = x_447_transpose_x_0, transpose_y = x_447_transpose_y_0, x = q_with_bias_v_41_cast_fp16, y = op_3528_to_fp16_quantized)[name = tensor<string, []>("x_447_cast_fp16")];
             tensor<int32, [8]> x_449_pad_0 = const()[name = tensor<string, []>("x_449_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -2924,12 +2924,12 @@ program(1.0)
             tensor<int32, [3]> var_3561 = const()[name = tensor<string, []>("op_3561"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_3560_cast_fp16 = transpose(perm = var_3560_perm_0, x = x_453_cast_fp16)[name = tensor<string, []>("transpose_167")];
             tensor<fp16, [1, 126, 1024]> input_1075_cast_fp16 = reshape(shape = var_3561, x = var_3560_cast_fp16)[name = tensor<string, []>("input_1075_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(505179904))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506228544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(505136896))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506185536))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_187_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_20_self_attn_linear_out_weight_to_fp16_quantized, x = input_1075_cast_fp16)[name = tensor<string, []>("linear_187_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1079_cast_fp16 = add(x = input_1071_cast_fp16, y = linear_187_cast_fp16)[name = tensor<string, []>("input_1079_cast_fp16")];
             tensor<int32, [1]> x_457_axes_0 = const()[name = tensor<string, []>("x_457_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_20_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506230656)))];
-            tensor<fp16, [1024]> model_layers_20_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506232768)))];
+            tensor<fp16, [1024]> model_layers_20_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506187648)))];
+            tensor<fp16, [1024]> model_layers_20_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506189760)))];
             tensor<fp16, [1, 126, 1024]> x_457_cast_fp16 = layer_norm(axes = x_457_axes_0, beta = model_layers_20_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_20_norm_conv_weight_to_fp16, x = input_1079_cast_fp16)[name = tensor<string, []>("x_457_cast_fp16")];
             tensor<int32, [3]> input_1081_perm_0 = const()[name = tensor<string, []>("input_1081_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_1083_pad_type_0 = const()[name = tensor<string, []>("input_1083_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2937,7 +2937,7 @@ program(1.0)
             tensor<int32, [2]> input_1083_pad_0 = const()[name = tensor<string, []>("input_1083_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1083_dilations_0 = const()[name = tensor<string, []>("input_1083_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_1083_groups_0 = const()[name = tensor<string, []>("input_1083_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506234880))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508332096))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506191872))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508289088))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_1081_cast_fp16 = transpose(perm = input_1081_perm_0, x = x_457_cast_fp16)[name = tensor<string, []>("transpose_166")];
             tensor<fp16, [1, 2048, 126]> input_1083_cast_fp16 = conv(dilations = input_1083_dilations_0, groups = input_1083_groups_0, pad = input_1083_pad_0, pad_type = input_1083_pad_type_0, strides = input_1083_strides_0, weight = model_layers_20_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1081_cast_fp16)[name = tensor<string, []>("input_1083_cast_fp16")];
             tensor<int32, []> x_459_split_num_splits_0 = const()[name = tensor<string, []>("x_459_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -2955,8 +2955,8 @@ program(1.0)
             tensor<int32, [1]> input_1089_strides_0 = const()[name = tensor<string, []>("input_1089_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_1089_pad_0 = const()[name = tensor<string, []>("input_1089_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1089_dilations_0 = const()[name = tensor<string, []>("input_1089_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_288_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_288_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508336256))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508345536))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_289_to_fp16 = const()[name = tensor<string, []>("const_289_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508347648)))];
+            tensor<fp16, [1024, 1, 9]> const_288_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_288_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508293248))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508302528))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_289_to_fp16 = const()[name = tensor<string, []>("const_289_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508304640)))];
             tensor<fp16, [1, 1024, 126]> input_1091_cast_fp16 = conv(bias = const_289_to_fp16, dilations = input_1089_dilations_0, groups = input_1089_groups_0, pad = input_1089_pad_0, pad_type = input_1089_pad_type_0, strides = input_1089_strides_0, weight = const_288_to_fp16_quantized, x = input_1087_cast_fp16)[name = tensor<string, []>("input_1091_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_1093_cast_fp16 = silu(x = input_1091_cast_fp16)[name = tensor<string, []>("input_1093_cast_fp16")];
             tensor<string, []> x_461_pad_type_0 = const()[name = tensor<string, []>("x_461_pad_type_0"), val = tensor<string, []>("valid")];
@@ -2964,64 +2964,64 @@ program(1.0)
             tensor<int32, [2]> x_461_pad_0 = const()[name = tensor<string, []>("x_461_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_461_dilations_0 = const()[name = tensor<string, []>("x_461_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_461_groups_0 = const()[name = tensor<string, []>("x_461_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508349760))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509398400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(508306752))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509355392))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_461_cast_fp16 = conv(dilations = x_461_dilations_0, groups = x_461_groups_0, pad = x_461_pad_0, pad_type = x_461_pad_type_0, strides = x_461_strides_0, weight = model_layers_20_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1093_cast_fp16)[name = tensor<string, []>("x_461_cast_fp16")];
             tensor<int32, [3]> input_1095_perm_0 = const()[name = tensor<string, []>("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_1095_cast_fp16 = transpose(perm = input_1095_perm_0, x = x_461_cast_fp16)[name = tensor<string, []>("transpose_165")];
             tensor<fp16, [1, 126, 1024]> input_1097_cast_fp16 = add(x = input_1079_cast_fp16, y = input_1095_cast_fp16)[name = tensor<string, []>("input_1097_cast_fp16")];
             tensor<int32, [1]> input_1099_axes_0 = const()[name = tensor<string, []>("input_1099_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509400512)))];
-            tensor<fp16, [1024]> model_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509402624)))];
+            tensor<fp16, [1024]> model_layers_20_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509357504)))];
+            tensor<fp16, [1024]> model_layers_20_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509359616)))];
             tensor<fp16, [1, 126, 1024]> input_1099_cast_fp16 = layer_norm(axes = input_1099_axes_0, beta = model_layers_20_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_20_norm_feed_forward2_weight_to_fp16, x = input_1097_cast_fp16)[name = tensor<string, []>("input_1099_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509404736))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513599104))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_20_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(509361728))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513556096))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_188_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1099_cast_fp16)[name = tensor<string, []>("linear_188_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1103_cast_fp16 = silu(x = linear_188_cast_fp16)[name = tensor<string, []>("input_1103_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513607360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517801728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_20_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_20_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513564352))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517758720))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_189_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_20_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1103_cast_fp16)[name = tensor<string, []>("linear_189_cast_fp16")];
             tensor<fp16, []> var_3621_to_fp16 = const()[name = tensor<string, []>("op_3621_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3622_cast_fp16 = mul(x = linear_189_cast_fp16, y = var_3621_to_fp16)[name = tensor<string, []>("op_3622_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1109_cast_fp16 = add(x = input_1097_cast_fp16, y = var_3622_cast_fp16)[name = tensor<string, []>("input_1109_cast_fp16")];
             tensor<int32, [1]> input_1111_axes_0 = const()[name = tensor<string, []>("input_1111_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_20_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517803840)))];
-            tensor<fp16, [1024]> model_layers_20_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517805952)))];
+            tensor<fp16, [1024]> model_layers_20_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517760832)))];
+            tensor<fp16, [1024]> model_layers_20_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_20_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517762944)))];
             tensor<fp16, [1, 126, 1024]> input_1111_cast_fp16 = layer_norm(axes = input_1111_axes_0, beta = model_layers_20_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_20_norm_out_weight_to_fp16, x = input_1109_cast_fp16)[name = tensor<string, []>("input_1111_cast_fp16")];
             tensor<int32, [1]> input_1113_axes_0 = const()[name = tensor<string, []>("input_1113_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517808064)))];
-            tensor<fp16, [1024]> model_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517810176)))];
+            tensor<fp16, [1024]> model_layers_21_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517765056)))];
+            tensor<fp16, [1024]> model_layers_21_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517767168)))];
             tensor<fp16, [1, 126, 1024]> input_1113_cast_fp16 = layer_norm(axes = input_1113_axes_0, beta = model_layers_21_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_21_norm_feed_forward1_weight_to_fp16, x = input_1111_cast_fp16)[name = tensor<string, []>("input_1113_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517812288))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522006656))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_21_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(517769280))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(521963648))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1113_cast_fp16)[name = tensor<string, []>("linear_190_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1117_cast_fp16 = silu(x = linear_190_cast_fp16)[name = tensor<string, []>("input_1117_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522014912))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526209280))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_21_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(521971904))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526166272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_191_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_21_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1117_cast_fp16)[name = tensor<string, []>("linear_191_cast_fp16")];
             tensor<fp16, []> var_3650_to_fp16 = const()[name = tensor<string, []>("op_3650_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3651_cast_fp16 = mul(x = linear_191_cast_fp16, y = var_3650_to_fp16)[name = tensor<string, []>("op_3651_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1123_cast_fp16 = add(x = input_1111_cast_fp16, y = var_3651_cast_fp16)[name = tensor<string, []>("input_1123_cast_fp16")];
             tensor<int32, [1]> query_43_axes_0 = const()[name = tensor<string, []>("query_43_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_21_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526211392)))];
-            tensor<fp16, [1024]> model_layers_21_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526213504)))];
+            tensor<fp16, [1024]> model_layers_21_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526168384)))];
+            tensor<fp16, [1024]> model_layers_21_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526170496)))];
             tensor<fp16, [1, 126, 1024]> query_43_cast_fp16 = layer_norm(axes = query_43_axes_0, beta = model_layers_21_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_21_norm_self_att_weight_to_fp16, x = input_1123_cast_fp16)[name = tensor<string, []>("query_43_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526215616))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527264256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526172608))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527221248))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_192_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_21_self_attn_linear_q_weight_to_fp16_quantized, x = query_43_cast_fp16)[name = tensor<string, []>("linear_192_cast_fp16")];
             tensor<int32, [4]> var_3667 = const()[name = tensor<string, []>("op_3667"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_127_cast_fp16 = reshape(shape = var_3667, x = linear_192_cast_fp16)[name = tensor<string, []>("q_127_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527266368))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528315008))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527223360))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528272000))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_193_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_21_self_attn_linear_k_weight_to_fp16_quantized, x = query_43_cast_fp16)[name = tensor<string, []>("linear_193_cast_fp16")];
             tensor<int32, [4]> var_3671 = const()[name = tensor<string, []>("op_3671"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_85_cast_fp16 = reshape(shape = var_3671, x = linear_193_cast_fp16)[name = tensor<string, []>("k_85_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528317120))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529365760))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(528274112))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529322752))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_194_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_21_self_attn_linear_v_weight_to_fp16_quantized, x = query_43_cast_fp16)[name = tensor<string, []>("linear_194_cast_fp16")];
             tensor<int32, [4]> var_3675 = const()[name = tensor<string, []>("op_3675"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_43_cast_fp16 = reshape(shape = var_3675, x = linear_194_cast_fp16)[name = tensor<string, []>("v_43_cast_fp16")];
             tensor<int32, [4]> value_43_perm_0 = const()[name = tensor<string, []>("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_21_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_21_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529367872)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3687_cast_fp16 = add(x = q_127_cast_fp16, y = model_layers_21_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3687_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_21_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_21_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529369984)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3689_cast_fp16 = add(x = q_127_cast_fp16, y = model_layers_21_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3689_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_21_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529324864))), scale = tensor<fp16, [8]>([0x1.cdp-8, 0x1.1bcp-7, 0x1.194p-8, 0x1.018p-8, 0x1.15cp-7, 0x1.3bcp-8, 0x1.c4cp-8, 0x1.68cp-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3687_cast_fp16 = add(x = q_127_cast_fp16, y = model_layers_21_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_3687_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_21_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529325952))), scale = tensor<fp16, [8]>([0x1.09cp-7, 0x1.c14p-8, 0x1.15cp-7, 0x1.404p-7, 0x1.edcp-8, 0x1.a24p-8, 0x1.66p-7, 0x1.20cp-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3689_cast_fp16 = add(x = q_127_cast_fp16, y = model_layers_21_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_3689_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_469_transpose_x_0 = const()[name = tensor<string, []>("x_469_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_469_transpose_y_0 = const()[name = tensor<string, []>("x_469_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_3691_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3691_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529372096))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529629184))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_3691_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3691_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529327040))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529584128))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_43_cast_fp16 = transpose(perm = q_with_bias_v_43_perm_0, x = var_3689_cast_fp16)[name = tensor<string, []>("transpose_164")];
             tensor<fp16, [1, 8, 126, 251]> x_469_cast_fp16 = matmul(transpose_x = x_469_transpose_x_0, transpose_y = x_469_transpose_y_0, x = q_with_bias_v_43_cast_fp16, y = op_3691_to_fp16_quantized)[name = tensor<string, []>("x_469_cast_fp16")];
             tensor<int32, [8]> x_471_pad_0 = const()[name = tensor<string, []>("x_471_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -3061,12 +3061,12 @@ program(1.0)
             tensor<int32, [3]> var_3724 = const()[name = tensor<string, []>("op_3724"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_3723_cast_fp16 = transpose(perm = var_3723_perm_0, x = x_475_cast_fp16)[name = tensor<string, []>("transpose_160")];
             tensor<fp16, [1, 126, 1024]> input_1127_cast_fp16 = reshape(shape = var_3724, x = var_3723_cast_fp16)[name = tensor<string, []>("input_1127_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529629760))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530678400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529584704))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530633344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_196_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_21_self_attn_linear_out_weight_to_fp16_quantized, x = input_1127_cast_fp16)[name = tensor<string, []>("linear_196_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1131_cast_fp16 = add(x = input_1123_cast_fp16, y = linear_196_cast_fp16)[name = tensor<string, []>("input_1131_cast_fp16")];
             tensor<int32, [1]> x_479_axes_0 = const()[name = tensor<string, []>("x_479_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_21_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530680512)))];
-            tensor<fp16, [1024]> model_layers_21_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530682624)))];
+            tensor<fp16, [1024]> model_layers_21_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530635456)))];
+            tensor<fp16, [1024]> model_layers_21_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530637568)))];
             tensor<fp16, [1, 126, 1024]> x_479_cast_fp16 = layer_norm(axes = x_479_axes_0, beta = model_layers_21_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_21_norm_conv_weight_to_fp16, x = input_1131_cast_fp16)[name = tensor<string, []>("x_479_cast_fp16")];
             tensor<int32, [3]> input_1133_perm_0 = const()[name = tensor<string, []>("input_1133_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_1135_pad_type_0 = const()[name = tensor<string, []>("input_1135_pad_type_0"), val = tensor<string, []>("valid")];
@@ -3074,7 +3074,7 @@ program(1.0)
             tensor<int32, [2]> input_1135_pad_0 = const()[name = tensor<string, []>("input_1135_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1135_dilations_0 = const()[name = tensor<string, []>("input_1135_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_1135_groups_0 = const()[name = tensor<string, []>("input_1135_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530684736))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532781952))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530639680))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532736896))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_1133_cast_fp16 = transpose(perm = input_1133_perm_0, x = x_479_cast_fp16)[name = tensor<string, []>("transpose_159")];
             tensor<fp16, [1, 2048, 126]> input_1135_cast_fp16 = conv(dilations = input_1135_dilations_0, groups = input_1135_groups_0, pad = input_1135_pad_0, pad_type = input_1135_pad_type_0, strides = input_1135_strides_0, weight = model_layers_21_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1133_cast_fp16)[name = tensor<string, []>("input_1135_cast_fp16")];
             tensor<int32, []> x_481_split_num_splits_0 = const()[name = tensor<string, []>("x_481_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -3092,8 +3092,8 @@ program(1.0)
             tensor<int32, [1]> input_1141_strides_0 = const()[name = tensor<string, []>("input_1141_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_1141_pad_0 = const()[name = tensor<string, []>("input_1141_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1141_dilations_0 = const()[name = tensor<string, []>("input_1141_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_290_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_290_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532786112))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532795392))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_291_to_fp16 = const()[name = tensor<string, []>("const_291_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532797504)))];
+            tensor<fp16, [1024, 1, 9]> const_290_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_290_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532741056))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532750336))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_291_to_fp16 = const()[name = tensor<string, []>("const_291_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532752448)))];
             tensor<fp16, [1, 1024, 126]> input_1143_cast_fp16 = conv(bias = const_291_to_fp16, dilations = input_1141_dilations_0, groups = input_1141_groups_0, pad = input_1141_pad_0, pad_type = input_1141_pad_type_0, strides = input_1141_strides_0, weight = const_290_to_fp16_quantized, x = input_1139_cast_fp16)[name = tensor<string, []>("input_1143_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_1145_cast_fp16 = silu(x = input_1143_cast_fp16)[name = tensor<string, []>("input_1145_cast_fp16")];
             tensor<string, []> x_483_pad_type_0 = const()[name = tensor<string, []>("x_483_pad_type_0"), val = tensor<string, []>("valid")];
@@ -3101,64 +3101,64 @@ program(1.0)
             tensor<int32, [2]> x_483_pad_0 = const()[name = tensor<string, []>("x_483_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_483_dilations_0 = const()[name = tensor<string, []>("x_483_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_483_groups_0 = const()[name = tensor<string, []>("x_483_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532799616))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533848256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(532754560))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533803200))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_483_cast_fp16 = conv(dilations = x_483_dilations_0, groups = x_483_groups_0, pad = x_483_pad_0, pad_type = x_483_pad_type_0, strides = x_483_strides_0, weight = model_layers_21_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1145_cast_fp16)[name = tensor<string, []>("x_483_cast_fp16")];
             tensor<int32, [3]> input_1147_perm_0 = const()[name = tensor<string, []>("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_1147_cast_fp16 = transpose(perm = input_1147_perm_0, x = x_483_cast_fp16)[name = tensor<string, []>("transpose_158")];
             tensor<fp16, [1, 126, 1024]> input_1149_cast_fp16 = add(x = input_1131_cast_fp16, y = input_1147_cast_fp16)[name = tensor<string, []>("input_1149_cast_fp16")];
             tensor<int32, [1]> input_1151_axes_0 = const()[name = tensor<string, []>("input_1151_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533850368)))];
-            tensor<fp16, [1024]> model_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533852480)))];
+            tensor<fp16, [1024]> model_layers_21_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533805312)))];
+            tensor<fp16, [1024]> model_layers_21_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533807424)))];
             tensor<fp16, [1, 126, 1024]> input_1151_cast_fp16 = layer_norm(axes = input_1151_axes_0, beta = model_layers_21_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_21_norm_feed_forward2_weight_to_fp16, x = input_1149_cast_fp16)[name = tensor<string, []>("input_1151_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_21_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533854592))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538048960))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_21_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(533809536))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538003904))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_197_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1151_cast_fp16)[name = tensor<string, []>("linear_197_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1155_cast_fp16 = silu(x = linear_197_cast_fp16)[name = tensor<string, []>("input_1155_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_21_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538057216))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542251584))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_21_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_21_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538012160))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542206528))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_198_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_21_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1155_cast_fp16)[name = tensor<string, []>("linear_198_cast_fp16")];
             tensor<fp16, []> var_3784_to_fp16 = const()[name = tensor<string, []>("op_3784_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3785_cast_fp16 = mul(x = linear_198_cast_fp16, y = var_3784_to_fp16)[name = tensor<string, []>("op_3785_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1161_cast_fp16 = add(x = input_1149_cast_fp16, y = var_3785_cast_fp16)[name = tensor<string, []>("input_1161_cast_fp16")];
             tensor<int32, [1]> input_1163_axes_0 = const()[name = tensor<string, []>("input_1163_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_21_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542253696)))];
-            tensor<fp16, [1024]> model_layers_21_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542255808)))];
+            tensor<fp16, [1024]> model_layers_21_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542208640)))];
+            tensor<fp16, [1024]> model_layers_21_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_21_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542210752)))];
             tensor<fp16, [1, 126, 1024]> input_1163_cast_fp16 = layer_norm(axes = input_1163_axes_0, beta = model_layers_21_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_21_norm_out_weight_to_fp16, x = input_1161_cast_fp16)[name = tensor<string, []>("input_1163_cast_fp16")];
             tensor<int32, [1]> input_1165_axes_0 = const()[name = tensor<string, []>("input_1165_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542257920)))];
-            tensor<fp16, [1024]> model_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542260032)))];
+            tensor<fp16, [1024]> model_layers_22_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542212864)))];
+            tensor<fp16, [1024]> model_layers_22_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542214976)))];
             tensor<fp16, [1, 126, 1024]> input_1165_cast_fp16 = layer_norm(axes = input_1165_axes_0, beta = model_layers_22_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_22_norm_feed_forward1_weight_to_fp16, x = input_1163_cast_fp16)[name = tensor<string, []>("input_1165_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_22_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542262144))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546456512))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_22_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542217088))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546411456))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_199_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1165_cast_fp16)[name = tensor<string, []>("linear_199_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1169_cast_fp16 = silu(x = linear_199_cast_fp16)[name = tensor<string, []>("input_1169_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_22_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546464768))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550659136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_22_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(546419712))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550614080))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_200_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_22_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1169_cast_fp16)[name = tensor<string, []>("linear_200_cast_fp16")];
             tensor<fp16, []> var_3813_to_fp16 = const()[name = tensor<string, []>("op_3813_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3814_cast_fp16 = mul(x = linear_200_cast_fp16, y = var_3813_to_fp16)[name = tensor<string, []>("op_3814_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1175_cast_fp16 = add(x = input_1163_cast_fp16, y = var_3814_cast_fp16)[name = tensor<string, []>("input_1175_cast_fp16")];
             tensor<int32, [1]> query_45_axes_0 = const()[name = tensor<string, []>("query_45_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_22_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550661248)))];
-            tensor<fp16, [1024]> model_layers_22_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550663360)))];
+            tensor<fp16, [1024]> model_layers_22_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550616192)))];
+            tensor<fp16, [1024]> model_layers_22_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550618304)))];
             tensor<fp16, [1, 126, 1024]> query_45_cast_fp16 = layer_norm(axes = query_45_axes_0, beta = model_layers_22_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_22_norm_self_att_weight_to_fp16, x = input_1175_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550665472))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551714112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550620416))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551669056))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_201_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_22_self_attn_linear_q_weight_to_fp16_quantized, x = query_45_cast_fp16)[name = tensor<string, []>("linear_201_cast_fp16")];
             tensor<int32, [4]> var_3830 = const()[name = tensor<string, []>("op_3830"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_133_cast_fp16 = reshape(shape = var_3830, x = linear_201_cast_fp16)[name = tensor<string, []>("q_133_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551716224))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552764864))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551671168))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552719808))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_202_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_22_self_attn_linear_k_weight_to_fp16_quantized, x = query_45_cast_fp16)[name = tensor<string, []>("linear_202_cast_fp16")];
             tensor<int32, [4]> var_3834 = const()[name = tensor<string, []>("op_3834"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_89_cast_fp16 = reshape(shape = var_3834, x = linear_202_cast_fp16)[name = tensor<string, []>("k_89_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552766976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553815616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(552721920))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553770560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_203_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_22_self_attn_linear_v_weight_to_fp16_quantized, x = query_45_cast_fp16)[name = tensor<string, []>("linear_203_cast_fp16")];
             tensor<int32, [4]> var_3838 = const()[name = tensor<string, []>("op_3838"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_45_cast_fp16 = reshape(shape = var_3838, x = linear_203_cast_fp16)[name = tensor<string, []>("v_45_cast_fp16")];
             tensor<int32, [4]> value_45_perm_0 = const()[name = tensor<string, []>("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_22_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_22_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553817728)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3850_cast_fp16 = add(x = q_133_cast_fp16, y = model_layers_22_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_3850_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_22_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_22_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553819840)))];
-            tensor<fp16, [1, 126, 8, 128]> var_3852_cast_fp16 = add(x = q_133_cast_fp16, y = model_layers_22_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_3852_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_22_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553772672))), scale = tensor<fp16, [8]>([0x1.184p-8, 0x1.3e8p-7, 0x1.584p-7, 0x1.becp-8, 0x1.ce8p-7, 0x1.48cp-8, 0x1.57p-8, 0x1.c8p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3850_cast_fp16 = add(x = q_133_cast_fp16, y = model_layers_22_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_3850_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_22_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553773760))), scale = tensor<fp16, [8]>([0x1.794p-8, 0x1.d48p-9, 0x1.2c4p-8, 0x1.e9cp-8, 0x1.01cp-7, 0x1.a14p-7, 0x1.06p-7, 0x1.81cp-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_3852_cast_fp16 = add(x = q_133_cast_fp16, y = model_layers_22_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_3852_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_491_transpose_x_0 = const()[name = tensor<string, []>("x_491_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_491_transpose_y_0 = const()[name = tensor<string, []>("x_491_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_3854_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3854_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553821952))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554079040))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_3854_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3854_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553774848))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554031936))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_45_cast_fp16 = transpose(perm = q_with_bias_v_45_perm_0, x = var_3852_cast_fp16)[name = tensor<string, []>("transpose_157")];
             tensor<fp16, [1, 8, 126, 251]> x_491_cast_fp16 = matmul(transpose_x = x_491_transpose_x_0, transpose_y = x_491_transpose_y_0, x = q_with_bias_v_45_cast_fp16, y = op_3854_to_fp16_quantized)[name = tensor<string, []>("x_491_cast_fp16")];
             tensor<int32, [8]> x_493_pad_0 = const()[name = tensor<string, []>("x_493_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -3198,12 +3198,12 @@ program(1.0)
             tensor<int32, [3]> var_3887 = const()[name = tensor<string, []>("op_3887"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_3886_cast_fp16 = transpose(perm = var_3886_perm_0, x = x_497_cast_fp16)[name = tensor<string, []>("transpose_153")];
             tensor<fp16, [1, 126, 1024]> input_1179_cast_fp16 = reshape(shape = var_3887, x = var_3886_cast_fp16)[name = tensor<string, []>("input_1179_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554079616))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555128256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554032512))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555081152))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_205_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_22_self_attn_linear_out_weight_to_fp16_quantized, x = input_1179_cast_fp16)[name = tensor<string, []>("linear_205_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1183_cast_fp16 = add(x = input_1175_cast_fp16, y = linear_205_cast_fp16)[name = tensor<string, []>("input_1183_cast_fp16")];
             tensor<int32, [1]> x_501_axes_0 = const()[name = tensor<string, []>("x_501_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_22_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555130368)))];
-            tensor<fp16, [1024]> model_layers_22_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555132480)))];
+            tensor<fp16, [1024]> model_layers_22_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555083264)))];
+            tensor<fp16, [1024]> model_layers_22_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555085376)))];
             tensor<fp16, [1, 126, 1024]> x_501_cast_fp16 = layer_norm(axes = x_501_axes_0, beta = model_layers_22_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_22_norm_conv_weight_to_fp16, x = input_1183_cast_fp16)[name = tensor<string, []>("x_501_cast_fp16")];
             tensor<int32, [3]> input_1185_perm_0 = const()[name = tensor<string, []>("input_1185_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_1187_pad_type_0 = const()[name = tensor<string, []>("input_1187_pad_type_0"), val = tensor<string, []>("valid")];
@@ -3211,7 +3211,7 @@ program(1.0)
             tensor<int32, [2]> input_1187_pad_0 = const()[name = tensor<string, []>("input_1187_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1187_dilations_0 = const()[name = tensor<string, []>("input_1187_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_1187_groups_0 = const()[name = tensor<string, []>("input_1187_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555134592))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557231808))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(555087488))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557184704))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_1185_cast_fp16 = transpose(perm = input_1185_perm_0, x = x_501_cast_fp16)[name = tensor<string, []>("transpose_152")];
             tensor<fp16, [1, 2048, 126]> input_1187_cast_fp16 = conv(dilations = input_1187_dilations_0, groups = input_1187_groups_0, pad = input_1187_pad_0, pad_type = input_1187_pad_type_0, strides = input_1187_strides_0, weight = model_layers_22_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1185_cast_fp16)[name = tensor<string, []>("input_1187_cast_fp16")];
             tensor<int32, []> x_503_split_num_splits_0 = const()[name = tensor<string, []>("x_503_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -3229,8 +3229,8 @@ program(1.0)
             tensor<int32, [1]> input_1193_strides_0 = const()[name = tensor<string, []>("input_1193_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_1193_pad_0 = const()[name = tensor<string, []>("input_1193_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1193_dilations_0 = const()[name = tensor<string, []>("input_1193_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_292_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_292_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557235968))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557245248))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_293_to_fp16 = const()[name = tensor<string, []>("const_293_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557247360)))];
+            tensor<fp16, [1024, 1, 9]> const_292_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_292_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557188864))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557198144))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_293_to_fp16 = const()[name = tensor<string, []>("const_293_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557200256)))];
             tensor<fp16, [1, 1024, 126]> input_1195_cast_fp16 = conv(bias = const_293_to_fp16, dilations = input_1193_dilations_0, groups = input_1193_groups_0, pad = input_1193_pad_0, pad_type = input_1193_pad_type_0, strides = input_1193_strides_0, weight = const_292_to_fp16_quantized, x = input_1191_cast_fp16)[name = tensor<string, []>("input_1195_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_1197_cast_fp16 = silu(x = input_1195_cast_fp16)[name = tensor<string, []>("input_1197_cast_fp16")];
             tensor<string, []> x_505_pad_type_0 = const()[name = tensor<string, []>("x_505_pad_type_0"), val = tensor<string, []>("valid")];
@@ -3238,64 +3238,64 @@ program(1.0)
             tensor<int32, [2]> x_505_pad_0 = const()[name = tensor<string, []>("x_505_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_505_dilations_0 = const()[name = tensor<string, []>("x_505_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_505_groups_0 = const()[name = tensor<string, []>("x_505_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557249472))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558298112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(557202368))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558251008))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_505_cast_fp16 = conv(dilations = x_505_dilations_0, groups = x_505_groups_0, pad = x_505_pad_0, pad_type = x_505_pad_type_0, strides = x_505_strides_0, weight = model_layers_22_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1197_cast_fp16)[name = tensor<string, []>("x_505_cast_fp16")];
             tensor<int32, [3]> input_1199_perm_0 = const()[name = tensor<string, []>("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_1199_cast_fp16 = transpose(perm = input_1199_perm_0, x = x_505_cast_fp16)[name = tensor<string, []>("transpose_151")];
             tensor<fp16, [1, 126, 1024]> input_1201_cast_fp16 = add(x = input_1183_cast_fp16, y = input_1199_cast_fp16)[name = tensor<string, []>("input_1201_cast_fp16")];
             tensor<int32, [1]> input_1203_axes_0 = const()[name = tensor<string, []>("input_1203_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558300224)))];
-            tensor<fp16, [1024]> model_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558302336)))];
+            tensor<fp16, [1024]> model_layers_22_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558253120)))];
+            tensor<fp16, [1024]> model_layers_22_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558255232)))];
             tensor<fp16, [1, 126, 1024]> input_1203_cast_fp16 = layer_norm(axes = input_1203_axes_0, beta = model_layers_22_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_22_norm_feed_forward2_weight_to_fp16, x = input_1201_cast_fp16)[name = tensor<string, []>("input_1203_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_22_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558304448))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(562498816))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_22_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558257344))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(562451712))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_206_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1203_cast_fp16)[name = tensor<string, []>("linear_206_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1207_cast_fp16 = silu(x = linear_206_cast_fp16)[name = tensor<string, []>("input_1207_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_22_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(562507072))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566701440))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_22_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_22_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(562459968))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566654336))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_207_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_22_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1207_cast_fp16)[name = tensor<string, []>("linear_207_cast_fp16")];
             tensor<fp16, []> var_3947_to_fp16 = const()[name = tensor<string, []>("op_3947_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3948_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3947_to_fp16)[name = tensor<string, []>("op_3948_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1213_cast_fp16 = add(x = input_1201_cast_fp16, y = var_3948_cast_fp16)[name = tensor<string, []>("input_1213_cast_fp16")];
             tensor<int32, [1]> input_1215_axes_0 = const()[name = tensor<string, []>("input_1215_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_22_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566703552)))];
-            tensor<fp16, [1024]> model_layers_22_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566705664)))];
+            tensor<fp16, [1024]> model_layers_22_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566656448)))];
+            tensor<fp16, [1024]> model_layers_22_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_22_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566658560)))];
             tensor<fp16, [1, 126, 1024]> input_1215_cast_fp16 = layer_norm(axes = input_1215_axes_0, beta = model_layers_22_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_22_norm_out_weight_to_fp16, x = input_1213_cast_fp16)[name = tensor<string, []>("input_1215_cast_fp16")];
             tensor<int32, [1]> input_1217_axes_0 = const()[name = tensor<string, []>("input_1217_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566707776)))];
-            tensor<fp16, [1024]> model_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566709888)))];
+            tensor<fp16, [1024]> model_layers_23_norm_feed_forward1_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566660672)))];
+            tensor<fp16, [1024]> model_layers_23_norm_feed_forward1_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566662784)))];
             tensor<fp16, [1, 126, 1024]> input_1217_cast_fp16 = layer_norm(axes = input_1217_axes_0, beta = model_layers_23_norm_feed_forward1_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_23_norm_feed_forward1_weight_to_fp16, x = input_1215_cast_fp16)[name = tensor<string, []>("input_1217_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_23_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566712000))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570906368))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_23_feed_forward1_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward1_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(566664896))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570859264))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_208_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_feed_forward1_linear1_weight_to_fp16_quantized, x = input_1217_cast_fp16)[name = tensor<string, []>("linear_208_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1221_cast_fp16 = silu(x = linear_208_cast_fp16)[name = tensor<string, []>("input_1221_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_23_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570914624))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575108992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_23_feed_forward1_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward1_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(570867520))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575061888))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_209_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_23_feed_forward1_linear2_weight_to_fp16_quantized, x = input_1221_cast_fp16)[name = tensor<string, []>("linear_209_cast_fp16")];
             tensor<fp16, []> var_3976_to_fp16 = const()[name = tensor<string, []>("op_3976_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_3977_cast_fp16 = mul(x = linear_209_cast_fp16, y = var_3976_to_fp16)[name = tensor<string, []>("op_3977_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1227_cast_fp16 = add(x = input_1215_cast_fp16, y = var_3977_cast_fp16)[name = tensor<string, []>("input_1227_cast_fp16")];
             tensor<int32, [1]> query_axes_0 = const()[name = tensor<string, []>("query_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_23_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575111104)))];
-            tensor<fp16, [1024]> model_layers_23_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575113216)))];
+            tensor<fp16, [1024]> model_layers_23_norm_self_att_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_self_att_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575064000)))];
+            tensor<fp16, [1024]> model_layers_23_norm_self_att_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_self_att_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575066112)))];
             tensor<fp16, [1, 126, 1024]> query_cast_fp16 = layer_norm(axes = query_axes_0, beta = model_layers_23_norm_self_att_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_23_norm_self_att_weight_to_fp16, x = input_1227_cast_fp16)[name = tensor<string, []>("query_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575115328))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(576163968))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_q_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_q_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575068224))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(576116864))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_210_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_23_self_attn_linear_q_weight_to_fp16_quantized, x = query_cast_fp16)[name = tensor<string, []>("linear_210_cast_fp16")];
             tensor<int32, [4]> var_3993 = const()[name = tensor<string, []>("op_3993"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> q_139_cast_fp16 = reshape(shape = var_3993, x = linear_210_cast_fp16)[name = tensor<string, []>("q_139_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(576166080))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577214720))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_k_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_k_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(576118976))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577167616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_211_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_23_self_attn_linear_k_weight_to_fp16_quantized, x = query_cast_fp16)[name = tensor<string, []>("linear_211_cast_fp16")];
             tensor<int32, [4]> var_3997 = const()[name = tensor<string, []>("op_3997"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> k_93_cast_fp16 = reshape(shape = var_3997, x = linear_211_cast_fp16)[name = tensor<string, []>("k_93_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577216832))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578265472))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_v_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_v_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577169728))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578218368))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_212_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_23_self_attn_linear_v_weight_to_fp16_quantized, x = query_cast_fp16)[name = tensor<string, []>("linear_212_cast_fp16")];
             tensor<int32, [4]> var_4001 = const()[name = tensor<string, []>("op_4001"), val = tensor<int32, [4]>([1, -1, 8, 128])];
             tensor<fp16, [1, 126, 8, 128]> v_cast_fp16 = reshape(shape = var_4001, x = linear_212_cast_fp16)[name = tensor<string, []>("v_cast_fp16")];
             tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
-            tensor<fp16, [8, 128]> model_layers_23_self_attn_pos_bias_u_to_fp16 = const()[name = tensor<string, []>("model_layers_23_self_attn_pos_bias_u_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578267584)))];
-            tensor<fp16, [1, 126, 8, 128]> var_4013_cast_fp16 = add(x = q_139_cast_fp16, y = model_layers_23_self_attn_pos_bias_u_to_fp16)[name = tensor<string, []>("op_4013_cast_fp16")];
-            tensor<fp16, [8, 128]> model_layers_23_self_attn_pos_bias_v_to_fp16 = const()[name = tensor<string, []>("model_layers_23_self_attn_pos_bias_v_to_fp16"), val = tensor<fp16, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578269696)))];
-            tensor<fp16, [1, 126, 8, 128]> var_4015_cast_fp16 = add(x = q_139_cast_fp16, y = model_layers_23_self_attn_pos_bias_v_to_fp16)[name = tensor<string, []>("op_4015_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_23_self_attn_pos_bias_u_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_pos_bias_u_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578220480))), scale = tensor<fp16, [8]>([0x1.21p-7, 0x1.8p-8, 0x1.544p-8, 0x1.a88p-8, 0x1.01p-7, 0x1.95p-8, 0x1.2dcp-8, 0x1.39p-7]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_4013_cast_fp16 = add(x = q_139_cast_fp16, y = model_layers_23_self_attn_pos_bias_u_to_fp16_quantized)[name = tensor<string, []>("op_4013_cast_fp16")];
+            tensor<fp16, [8, 128]> model_layers_23_self_attn_pos_bias_v_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_pos_bias_v_to_fp16_quantized"), quantized_data = tensor<int8, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578221568))), scale = tensor<fp16, [8]>([0x1.42cp-7, 0x1.bfcp-8, 0x1.864p-8, 0x1.0f4p-7, 0x1.7d8p-9, 0x1.c1cp-8, 0x1.ce8p-9, 0x1.81cp-8]), zero_point = tensor<int8, [8]>([0, 0, 0, 0, 0, 0, 0, 0])];
+            tensor<fp16, [1, 126, 8, 128]> var_4015_cast_fp16 = add(x = q_139_cast_fp16, y = model_layers_23_self_attn_pos_bias_v_to_fp16_quantized)[name = tensor<string, []>("op_4015_cast_fp16")];
             tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
             tensor<bool, []> x_513_transpose_x_0 = const()[name = tensor<string, []>("x_513_transpose_x_0"), val = tensor<bool, []>(false)];
             tensor<bool, []> x_513_transpose_y_0 = const()[name = tensor<string, []>("x_513_transpose_y_0"), val = tensor<bool, []>(false)];
-            tensor<fp16, [1, 8, 128, 251]> op_4017_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4017_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578271808))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578528896))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16179776)))];
+            tensor<fp16, [1, 8, 128, 251]> op_4017_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4017_to_fp16_quantized"), quantized_data = tensor<int8, [1, 8, 128, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578222656))), scale = tensor<fp16, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578479744))), zero_point = tensor<int8, [251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16177728)))];
             tensor<fp16, [1, 8, 126, 128]> q_with_bias_v_cast_fp16 = transpose(perm = q_with_bias_v_perm_0, x = var_4015_cast_fp16)[name = tensor<string, []>("transpose_150")];
             tensor<fp16, [1, 8, 126, 251]> x_513_cast_fp16 = matmul(transpose_x = x_513_transpose_x_0, transpose_y = x_513_transpose_y_0, x = q_with_bias_v_cast_fp16, y = op_4017_to_fp16_quantized)[name = tensor<string, []>("x_513_cast_fp16")];
             tensor<int32, [8]> x_515_pad_0 = const()[name = tensor<string, []>("x_515_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
@@ -3335,12 +3335,12 @@ program(1.0)
             tensor<int32, [3]> var_4050 = const()[name = tensor<string, []>("op_4050"), val = tensor<int32, [3]>([1, -1, 1024])];
             tensor<fp16, [1, 126, 8, 128]> var_4049_cast_fp16 = transpose(perm = var_4049_perm_0, x = x_519_cast_fp16)[name = tensor<string, []>("transpose_146")];
             tensor<fp16, [1, 126, 1024]> input_1231_cast_fp16 = reshape(shape = var_4050, x = var_4049_cast_fp16)[name = tensor<string, []>("input_1231_cast_fp16")];
-            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578529472))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579578112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_linear_out_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_self_attn_linear_out_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578480320))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579528960))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_214_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_23_self_attn_linear_out_weight_to_fp16_quantized, x = input_1231_cast_fp16)[name = tensor<string, []>("linear_214_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_1235_cast_fp16 = add(x = input_1227_cast_fp16, y = linear_214_cast_fp16)[name = tensor<string, []>("input_1235_cast_fp16")];
             tensor<int32, [1]> x_523_axes_0 = const()[name = tensor<string, []>("x_523_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_23_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579580224)))];
-            tensor<fp16, [1024]> model_layers_23_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579582336)))];
+            tensor<fp16, [1024]> model_layers_23_norm_conv_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_conv_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579531072)))];
+            tensor<fp16, [1024]> model_layers_23_norm_conv_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_conv_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579533184)))];
             tensor<fp16, [1, 126, 1024]> x_523_cast_fp16 = layer_norm(axes = x_523_axes_0, beta = model_layers_23_norm_conv_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_23_norm_conv_weight_to_fp16, x = input_1235_cast_fp16)[name = tensor<string, []>("x_523_cast_fp16")];
             tensor<int32, [3]> input_1237_perm_0 = const()[name = tensor<string, []>("input_1237_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<string, []> input_1239_pad_type_0 = const()[name = tensor<string, []>("input_1239_pad_type_0"), val = tensor<string, []>("valid")];
@@ -3348,7 +3348,7 @@ program(1.0)
             tensor<int32, [2]> input_1239_pad_0 = const()[name = tensor<string, []>("input_1239_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1239_dilations_0 = const()[name = tensor<string, []>("input_1239_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_1239_groups_0 = const()[name = tensor<string, []>("input_1239_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [2048, 1024, 1]> model_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579584448))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581681664))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19332864)))];
+            tensor<fp16, [2048, 1024, 1]> model_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579535296))), scale = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581632512))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19330816)))];
             tensor<fp16, [1, 1024, 126]> input_1237_cast_fp16 = transpose(perm = input_1237_perm_0, x = x_523_cast_fp16)[name = tensor<string, []>("transpose_145")];
             tensor<fp16, [1, 2048, 126]> input_1239_cast_fp16 = conv(dilations = input_1239_dilations_0, groups = input_1239_groups_0, pad = input_1239_pad_0, pad_type = input_1239_pad_type_0, strides = input_1239_strides_0, weight = model_layers_23_conv_pointwise_conv1_weight_to_fp16_quantized, x = input_1237_cast_fp16)[name = tensor<string, []>("input_1239_cast_fp16")];
             tensor<int32, []> x_525_split_num_splits_0 = const()[name = tensor<string, []>("x_525_split_num_splits_0"), val = tensor<int32, []>(2)];
@@ -3366,8 +3366,8 @@ program(1.0)
             tensor<int32, [1]> input_1245_strides_0 = const()[name = tensor<string, []>("input_1245_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_1245_pad_0 = const()[name = tensor<string, []>("input_1245_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_1245_dilations_0 = const()[name = tensor<string, []>("input_1245_dilations_0"), val = tensor<int32, [1]>([1])];
-            tensor<fp16, [1024, 1, 9]> const_294_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_294_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581685824))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581695104))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
-            tensor<fp16, [1024]> const_295_to_fp16 = const()[name = tensor<string, []>("const_295_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581697216)))];
+            tensor<fp16, [1024, 1, 9]> const_294_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("const_294_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581636672))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581645952))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024]> const_295_to_fp16 = const()[name = tensor<string, []>("const_295_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581648064)))];
             tensor<fp16, [1, 1024, 126]> input_1247_cast_fp16 = conv(bias = const_295_to_fp16, dilations = input_1245_dilations_0, groups = input_1245_groups_0, pad = input_1245_pad_0, pad_type = input_1245_pad_type_0, strides = input_1245_strides_0, weight = const_294_to_fp16_quantized, x = input_1243_cast_fp16)[name = tensor<string, []>("input_1247_cast_fp16")];
             tensor<fp16, [1, 1024, 126]> input_1249_cast_fp16 = silu(x = input_1247_cast_fp16)[name = tensor<string, []>("input_1249_cast_fp16")];
             tensor<string, []> x_527_pad_type_0 = const()[name = tensor<string, []>("x_527_pad_type_0"), val = tensor<string, []>("valid")];
@@ -3375,26 +3375,26 @@ program(1.0)
             tensor<int32, [2]> x_527_pad_0 = const()[name = tensor<string, []>("x_527_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> x_527_dilations_0 = const()[name = tensor<string, []>("x_527_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> x_527_groups_0 = const()[name = tensor<string, []>("x_527_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp16, [1024, 1024, 1]> model_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581699328))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582747968))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 1024, 1]> model_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(581650176))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582698816))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 1024, 126]> x_527_cast_fp16 = conv(dilations = x_527_dilations_0, groups = x_527_groups_0, pad = x_527_pad_0, pad_type = x_527_pad_type_0, strides = x_527_strides_0, weight = model_layers_23_conv_pointwise_conv2_weight_to_fp16_quantized, x = input_1249_cast_fp16)[name = tensor<string, []>("x_527_cast_fp16")];
             tensor<int32, [3]> input_1251_perm_0 = const()[name = tensor<string, []>("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<fp16, [1, 126, 1024]> input_1251_cast_fp16 = transpose(perm = input_1251_perm_0, x = x_527_cast_fp16)[name = tensor<string, []>("transpose_144")];
             tensor<fp16, [1, 126, 1024]> input_1253_cast_fp16 = add(x = input_1235_cast_fp16, y = input_1251_cast_fp16)[name = tensor<string, []>("input_1253_cast_fp16")];
             tensor<int32, [1]> input_1255_axes_0 = const()[name = tensor<string, []>("input_1255_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582750080)))];
-            tensor<fp16, [1024]> model_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582752192)))];
+            tensor<fp16, [1024]> model_layers_23_norm_feed_forward2_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582700928)))];
+            tensor<fp16, [1024]> model_layers_23_norm_feed_forward2_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_feed_forward2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582703040)))];
             tensor<fp16, [1, 126, 1024]> input_1255_cast_fp16 = layer_norm(axes = input_1255_axes_0, beta = model_layers_23_norm_feed_forward2_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_23_norm_feed_forward2_weight_to_fp16, x = input_1253_cast_fp16)[name = tensor<string, []>("input_1255_cast_fp16")];
-            tensor<fp16, [4096, 1024]> model_layers_23_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582754304))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586948672))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
+            tensor<fp16, [4096, 1024]> model_layers_23_feed_forward2_linear1_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward2_linear1_weight_to_fp16_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(582705152))), scale = tensor<fp16, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586899520))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8542720)))];
             tensor<fp16, [1, 126, 4096]> linear_215_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_feed_forward2_linear1_weight_to_fp16_quantized, x = input_1255_cast_fp16)[name = tensor<string, []>("linear_215_cast_fp16")];
             tensor<fp16, [1, 126, 4096]> input_1259_cast_fp16 = silu(x = linear_215_cast_fp16)[name = tensor<string, []>("input_1259_cast_fp16")];
-            tensor<fp16, [1024, 4096]> model_layers_23_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586956928))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591151296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
+            tensor<fp16, [1024, 4096]> model_layers_23_feed_forward2_linear2_weight_to_fp16_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("model_layers_23_feed_forward2_linear2_weight_to_fp16_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(586907776))), scale = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591102144))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4338816)))];
             tensor<fp16, [1, 126, 1024]> linear_216_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = model_layers_23_feed_forward2_linear2_weight_to_fp16_quantized, x = input_1259_cast_fp16)[name = tensor<string, []>("linear_216_cast_fp16")];
             tensor<fp16, []> var_4110_to_fp16 = const()[name = tensor<string, []>("op_4110_to_fp16"), val = tensor<fp16, []>(0x1p-1)];
             tensor<fp16, [1, 126, 1024]> var_4111_cast_fp16 = mul(x = linear_216_cast_fp16, y = var_4110_to_fp16)[name = tensor<string, []>("op_4111_cast_fp16")];
             tensor<fp16, [1, 126, 1024]> input_cast_fp16 = add(x = input_1253_cast_fp16, y = var_4111_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
             tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
-            tensor<fp16, [1024]> model_layers_23_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591153408)))];
-            tensor<fp16, [1024]> model_layers_23_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591155520)))];
+            tensor<fp16, [1024]> model_layers_23_norm_out_weight_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_out_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591104256)))];
+            tensor<fp16, [1024]> model_layers_23_norm_out_bias_to_fp16 = const()[name = tensor<string, []>("model_layers_23_norm_out_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591106368)))];
             tensor<fp16, [1, 126, 1024]> var_4124_cast_fp16 = layer_norm(axes = audio_signal_axes_0, beta = model_layers_23_norm_out_bias_to_fp16, epsilon = var_4_to_fp16, gamma = model_layers_23_norm_out_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("op_4124_cast_fp16")];
             tensor<string, []> var_4124_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4124_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
             tensor<fp32, [1, 126, 1024]> encoder_output = cast(dtype = var_4124_cast_fp16_to_fp32_dtype_0, x = var_4124_cast_fp16)[name = tensor<string, []>("cast_0")];