diff --git "a/nvidia_parakeet-v2_476MB/AudioEncoder.mlmodelc/model.mil" "b/nvidia_parakeet-v2_476MB/AudioEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/nvidia_parakeet-v2_476MB/AudioEncoder.mlmodelc/model.mil" @@ -0,0 +1,4395 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}})] +{ + func main(tensor input_1, tensor melspectrogram_features) { + tensor pos_emb_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288128))), name = tensor("pos_emb_to_fp16_palettized"), shape = tensor([1, 1024, 1, 375])]; + tensor var_55_cast_fp16 = sub(x = pos_emb_to_fp16_palettized, y = input_1)[name = tensor("op_55_cast_fp16")]; + tensor obj_3_cast_fp16 = mul(x = var_55_cast_fp16, y = input_1)[name = tensor("obj_3_cast_fp16")]; + tensor input_1_pad_type_0 = const()[name = tensor("input_1_pad_type_0"), val = tensor("custom")]; + tensor input_1_pad_0 = const()[name = tensor("input_1_pad_0"), val = tensor([1, 1, 1, 1])]; + tensor input_1_strides_0 = const()[name = tensor("input_1_strides_0"), val = tensor([2, 2])]; + tensor input_1_dilations_0 = const()[name = tensor("input_1_dilations_0"), val = tensor([1, 1])]; + tensor input_1_groups_0 = const()[name = tensor("input_1_groups_0"), val = tensor(1)]; + tensor pre_encode_conv_0_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290112))), name = tensor("pre_encode_conv_0_weight_to_fp16_palettized"), shape = tensor([256, 1, 3, 3])]; + tensor pre_encode_conv_0_bias_to_fp16 = const()[name = tensor("pre_encode_conv_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290304)))]; + tensor input_1_cast_fp16 = conv(bias = pre_encode_conv_0_bias_to_fp16, dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = pre_encode_conv_0_weight_to_fp16_palettized, x = melspectrogram_features)[name = tensor("input_1_cast_fp16")]; + tensor input_3_cast_fp16 = relu(x = input_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; + tensor input_5_pad_type_0 = const()[name = tensor("input_5_pad_type_0"), val = tensor("custom")]; + tensor input_5_pad_0 = const()[name = tensor("input_5_pad_0"), val = tensor([1, 1, 1, 1])]; + tensor input_5_strides_0 = const()[name = tensor("input_5_strides_0"), val = tensor([2, 2])]; + tensor input_5_groups_0 = const()[name = tensor("input_5_groups_0"), val = tensor(256)]; + tensor input_5_dilations_0 = const()[name = tensor("input_5_dilations_0"), val = tensor([1, 1])]; + tensor pre_encode_conv_2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292672))), name = tensor("pre_encode_conv_2_weight_to_fp16_palettized"), shape = tensor([256, 1, 3, 3])]; + tensor pre_encode_conv_2_bias_to_fp16 = const()[name = tensor("pre_encode_conv_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292864)))]; + tensor input_5_cast_fp16 = conv(bias = pre_encode_conv_2_bias_to_fp16, dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = pre_encode_conv_2_weight_to_fp16_palettized, x = input_3_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor input_7_pad_type_0 = const()[name = tensor("input_7_pad_type_0"), val = tensor("valid")]; + tensor input_7_strides_0 = const()[name = tensor("input_7_strides_0"), val = tensor([1, 1])]; + tensor input_7_pad_0 = const()[name = tensor("input_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_7_dilations_0 = const()[name = tensor("input_7_dilations_0"), val = tensor([1, 1])]; + tensor input_7_groups_0 = const()[name = tensor("input_7_groups_0"), val = tensor(1)]; + tensor pre_encode_conv_3_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342656))), name = tensor("pre_encode_conv_3_weight_to_fp16_palettized"), shape = tensor([256, 256, 1, 1])]; + tensor pre_encode_conv_3_bias_to_fp16 = const()[name = tensor("pre_encode_conv_3_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(342848)))]; + tensor input_7_cast_fp16 = conv(bias = pre_encode_conv_3_bias_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = pre_encode_conv_3_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_cast_fp16 = relu(x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_pad_type_0 = const()[name = tensor("input_11_pad_type_0"), val = tensor("custom")]; + tensor input_11_pad_0 = const()[name = tensor("input_11_pad_0"), val = tensor([1, 1, 1, 1])]; + tensor input_11_strides_0 = const()[name = tensor("input_11_strides_0"), val = tensor([2, 2])]; + tensor input_11_groups_0 = const()[name = tensor("input_11_groups_0"), val = tensor(256)]; + tensor input_11_dilations_0 = const()[name = tensor("input_11_dilations_0"), val = tensor([1, 1])]; + tensor pre_encode_conv_5_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345216))), name = tensor("pre_encode_conv_5_weight_to_fp16_palettized"), shape = tensor([256, 1, 3, 3])]; + tensor pre_encode_conv_5_bias_to_fp16 = const()[name = tensor("pre_encode_conv_5_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345408)))]; + tensor input_11_cast_fp16 = conv(bias = pre_encode_conv_5_bias_to_fp16, dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = pre_encode_conv_5_weight_to_fp16_palettized, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor input_13_pad_type_0 = const()[name = tensor("input_13_pad_type_0"), val = tensor("valid")]; + tensor input_13_strides_0 = const()[name = tensor("input_13_strides_0"), val = tensor([1, 1])]; + tensor input_13_pad_0 = const()[name = tensor("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_13_dilations_0 = const()[name = tensor("input_13_dilations_0"), val = tensor([1, 1])]; + tensor input_13_groups_0 = const()[name = tensor("input_13_groups_0"), val = tensor(1)]; + tensor pre_encode_conv_6_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395200))), name = tensor("pre_encode_conv_6_weight_to_fp16_palettized"), shape = tensor([256, 256, 1, 1])]; + tensor pre_encode_conv_6_bias_to_fp16 = const()[name = tensor("pre_encode_conv_6_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395392)))]; + tensor input_13_cast_fp16 = conv(bias = pre_encode_conv_6_bias_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = pre_encode_conv_6_weight_to_fp16_palettized, x = input_11_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_1_cast_fp16 = relu(x = input_13_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor var_115_perm_0 = const()[name = tensor("op_115_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor var_118 = const()[name = tensor("op_118"), val = tensor([1, 4096, 1, 188])]; + tensor var_115_cast_fp16 = transpose(perm = var_115_perm_0, x = x_1_cast_fp16)[name = tensor("transpose_0")]; + tensor input_15_cast_fp16 = reshape(shape = var_118, x = var_115_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor inputs_1_pad_type_0 = const()[name = tensor("inputs_1_pad_type_0"), val = tensor("valid")]; + tensor inputs_1_strides_0 = const()[name = tensor("inputs_1_strides_0"), val = tensor([1, 1])]; + tensor inputs_1_pad_0 = const()[name = tensor("inputs_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor inputs_1_dilations_0 = const()[name = tensor("inputs_1_dilations_0"), val = tensor([1, 1])]; + tensor inputs_1_groups_0 = const()[name = tensor("inputs_1_groups_0"), val = tensor(1)]; + tensor pre_encode_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3541760))), name = tensor("pre_encode_out_weight_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor pre_encode_out_bias_to_fp16 = const()[name = tensor("pre_encode_out_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3541952)))]; + tensor inputs_1_cast_fp16 = conv(bias = pre_encode_out_bias_to_fp16, dilations = inputs_1_dilations_0, groups = inputs_1_groups_0, pad = inputs_1_pad_0, pad_type = inputs_1_pad_type_0, strides = inputs_1_strides_0, weight = pre_encode_out_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_131 = const()[name = tensor("op_131"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_162_to_fp16 = const()[name = tensor("op_162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_162_to_fp16, x = inputs_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor input_17_mean_0_to_fp16 = const()[name = tensor("input_17_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3544064)))]; + tensor input_17_variance_0_to_fp16 = const()[name = tensor("input_17_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3546176)))]; + tensor input_17_gamma_0_to_fp16 = const()[name = tensor("input_17_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3548288)))]; + tensor input_17_beta_0_to_fp16 = const()[name = tensor("input_17_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3550400)))]; + tensor input_17_epsilon_0_to_fp16 = const()[name = tensor("input_17_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = batch_norm(beta = input_17_beta_0_to_fp16, epsilon = input_17_epsilon_0_to_fp16, gamma = input_17_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; + tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; + tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; + tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; + tensor layers_0_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3552512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6698304))), name = tensor("layers_0_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor layers_0_feed_forward1_fc1_bias_to_fp16 = const()[name = tensor("layers_0_feed_forward1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6698496)))]; + tensor input_19_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_0_feed_forward1_fc1_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor x_3_pad_type_0 = const()[name = tensor("x_3_pad_type_0"), val = tensor("valid")]; + tensor x_3_strides_0 = const()[name = tensor("x_3_strides_0"), val = tensor([1, 1])]; + tensor x_3_pad_0 = const()[name = tensor("x_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_3_dilations_0 = const()[name = tensor("x_3_dilations_0"), val = tensor([1, 1])]; + tensor x_3_groups_0 = const()[name = tensor("x_3_groups_0"), val = tensor(1)]; + tensor op_190_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6706752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9852544))), name = tensor("op_190_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_190_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_3_dilations_0, groups = x_3_groups_0, pad = x_3_pad_0, pad_type = x_3_pad_type_0, strides = x_3_strides_0, weight = op_190_weight_0_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor("op_190_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_190_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([1])]; + tensor var_200_to_fp16 = const()[name = tensor("op_200_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_200_to_fp16, x = inputs_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; + tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9852736)))]; + tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9854848)))]; + tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_1_cast_fp16")]; + tensor query_1_pad_type_0 = const()[name = tensor("query_1_pad_type_0"), val = tensor("valid")]; + tensor query_1_strides_0 = const()[name = tensor("query_1_strides_0"), val = tensor([1, 1])]; + tensor query_1_pad_0 = const()[name = tensor("query_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_1_dilations_0 = const()[name = tensor("query_1_dilations_0"), val = tensor([1, 1])]; + tensor query_1_groups_0 = const()[name = tensor("query_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9856960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10643456))), name = tensor("layers_0_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_1_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_1_dilations_0, groups = query_1_groups_0, pad = query_1_pad_0, pad_type = query_1_pad_type_0, strides = query_1_strides_0, weight = layers_0_self_attn_q_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; + tensor key_1_pad_type_0 = const()[name = tensor("key_1_pad_type_0"), val = tensor("valid")]; + tensor key_1_strides_0 = const()[name = tensor("key_1_strides_0"), val = tensor([1, 1])]; + tensor key_1_pad_0 = const()[name = tensor("key_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_1_dilations_0 = const()[name = tensor("key_1_dilations_0"), val = tensor([1, 1])]; + tensor key_1_groups_0 = const()[name = tensor("key_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10643648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11430144))), name = tensor("layers_0_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_1_cast_fp16 = conv(dilations = key_1_dilations_0, groups = key_1_groups_0, pad = key_1_pad_0, pad_type = key_1_pad_type_0, strides = key_1_strides_0, weight = layers_0_self_attn_k_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("key_1_cast_fp16")]; + tensor value_1_pad_type_0 = const()[name = tensor("value_1_pad_type_0"), val = tensor("valid")]; + tensor value_1_strides_0 = const()[name = tensor("value_1_strides_0"), val = tensor([1, 1])]; + tensor value_1_pad_0 = const()[name = tensor("value_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_1_dilations_0 = const()[name = tensor("value_1_dilations_0"), val = tensor([1, 1])]; + tensor value_1_groups_0 = const()[name = tensor("value_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11430336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12216832))), name = tensor("layers_0_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_1_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_1_dilations_0, groups = value_1_groups_0, pad = value_1_pad_0, pad_type = value_1_pad_type_0, strides = value_1_strides_0, weight = layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_238_to_fp16 = const()[name = tensor("op_238_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12217024)))]; + tensor query_3_cast_fp16 = add(x = query_1_cast_fp16, y = var_238_to_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_241_to_fp16 = const()[name = tensor("op_241_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12219136)))]; + tensor q_with_bias_v_1_cast_fp16 = add(x = query_1_cast_fp16, y = var_241_to_fp16)[name = tensor("q_with_bias_v_1_cast_fp16")]; + tensor p_1_pad_type_0 = const()[name = tensor("p_1_pad_type_0"), val = tensor("valid")]; + tensor p_1_strides_0 = const()[name = tensor("p_1_strides_0"), val = tensor([1, 1])]; + tensor p_1_pad_0 = const()[name = tensor("p_1_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_1_dilations_0 = const()[name = tensor("p_1_dilations_0"), val = tensor([1, 1])]; + tensor p_1_groups_0 = const()[name = tensor("p_1_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12221248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13007744))), name = tensor("layers_0_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_1_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_1_dilations_0, groups = p_1_groups_0, pad = p_1_pad_0, pad_type = p_1_pad_type_0, strides = p_1_strides_0, weight = layers_0_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_1_cast_fp16")]; + tensor var_252 = const()[name = tensor("op_252"), val = tensor([1, 8, 128, 188])]; + tensor var_253_cast_fp16 = reshape(shape = var_252, x = q_with_bias_v_1_cast_fp16)[name = tensor("op_253_cast_fp16")]; + tensor var_254 = const()[name = tensor("op_254"), val = tensor([1, 8, 128, -1])]; + tensor var_255_cast_fp16 = reshape(shape = var_254, x = p_1_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor matrix_bd_1_transpose_x_0 = const()[name = tensor("matrix_bd_1_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_1_transpose_y_0 = const()[name = tensor("matrix_bd_1_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_1_cast_fp16 = matmul(transpose_x = matrix_bd_1_transpose_x_0, transpose_y = matrix_bd_1_transpose_y_0, x = var_253_cast_fp16, y = var_255_cast_fp16)[name = tensor("matrix_bd_1_cast_fp16")]; + tensor matrix_bd_3_pad_0 = const()[name = tensor("matrix_bd_3_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_3_mode_0 = const()[name = tensor("matrix_bd_3_mode_0"), val = tensor("constant")]; + tensor const_10_to_fp16 = const()[name = tensor("const_10_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_3_cast_fp16 = pad(constant_val = const_10_to_fp16, mode = matrix_bd_3_mode_0, pad = matrix_bd_3_pad_0, x = matrix_bd_1_cast_fp16)[name = tensor("matrix_bd_3_cast_fp16")]; + tensor var_264 = const()[name = tensor("op_264"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_264, x = matrix_bd_3_cast_fp16)[name = tensor("matrix_bd_5_cast_fp16")]; + tensor var_268_begin_0 = const()[name = tensor("op_268_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_268_end_0 = const()[name = tensor("op_268_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_268_end_mask_0 = const()[name = tensor("op_268_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_268_cast_fp16 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor("op_268_cast_fp16")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_7_cast_fp16 = reshape(shape = var_269, x = var_268_cast_fp16)[name = tensor("matrix_bd_7_cast_fp16")]; + tensor var_274_begin_0 = const()[name = tensor("op_274_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_274_end_0 = const()[name = tensor("op_274_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_274_end_mask_0 = const()[name = tensor("op_274_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_274_cast_fp16 = slice_by_index(begin = var_274_begin_0, end = var_274_end_0, end_mask = var_274_end_mask_0, x = matrix_bd_7_cast_fp16)[name = tensor("op_274_cast_fp16")]; + tensor var_275_to_fp16 = const()[name = tensor("op_275_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_1_cast_fp16 = mul(x = var_274_cast_fp16, y = var_275_to_fp16)[name = tensor("qk_mask_1_cast_fp16")]; + tensor var_279 = const()[name = tensor("op_279"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_279, x = query_3_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_281_to_fp16 = const()[name = tensor("op_281_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_282_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_281_to_fp16)[name = tensor("op_282_cast_fp16")]; + tensor var_285 = const()[name = tensor("op_285"), val = tensor([1, 8, 128, 188])]; + tensor var_286_cast_fp16 = reshape(shape = var_285, x = key_1_cast_fp16)[name = tensor("op_286_cast_fp16")]; + tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; + tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; + tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_282_cast_fp16, y = var_286_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; + tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = qk_mask_1_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; + tensor var_290_cast_fp16 = softmax(axis = var_131, x = mh_w_3_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 8, 128, 188])]; + tensor var_292_cast_fp16 = reshape(shape = var_291, x = value_1_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; + tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; + tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_292_cast_fp16, y = var_290_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_295 = const()[name = tensor("op_295"), val = tensor([1, 1024, 1, 188])]; + tensor input_23_cast_fp16 = reshape(shape = var_295, x = attn_1_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor obj_5_pad_type_0 = const()[name = tensor("obj_5_pad_type_0"), val = tensor("valid")]; + tensor obj_5_strides_0 = const()[name = tensor("obj_5_strides_0"), val = tensor([1, 1])]; + tensor obj_5_pad_0 = const()[name = tensor("obj_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_5_dilations_0 = const()[name = tensor("obj_5_dilations_0"), val = tensor([1, 1])]; + tensor obj_5_groups_0 = const()[name = tensor("obj_5_groups_0"), val = tensor(1)]; + tensor layers_0_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13007936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13794432))), name = tensor("layers_0_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_5_dilations_0, groups = obj_5_groups_0, pad = obj_5_pad_0, pad_type = obj_5_pad_type_0, strides = obj_5_strides_0, weight = layers_0_self_attn_o_proj_weight_to_fp16_palettized, x = input_23_cast_fp16)[name = tensor("obj_5_cast_fp16")]; + tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_5_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; + tensor out_5_axes_0 = const()[name = tensor("out_5_axes_0"), val = tensor([1])]; + tensor var_313_to_fp16 = const()[name = tensor("op_313_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_313_to_fp16, x = inputs_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_25_gamma_0_to_fp16 = const()[name = tensor("input_25_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13794624)))]; + tensor input_25_beta_0_to_fp16 = const()[name = tensor("input_25_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13796736)))]; + tensor input_25_epsilon_0_to_fp16 = const()[name = tensor("input_25_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_25_cast_fp16 = batch_norm(beta = input_25_beta_0_to_fp16, epsilon = input_25_epsilon_0_to_fp16, gamma = input_25_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_pad_type_0 = const()[name = tensor("input_27_pad_type_0"), val = tensor("valid")]; + tensor input_27_strides_0 = const()[name = tensor("input_27_strides_0"), val = tensor([1, 1])]; + tensor input_27_pad_0 = const()[name = tensor("input_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_27_dilations_0 = const()[name = tensor("input_27_dilations_0"), val = tensor([1, 1])]; + tensor input_27_groups_0 = const()[name = tensor("input_27_groups_0"), val = tensor(1)]; + tensor layers_0_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13798848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15371776))), name = tensor("layers_0_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_27_cast_fp16 = conv(dilations = input_27_dilations_0, groups = input_27_groups_0, pad = input_27_pad_0, pad_type = input_27_pad_type_0, strides = input_27_strides_0, weight = layers_0_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_split_num_splits_0 = const()[name = tensor("input_29_split_num_splits_0"), val = tensor(2)]; + tensor input_29_split_axis_0 = const()[name = tensor("input_29_split_axis_0"), val = tensor(1)]; + tensor input_29_split_cast_fp16_0, tensor input_29_split_cast_fp16_1 = split(axis = input_29_split_axis_0, num_splits = input_29_split_num_splits_0, x = input_27_cast_fp16)[name = tensor("input_29_split_cast_fp16")]; + tensor input_29_split_1_sigmoid_cast_fp16 = sigmoid(x = input_29_split_cast_fp16_1)[name = tensor("input_29_split_1_sigmoid_cast_fp16")]; + tensor input_29_cast_fp16 = mul(x = input_29_split_cast_fp16_0, y = input_29_split_1_sigmoid_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("custom")]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1024)]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1, 1])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1, 1])]; + tensor const_268_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15371968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15378944))), name = tensor("const_268_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_269_to_fp16 = const()[name = tensor("const_269_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15379136)))]; + tensor input_33_cast_fp16 = conv(bias = const_269_to_fp16, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = const_268_to_fp16_palettized, x = input_29_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_cast_fp16 = silu(x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor x_5_pad_type_0 = const()[name = tensor("x_5_pad_type_0"), val = tensor("valid")]; + tensor x_5_strides_0 = const()[name = tensor("x_5_strides_0"), val = tensor([1, 1])]; + tensor x_5_pad_0 = const()[name = tensor("x_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_5_dilations_0 = const()[name = tensor("x_5_dilations_0"), val = tensor([1, 1])]; + tensor x_5_groups_0 = const()[name = tensor("x_5_groups_0"), val = tensor(1)]; + tensor layers_0_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15381248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16167744))), name = tensor("layers_0_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_5_cast_fp16 = conv(dilations = x_5_dilations_0, groups = x_5_groups_0, pad = x_5_pad_0, pad_type = x_5_pad_type_0, strides = x_5_strides_0, weight = layers_0_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_35_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = x_5_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; + tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; + tensor var_361_to_fp16 = const()[name = tensor("op_361_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_361_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor input_37_gamma_0_to_fp16 = const()[name = tensor("input_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16167936)))]; + tensor input_37_beta_0_to_fp16 = const()[name = tensor("input_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16170048)))]; + tensor input_37_epsilon_0_to_fp16 = const()[name = tensor("input_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_37_cast_fp16 = batch_norm(beta = input_37_beta_0_to_fp16, epsilon = input_37_epsilon_0_to_fp16, gamma = input_37_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; + tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; + tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; + tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; + tensor layers_0_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16172160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19317952))), name = tensor("layers_0_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_39_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = layers_0_feed_forward2_fc1_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_cast_fp16 = silu(x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor x_7_pad_type_0 = const()[name = tensor("x_7_pad_type_0"), val = tensor("valid")]; + tensor x_7_strides_0 = const()[name = tensor("x_7_strides_0"), val = tensor([1, 1])]; + tensor x_7_pad_0 = const()[name = tensor("x_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_7_dilations_0 = const()[name = tensor("x_7_dilations_0"), val = tensor([1, 1])]; + tensor x_7_groups_0 = const()[name = tensor("x_7_groups_0"), val = tensor(1)]; + tensor op_389_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19318144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22463936))), name = tensor("op_389_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_389_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_7_dilations_0, groups = x_7_groups_0, pad = x_7_pad_0, pad_type = x_7_pad_type_0, strides = x_7_strides_0, weight = op_389_weight_0_to_fp16_palettized, x = input_41_cast_fp16)[name = tensor("op_389_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_389_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_399_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; + tensor inputs_11_gamma_0_to_fp16 = const()[name = tensor("inputs_11_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22464128)))]; + tensor inputs_11_beta_0_to_fp16 = const()[name = tensor("inputs_11_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22466240)))]; + tensor inputs_11_epsilon_0_to_fp16 = const()[name = tensor("inputs_11_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_11_cast_fp16 = batch_norm(beta = inputs_11_beta_0_to_fp16, epsilon = inputs_11_epsilon_0_to_fp16, gamma = inputs_11_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_413 = const()[name = tensor("op_413"), val = tensor(3)]; + tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; + tensor var_444_to_fp16 = const()[name = tensor("op_444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_444_to_fp16, x = inputs_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_43_gamma_0_to_fp16 = const()[name = tensor("input_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22468352)))]; + tensor input_43_beta_0_to_fp16 = const()[name = tensor("input_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22470464)))]; + tensor input_43_epsilon_0_to_fp16 = const()[name = tensor("input_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_43_cast_fp16 = batch_norm(beta = input_43_beta_0_to_fp16, epsilon = input_43_epsilon_0_to_fp16, gamma = input_43_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_pad_type_0 = const()[name = tensor("input_45_pad_type_0"), val = tensor("valid")]; + tensor input_45_strides_0 = const()[name = tensor("input_45_strides_0"), val = tensor([1, 1])]; + tensor input_45_pad_0 = const()[name = tensor("input_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_45_dilations_0 = const()[name = tensor("input_45_dilations_0"), val = tensor([1, 1])]; + tensor input_45_groups_0 = const()[name = tensor("input_45_groups_0"), val = tensor(1)]; + tensor layers_1_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22472576))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25618368))), name = tensor("layers_1_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_45_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_1_feed_forward1_fc1_weight_to_fp16_palettized, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_cast_fp16 = silu(x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor x_9_pad_type_0 = const()[name = tensor("x_9_pad_type_0"), val = tensor("valid")]; + tensor x_9_strides_0 = const()[name = tensor("x_9_strides_0"), val = tensor([1, 1])]; + tensor x_9_pad_0 = const()[name = tensor("x_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_9_dilations_0 = const()[name = tensor("x_9_dilations_0"), val = tensor([1, 1])]; + tensor x_9_groups_0 = const()[name = tensor("x_9_groups_0"), val = tensor(1)]; + tensor op_472_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25618560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28764352))), name = tensor("op_472_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_472_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_9_dilations_0, groups = x_9_groups_0, pad = x_9_pad_0, pad_type = x_9_pad_type_0, strides = x_9_strides_0, weight = op_472_weight_0_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_472_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; + tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; + tensor var_482_to_fp16 = const()[name = tensor("op_482_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_482_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; + tensor obj_7_gamma_0_to_fp16 = const()[name = tensor("obj_7_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28764544)))]; + tensor obj_7_beta_0_to_fp16 = const()[name = tensor("obj_7_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28766656)))]; + tensor obj_7_epsilon_0_to_fp16 = const()[name = tensor("obj_7_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_7_cast_fp16 = batch_norm(beta = obj_7_beta_0_to_fp16, epsilon = obj_7_epsilon_0_to_fp16, gamma = obj_7_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_7_cast_fp16")]; + tensor query_5_pad_type_0 = const()[name = tensor("query_5_pad_type_0"), val = tensor("valid")]; + tensor query_5_strides_0 = const()[name = tensor("query_5_strides_0"), val = tensor([1, 1])]; + tensor query_5_pad_0 = const()[name = tensor("query_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_5_dilations_0 = const()[name = tensor("query_5_dilations_0"), val = tensor([1, 1])]; + tensor query_5_groups_0 = const()[name = tensor("query_5_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28768768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29555264))), name = tensor("layers_1_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_5_dilations_0, groups = query_5_groups_0, pad = query_5_pad_0, pad_type = query_5_pad_type_0, strides = query_5_strides_0, weight = layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = obj_7_cast_fp16)[name = tensor("query_5_cast_fp16")]; + tensor key_3_pad_type_0 = const()[name = tensor("key_3_pad_type_0"), val = tensor("valid")]; + tensor key_3_strides_0 = const()[name = tensor("key_3_strides_0"), val = tensor([1, 1])]; + tensor key_3_pad_0 = const()[name = tensor("key_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_3_dilations_0 = const()[name = tensor("key_3_dilations_0"), val = tensor([1, 1])]; + tensor key_3_groups_0 = const()[name = tensor("key_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29555456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30341952))), name = tensor("layers_1_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_3_cast_fp16 = conv(dilations = key_3_dilations_0, groups = key_3_groups_0, pad = key_3_pad_0, pad_type = key_3_pad_type_0, strides = key_3_strides_0, weight = layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = obj_7_cast_fp16)[name = tensor("key_3_cast_fp16")]; + tensor value_3_pad_type_0 = const()[name = tensor("value_3_pad_type_0"), val = tensor("valid")]; + tensor value_3_strides_0 = const()[name = tensor("value_3_strides_0"), val = tensor([1, 1])]; + tensor value_3_pad_0 = const()[name = tensor("value_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_3_dilations_0 = const()[name = tensor("value_3_dilations_0"), val = tensor([1, 1])]; + tensor value_3_groups_0 = const()[name = tensor("value_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30342144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31128640))), name = tensor("layers_1_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_3_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_3_dilations_0, groups = value_3_groups_0, pad = value_3_pad_0, pad_type = value_3_pad_type_0, strides = value_3_strides_0, weight = layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = obj_7_cast_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_520_to_fp16 = const()[name = tensor("op_520_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31128832)))]; + tensor query_7_cast_fp16 = add(x = query_5_cast_fp16, y = var_520_to_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_523_to_fp16 = const()[name = tensor("op_523_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31130944)))]; + tensor q_with_bias_v_3_cast_fp16 = add(x = query_5_cast_fp16, y = var_523_to_fp16)[name = tensor("q_with_bias_v_3_cast_fp16")]; + tensor p_3_pad_type_0 = const()[name = tensor("p_3_pad_type_0"), val = tensor("valid")]; + tensor p_3_strides_0 = const()[name = tensor("p_3_strides_0"), val = tensor([1, 1])]; + tensor p_3_pad_0 = const()[name = tensor("p_3_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_3_dilations_0 = const()[name = tensor("p_3_dilations_0"), val = tensor([1, 1])]; + tensor p_3_groups_0 = const()[name = tensor("p_3_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31133056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31919552))), name = tensor("layers_1_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_3_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_3_dilations_0, groups = p_3_groups_0, pad = p_3_pad_0, pad_type = p_3_pad_type_0, strides = p_3_strides_0, weight = layers_1_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_3_cast_fp16")]; + tensor var_534 = const()[name = tensor("op_534"), val = tensor([1, 8, 128, 188])]; + tensor var_535_cast_fp16 = reshape(shape = var_534, x = q_with_bias_v_3_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor var_536 = const()[name = tensor("op_536"), val = tensor([1, 8, 128, -1])]; + tensor var_537_cast_fp16 = reshape(shape = var_536, x = p_3_cast_fp16)[name = tensor("op_537_cast_fp16")]; + tensor matrix_bd_9_transpose_x_0 = const()[name = tensor("matrix_bd_9_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_9_transpose_y_0 = const()[name = tensor("matrix_bd_9_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_9_cast_fp16 = matmul(transpose_x = matrix_bd_9_transpose_x_0, transpose_y = matrix_bd_9_transpose_y_0, x = var_535_cast_fp16, y = var_537_cast_fp16)[name = tensor("matrix_bd_9_cast_fp16")]; + tensor matrix_bd_11_pad_0 = const()[name = tensor("matrix_bd_11_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_11_mode_0 = const()[name = tensor("matrix_bd_11_mode_0"), val = tensor("constant")]; + tensor const_21_to_fp16 = const()[name = tensor("const_21_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_11_cast_fp16 = pad(constant_val = const_21_to_fp16, mode = matrix_bd_11_mode_0, pad = matrix_bd_11_pad_0, x = matrix_bd_9_cast_fp16)[name = tensor("matrix_bd_11_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_546, x = matrix_bd_11_cast_fp16)[name = tensor("matrix_bd_13_cast_fp16")]; + tensor var_550_begin_0 = const()[name = tensor("op_550_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_550_end_0 = const()[name = tensor("op_550_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_550_end_mask_0 = const()[name = tensor("op_550_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_550_cast_fp16 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor("op_550_cast_fp16")]; + tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_15_cast_fp16 = reshape(shape = var_551, x = var_550_cast_fp16)[name = tensor("matrix_bd_15_cast_fp16")]; + tensor var_556_begin_0 = const()[name = tensor("op_556_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_556_end_0 = const()[name = tensor("op_556_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_556_end_mask_0 = const()[name = tensor("op_556_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_556_cast_fp16 = slice_by_index(begin = var_556_begin_0, end = var_556_end_0, end_mask = var_556_end_mask_0, x = matrix_bd_15_cast_fp16)[name = tensor("op_556_cast_fp16")]; + tensor var_557_to_fp16 = const()[name = tensor("op_557_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_3_cast_fp16 = mul(x = var_556_cast_fp16, y = var_557_to_fp16)[name = tensor("qk_mask_3_cast_fp16")]; + tensor var_561 = const()[name = tensor("op_561"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_561, x = query_7_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_563_to_fp16 = const()[name = tensor("op_563_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_564_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_563_to_fp16)[name = tensor("op_564_cast_fp16")]; + tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 8, 128, 188])]; + tensor var_568_cast_fp16 = reshape(shape = var_567, x = key_3_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; + tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; + tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_564_cast_fp16, y = var_568_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; + tensor mh_w_7_cast_fp16 = add(x = mh_w_5_cast_fp16, y = qk_mask_3_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; + tensor var_572_cast_fp16 = softmax(axis = var_413, x = mh_w_7_cast_fp16)[name = tensor("op_572_cast_fp16")]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 8, 128, 188])]; + tensor var_574_cast_fp16 = reshape(shape = var_573, x = value_3_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; + tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; + tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_574_cast_fp16, y = var_572_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 1024, 1, 188])]; + tensor input_49_cast_fp16 = reshape(shape = var_577, x = attn_3_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor obj_9_pad_type_0 = const()[name = tensor("obj_9_pad_type_0"), val = tensor("valid")]; + tensor obj_9_strides_0 = const()[name = tensor("obj_9_strides_0"), val = tensor([1, 1])]; + tensor obj_9_pad_0 = const()[name = tensor("obj_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_9_dilations_0 = const()[name = tensor("obj_9_dilations_0"), val = tensor([1, 1])]; + tensor obj_9_groups_0 = const()[name = tensor("obj_9_groups_0"), val = tensor(1)]; + tensor layers_1_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31919744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32706240))), name = tensor("layers_1_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_9_dilations_0, groups = obj_9_groups_0, pad = obj_9_pad_0, pad_type = obj_9_pad_type_0, strides = obj_9_strides_0, weight = layers_1_self_attn_o_proj_weight_to_fp16_palettized, x = input_49_cast_fp16)[name = tensor("obj_9_cast_fp16")]; + tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_9_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; + tensor var_595_to_fp16 = const()[name = tensor("op_595_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_595_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor input_51_gamma_0_to_fp16 = const()[name = tensor("input_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32706432)))]; + tensor input_51_beta_0_to_fp16 = const()[name = tensor("input_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32708544)))]; + tensor input_51_epsilon_0_to_fp16 = const()[name = tensor("input_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1, 1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1, 1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor layers_1_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32710656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34283584))), name = tensor("layers_1_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_53_cast_fp16 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_1_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor input_55_split_num_splits_0 = const()[name = tensor("input_55_split_num_splits_0"), val = tensor(2)]; + tensor input_55_split_axis_0 = const()[name = tensor("input_55_split_axis_0"), val = tensor(1)]; + tensor input_55_split_cast_fp16_0, tensor input_55_split_cast_fp16_1 = split(axis = input_55_split_axis_0, num_splits = input_55_split_num_splits_0, x = input_53_cast_fp16)[name = tensor("input_55_split_cast_fp16")]; + tensor input_55_split_1_sigmoid_cast_fp16 = sigmoid(x = input_55_split_cast_fp16_1)[name = tensor("input_55_split_1_sigmoid_cast_fp16")]; + tensor input_55_cast_fp16 = mul(x = input_55_split_cast_fp16_0, y = input_55_split_1_sigmoid_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_pad_type_0 = const()[name = tensor("input_57_pad_type_0"), val = tensor("custom")]; + tensor input_57_pad_0 = const()[name = tensor("input_57_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_57_groups_0 = const()[name = tensor("input_57_groups_0"), val = tensor(1024)]; + tensor input_57_strides_0 = const()[name = tensor("input_57_strides_0"), val = tensor([1, 1])]; + tensor input_57_dilations_0 = const()[name = tensor("input_57_dilations_0"), val = tensor([1, 1])]; + tensor const_270_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34283776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34290752))), name = tensor("const_270_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_271_to_fp16 = const()[name = tensor("const_271_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34290944)))]; + tensor input_59_cast_fp16 = conv(bias = const_271_to_fp16, dilations = input_57_dilations_0, groups = input_57_groups_0, pad = input_57_pad_0, pad_type = input_57_pad_type_0, strides = input_57_strides_0, weight = const_270_to_fp16_palettized, x = input_55_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = input_59_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor x_11_pad_type_0 = const()[name = tensor("x_11_pad_type_0"), val = tensor("valid")]; + tensor x_11_strides_0 = const()[name = tensor("x_11_strides_0"), val = tensor([1, 1])]; + tensor x_11_pad_0 = const()[name = tensor("x_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_11_dilations_0 = const()[name = tensor("x_11_dilations_0"), val = tensor([1, 1])]; + tensor x_11_groups_0 = const()[name = tensor("x_11_groups_0"), val = tensor(1)]; + tensor layers_1_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34293056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35079552))), name = tensor("layers_1_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_11_cast_fp16 = conv(dilations = x_11_dilations_0, groups = x_11_groups_0, pad = x_11_pad_0, pad_type = x_11_pad_type_0, strides = x_11_strides_0, weight = layers_1_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_61_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = x_11_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; + tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; + tensor var_643_to_fp16 = const()[name = tensor("op_643_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_643_to_fp16, x = inputs_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_63_gamma_0_to_fp16 = const()[name = tensor("input_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35079744)))]; + tensor input_63_beta_0_to_fp16 = const()[name = tensor("input_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35081856)))]; + tensor input_63_epsilon_0_to_fp16 = const()[name = tensor("input_63_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = batch_norm(beta = input_63_beta_0_to_fp16, epsilon = input_63_epsilon_0_to_fp16, gamma = input_63_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_pad_type_0 = const()[name = tensor("input_65_pad_type_0"), val = tensor("valid")]; + tensor input_65_strides_0 = const()[name = tensor("input_65_strides_0"), val = tensor([1, 1])]; + tensor input_65_pad_0 = const()[name = tensor("input_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_65_dilations_0 = const()[name = tensor("input_65_dilations_0"), val = tensor([1, 1])]; + tensor input_65_groups_0 = const()[name = tensor("input_65_groups_0"), val = tensor(1)]; + tensor layers_1_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35083968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38229760))), name = tensor("layers_1_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_65_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_65_dilations_0, groups = input_65_groups_0, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = input_65_strides_0, weight = layers_1_feed_forward2_fc1_weight_to_fp16_palettized, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor input_67_cast_fp16 = silu(x = input_65_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor x_13_pad_type_0 = const()[name = tensor("x_13_pad_type_0"), val = tensor("valid")]; + tensor x_13_strides_0 = const()[name = tensor("x_13_strides_0"), val = tensor([1, 1])]; + tensor x_13_pad_0 = const()[name = tensor("x_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_13_dilations_0 = const()[name = tensor("x_13_dilations_0"), val = tensor([1, 1])]; + tensor x_13_groups_0 = const()[name = tensor("x_13_groups_0"), val = tensor(1)]; + tensor op_671_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38229952))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41375744))), name = tensor("op_671_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_671_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_13_dilations_0, groups = x_13_groups_0, pad = x_13_pad_0, pad_type = x_13_pad_type_0, strides = x_13_strides_0, weight = op_671_weight_0_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("op_671_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_671_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; + tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; + tensor var_681_to_fp16 = const()[name = tensor("op_681_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_681_to_fp16, x = inputs_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; + tensor inputs_21_gamma_0_to_fp16 = const()[name = tensor("inputs_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41375936)))]; + tensor inputs_21_beta_0_to_fp16 = const()[name = tensor("inputs_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41378048)))]; + tensor inputs_21_epsilon_0_to_fp16 = const()[name = tensor("inputs_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_21_cast_fp16 = batch_norm(beta = inputs_21_beta_0_to_fp16, epsilon = inputs_21_epsilon_0_to_fp16, gamma = inputs_21_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor(3)]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; + tensor var_726_to_fp16 = const()[name = tensor("op_726_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_726_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor input_69_gamma_0_to_fp16 = const()[name = tensor("input_69_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41380160)))]; + tensor input_69_beta_0_to_fp16 = const()[name = tensor("input_69_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41382272)))]; + tensor input_69_epsilon_0_to_fp16 = const()[name = tensor("input_69_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = batch_norm(beta = input_69_beta_0_to_fp16, epsilon = input_69_epsilon_0_to_fp16, gamma = input_69_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1, 1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1, 1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor layers_2_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41384384))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44530176))), name = tensor("layers_2_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_71_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = layers_2_feed_forward1_fc1_weight_to_fp16_palettized, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor input_73_cast_fp16 = silu(x = input_71_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_15_pad_type_0 = const()[name = tensor("x_15_pad_type_0"), val = tensor("valid")]; + tensor x_15_strides_0 = const()[name = tensor("x_15_strides_0"), val = tensor([1, 1])]; + tensor x_15_pad_0 = const()[name = tensor("x_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_15_dilations_0 = const()[name = tensor("x_15_dilations_0"), val = tensor([1, 1])]; + tensor x_15_groups_0 = const()[name = tensor("x_15_groups_0"), val = tensor(1)]; + tensor op_754_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44530368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47676160))), name = tensor("op_754_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_754_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_15_dilations_0, groups = x_15_groups_0, pad = x_15_pad_0, pad_type = x_15_pad_type_0, strides = x_15_strides_0, weight = op_754_weight_0_to_fp16_palettized, x = input_73_cast_fp16)[name = tensor("op_754_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_754_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; + tensor out_23_axes_0 = const()[name = tensor("out_23_axes_0"), val = tensor([1])]; + tensor var_764_to_fp16 = const()[name = tensor("op_764_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_764_to_fp16, x = inputs_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; + tensor obj_11_gamma_0_to_fp16 = const()[name = tensor("obj_11_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47676352)))]; + tensor obj_11_beta_0_to_fp16 = const()[name = tensor("obj_11_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47678464)))]; + tensor obj_11_epsilon_0_to_fp16 = const()[name = tensor("obj_11_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_11_cast_fp16 = batch_norm(beta = obj_11_beta_0_to_fp16, epsilon = obj_11_epsilon_0_to_fp16, gamma = obj_11_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("obj_11_cast_fp16")]; + tensor query_9_pad_type_0 = const()[name = tensor("query_9_pad_type_0"), val = tensor("valid")]; + tensor query_9_strides_0 = const()[name = tensor("query_9_strides_0"), val = tensor([1, 1])]; + tensor query_9_pad_0 = const()[name = tensor("query_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_9_dilations_0 = const()[name = tensor("query_9_dilations_0"), val = tensor([1, 1])]; + tensor query_9_groups_0 = const()[name = tensor("query_9_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(47680576))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48467072))), name = tensor("layers_2_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_9_dilations_0, groups = query_9_groups_0, pad = query_9_pad_0, pad_type = query_9_pad_type_0, strides = query_9_strides_0, weight = layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = obj_11_cast_fp16)[name = tensor("query_9_cast_fp16")]; + tensor key_5_pad_type_0 = const()[name = tensor("key_5_pad_type_0"), val = tensor("valid")]; + tensor key_5_strides_0 = const()[name = tensor("key_5_strides_0"), val = tensor([1, 1])]; + tensor key_5_pad_0 = const()[name = tensor("key_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_5_dilations_0 = const()[name = tensor("key_5_dilations_0"), val = tensor([1, 1])]; + tensor key_5_groups_0 = const()[name = tensor("key_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(48467264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49253760))), name = tensor("layers_2_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_5_cast_fp16 = conv(dilations = key_5_dilations_0, groups = key_5_groups_0, pad = key_5_pad_0, pad_type = key_5_pad_type_0, strides = key_5_strides_0, weight = layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = obj_11_cast_fp16)[name = tensor("key_5_cast_fp16")]; + tensor value_5_pad_type_0 = const()[name = tensor("value_5_pad_type_0"), val = tensor("valid")]; + tensor value_5_strides_0 = const()[name = tensor("value_5_strides_0"), val = tensor([1, 1])]; + tensor value_5_pad_0 = const()[name = tensor("value_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_5_dilations_0 = const()[name = tensor("value_5_dilations_0"), val = tensor([1, 1])]; + tensor value_5_groups_0 = const()[name = tensor("value_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(49253952))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50040448))), name = tensor("layers_2_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_5_dilations_0, groups = value_5_groups_0, pad = value_5_pad_0, pad_type = value_5_pad_type_0, strides = value_5_strides_0, weight = layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = obj_11_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_802_to_fp16 = const()[name = tensor("op_802_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50040640)))]; + tensor query_11_cast_fp16 = add(x = query_9_cast_fp16, y = var_802_to_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_805_to_fp16 = const()[name = tensor("op_805_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50042752)))]; + tensor q_with_bias_v_5_cast_fp16 = add(x = query_9_cast_fp16, y = var_805_to_fp16)[name = tensor("q_with_bias_v_5_cast_fp16")]; + tensor p_5_pad_type_0 = const()[name = tensor("p_5_pad_type_0"), val = tensor("valid")]; + tensor p_5_strides_0 = const()[name = tensor("p_5_strides_0"), val = tensor([1, 1])]; + tensor p_5_pad_0 = const()[name = tensor("p_5_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_5_dilations_0 = const()[name = tensor("p_5_dilations_0"), val = tensor([1, 1])]; + tensor p_5_groups_0 = const()[name = tensor("p_5_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50044864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50831360))), name = tensor("layers_2_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_5_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_5_dilations_0, groups = p_5_groups_0, pad = p_5_pad_0, pad_type = p_5_pad_type_0, strides = p_5_strides_0, weight = layers_2_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_5_cast_fp16")]; + tensor var_816 = const()[name = tensor("op_816"), val = tensor([1, 8, 128, 188])]; + tensor var_817_cast_fp16 = reshape(shape = var_816, x = q_with_bias_v_5_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 8, 128, -1])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = p_5_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor matrix_bd_17_transpose_x_0 = const()[name = tensor("matrix_bd_17_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_17_transpose_y_0 = const()[name = tensor("matrix_bd_17_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_17_cast_fp16 = matmul(transpose_x = matrix_bd_17_transpose_x_0, transpose_y = matrix_bd_17_transpose_y_0, x = var_817_cast_fp16, y = var_819_cast_fp16)[name = tensor("matrix_bd_17_cast_fp16")]; + tensor matrix_bd_19_pad_0 = const()[name = tensor("matrix_bd_19_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_19_mode_0 = const()[name = tensor("matrix_bd_19_mode_0"), val = tensor("constant")]; + tensor const_32_to_fp16 = const()[name = tensor("const_32_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_19_cast_fp16 = pad(constant_val = const_32_to_fp16, mode = matrix_bd_19_mode_0, pad = matrix_bd_19_pad_0, x = matrix_bd_17_cast_fp16)[name = tensor("matrix_bd_19_cast_fp16")]; + tensor var_828 = const()[name = tensor("op_828"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_828, x = matrix_bd_19_cast_fp16)[name = tensor("matrix_bd_21_cast_fp16")]; + tensor var_832_begin_0 = const()[name = tensor("op_832_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_832_end_0 = const()[name = tensor("op_832_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_832_end_mask_0 = const()[name = tensor("op_832_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_832_cast_fp16 = slice_by_index(begin = var_832_begin_0, end = var_832_end_0, end_mask = var_832_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor("op_832_cast_fp16")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_23_cast_fp16 = reshape(shape = var_833, x = var_832_cast_fp16)[name = tensor("matrix_bd_23_cast_fp16")]; + tensor var_838_begin_0 = const()[name = tensor("op_838_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_838_end_0 = const()[name = tensor("op_838_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_838_end_mask_0 = const()[name = tensor("op_838_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_838_cast_fp16 = slice_by_index(begin = var_838_begin_0, end = var_838_end_0, end_mask = var_838_end_mask_0, x = matrix_bd_23_cast_fp16)[name = tensor("op_838_cast_fp16")]; + tensor var_839_to_fp16 = const()[name = tensor("op_839_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_5_cast_fp16 = mul(x = var_838_cast_fp16, y = var_839_to_fp16)[name = tensor("qk_mask_5_cast_fp16")]; + tensor var_843 = const()[name = tensor("op_843"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_843, x = query_11_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_845_to_fp16 = const()[name = tensor("op_845_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_846_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_845_to_fp16)[name = tensor("op_846_cast_fp16")]; + tensor var_849 = const()[name = tensor("op_849"), val = tensor([1, 8, 128, 188])]; + tensor var_850_cast_fp16 = reshape(shape = var_849, x = key_5_cast_fp16)[name = tensor("op_850_cast_fp16")]; + tensor mh_w_9_transpose_x_0 = const()[name = tensor("mh_w_9_transpose_x_0"), val = tensor(true)]; + tensor mh_w_9_transpose_y_0 = const()[name = tensor("mh_w_9_transpose_y_0"), val = tensor(false)]; + tensor mh_w_9_cast_fp16 = matmul(transpose_x = mh_w_9_transpose_x_0, transpose_y = mh_w_9_transpose_y_0, x = var_846_cast_fp16, y = var_850_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; + tensor mh_w_11_cast_fp16 = add(x = mh_w_9_cast_fp16, y = qk_mask_5_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; + tensor var_854_cast_fp16 = softmax(axis = var_695, x = mh_w_11_cast_fp16)[name = tensor("op_854_cast_fp16")]; + tensor var_855 = const()[name = tensor("op_855"), val = tensor([1, 8, 128, 188])]; + tensor var_856_cast_fp16 = reshape(shape = var_855, x = value_5_cast_fp16)[name = tensor("op_856_cast_fp16")]; + tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; + tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; + tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_856_cast_fp16, y = var_854_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor([1, 1024, 1, 188])]; + tensor input_75_cast_fp16 = reshape(shape = var_859, x = attn_5_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor obj_13_pad_type_0 = const()[name = tensor("obj_13_pad_type_0"), val = tensor("valid")]; + tensor obj_13_strides_0 = const()[name = tensor("obj_13_strides_0"), val = tensor([1, 1])]; + tensor obj_13_pad_0 = const()[name = tensor("obj_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_13_dilations_0 = const()[name = tensor("obj_13_dilations_0"), val = tensor([1, 1])]; + tensor obj_13_groups_0 = const()[name = tensor("obj_13_groups_0"), val = tensor(1)]; + tensor layers_2_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50831552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51618048))), name = tensor("layers_2_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_13_dilations_0, groups = obj_13_groups_0, pad = obj_13_pad_0, pad_type = obj_13_pad_type_0, strides = obj_13_strides_0, weight = layers_2_self_attn_o_proj_weight_to_fp16_palettized, x = input_75_cast_fp16)[name = tensor("obj_13_cast_fp16")]; + tensor inputs_25_cast_fp16 = add(x = inputs_23_cast_fp16, y = obj_13_cast_fp16)[name = tensor("inputs_25_cast_fp16")]; + tensor out_25_axes_0 = const()[name = tensor("out_25_axes_0"), val = tensor([1])]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_877_to_fp16, x = inputs_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor input_77_gamma_0_to_fp16 = const()[name = tensor("input_77_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51618240)))]; + tensor input_77_beta_0_to_fp16 = const()[name = tensor("input_77_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51620352)))]; + tensor input_77_epsilon_0_to_fp16 = const()[name = tensor("input_77_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = batch_norm(beta = input_77_beta_0_to_fp16, epsilon = input_77_epsilon_0_to_fp16, gamma = input_77_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; + tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; + tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; + tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; + tensor layers_2_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51622464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53195392))), name = tensor("layers_2_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_79_cast_fp16 = conv(dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = layers_2_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_split_num_splits_0 = const()[name = tensor("input_81_split_num_splits_0"), val = tensor(2)]; + tensor input_81_split_axis_0 = const()[name = tensor("input_81_split_axis_0"), val = tensor(1)]; + tensor input_81_split_cast_fp16_0, tensor input_81_split_cast_fp16_1 = split(axis = input_81_split_axis_0, num_splits = input_81_split_num_splits_0, x = input_79_cast_fp16)[name = tensor("input_81_split_cast_fp16")]; + tensor input_81_split_1_sigmoid_cast_fp16 = sigmoid(x = input_81_split_cast_fp16_1)[name = tensor("input_81_split_1_sigmoid_cast_fp16")]; + tensor input_81_cast_fp16 = mul(x = input_81_split_cast_fp16_0, y = input_81_split_1_sigmoid_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_pad_type_0 = const()[name = tensor("input_83_pad_type_0"), val = tensor("custom")]; + tensor input_83_pad_0 = const()[name = tensor("input_83_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_83_groups_0 = const()[name = tensor("input_83_groups_0"), val = tensor(1024)]; + tensor input_83_strides_0 = const()[name = tensor("input_83_strides_0"), val = tensor([1, 1])]; + tensor input_83_dilations_0 = const()[name = tensor("input_83_dilations_0"), val = tensor([1, 1])]; + tensor const_272_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53195584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53202560))), name = tensor("const_272_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_273_to_fp16 = const()[name = tensor("const_273_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53202752)))]; + tensor input_85_cast_fp16 = conv(bias = const_273_to_fp16, dilations = input_83_dilations_0, groups = input_83_groups_0, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = input_83_strides_0, weight = const_272_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_cast_fp16 = silu(x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor x_17_pad_type_0 = const()[name = tensor("x_17_pad_type_0"), val = tensor("valid")]; + tensor x_17_strides_0 = const()[name = tensor("x_17_strides_0"), val = tensor([1, 1])]; + tensor x_17_pad_0 = const()[name = tensor("x_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_17_dilations_0 = const()[name = tensor("x_17_dilations_0"), val = tensor([1, 1])]; + tensor x_17_groups_0 = const()[name = tensor("x_17_groups_0"), val = tensor(1)]; + tensor layers_2_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53204864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53991360))), name = tensor("layers_2_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_17_cast_fp16 = conv(dilations = x_17_dilations_0, groups = x_17_groups_0, pad = x_17_pad_0, pad_type = x_17_pad_type_0, strides = x_17_strides_0, weight = layers_2_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor inputs_27_cast_fp16 = add(x = inputs_25_cast_fp16, y = x_17_cast_fp16)[name = tensor("inputs_27_cast_fp16")]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([1])]; + tensor var_925_to_fp16 = const()[name = tensor("op_925_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_925_to_fp16, x = inputs_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor input_89_gamma_0_to_fp16 = const()[name = tensor("input_89_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53991552)))]; + tensor input_89_beta_0_to_fp16 = const()[name = tensor("input_89_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53993664)))]; + tensor input_89_epsilon_0_to_fp16 = const()[name = tensor("input_89_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = batch_norm(beta = input_89_beta_0_to_fp16, epsilon = input_89_epsilon_0_to_fp16, gamma = input_89_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_pad_type_0 = const()[name = tensor("input_91_pad_type_0"), val = tensor("valid")]; + tensor input_91_strides_0 = const()[name = tensor("input_91_strides_0"), val = tensor([1, 1])]; + tensor input_91_pad_0 = const()[name = tensor("input_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_91_dilations_0 = const()[name = tensor("input_91_dilations_0"), val = tensor([1, 1])]; + tensor input_91_groups_0 = const()[name = tensor("input_91_groups_0"), val = tensor(1)]; + tensor layers_2_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53995776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57141568))), name = tensor("layers_2_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_91_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_91_dilations_0, groups = input_91_groups_0, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = input_91_strides_0, weight = layers_2_feed_forward2_fc1_weight_to_fp16_palettized, x = input_89_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_cast_fp16 = silu(x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor x_19_pad_type_0 = const()[name = tensor("x_19_pad_type_0"), val = tensor("valid")]; + tensor x_19_strides_0 = const()[name = tensor("x_19_strides_0"), val = tensor([1, 1])]; + tensor x_19_pad_0 = const()[name = tensor("x_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_19_dilations_0 = const()[name = tensor("x_19_dilations_0"), val = tensor([1, 1])]; + tensor x_19_groups_0 = const()[name = tensor("x_19_groups_0"), val = tensor(1)]; + tensor op_953_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57141760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60287552))), name = tensor("op_953_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_953_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_19_dilations_0, groups = x_19_groups_0, pad = x_19_pad_0, pad_type = x_19_pad_type_0, strides = x_19_strides_0, weight = op_953_weight_0_to_fp16_palettized, x = input_93_cast_fp16)[name = tensor("op_953_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_953_cast_fp16)[name = tensor("inputs_29_cast_fp16")]; + tensor out_29_axes_0 = const()[name = tensor("out_29_axes_0"), val = tensor([1])]; + tensor var_963_to_fp16 = const()[name = tensor("op_963_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_963_to_fp16, x = inputs_29_cast_fp16)[name = tensor("out_29_cast_fp16")]; + tensor inputs_31_gamma_0_to_fp16 = const()[name = tensor("inputs_31_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60287744)))]; + tensor inputs_31_beta_0_to_fp16 = const()[name = tensor("inputs_31_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60289856)))]; + tensor inputs_31_epsilon_0_to_fp16 = const()[name = tensor("inputs_31_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_31_cast_fp16 = batch_norm(beta = inputs_31_beta_0_to_fp16, epsilon = inputs_31_epsilon_0_to_fp16, gamma = inputs_31_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_977 = const()[name = tensor("op_977"), val = tensor(3)]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1008_to_fp16 = const()[name = tensor("op_1008_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1008_to_fp16, x = inputs_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor input_95_gamma_0_to_fp16 = const()[name = tensor("input_95_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60291968)))]; + tensor input_95_beta_0_to_fp16 = const()[name = tensor("input_95_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60294080)))]; + tensor input_95_epsilon_0_to_fp16 = const()[name = tensor("input_95_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = batch_norm(beta = input_95_beta_0_to_fp16, epsilon = input_95_epsilon_0_to_fp16, gamma = input_95_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor input_97_pad_type_0 = const()[name = tensor("input_97_pad_type_0"), val = tensor("valid")]; + tensor input_97_strides_0 = const()[name = tensor("input_97_strides_0"), val = tensor([1, 1])]; + tensor input_97_pad_0 = const()[name = tensor("input_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_97_dilations_0 = const()[name = tensor("input_97_dilations_0"), val = tensor([1, 1])]; + tensor input_97_groups_0 = const()[name = tensor("input_97_groups_0"), val = tensor(1)]; + tensor layers_3_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(60296192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63441984))), name = tensor("layers_3_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_97_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = layers_3_feed_forward1_fc1_weight_to_fp16_palettized, x = input_95_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_cast_fp16 = silu(x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor x_21_pad_type_0 = const()[name = tensor("x_21_pad_type_0"), val = tensor("valid")]; + tensor x_21_strides_0 = const()[name = tensor("x_21_strides_0"), val = tensor([1, 1])]; + tensor x_21_pad_0 = const()[name = tensor("x_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_21_dilations_0 = const()[name = tensor("x_21_dilations_0"), val = tensor([1, 1])]; + tensor x_21_groups_0 = const()[name = tensor("x_21_groups_0"), val = tensor(1)]; + tensor op_1036_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63442176))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66587968))), name = tensor("op_1036_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_1036_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, 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 = op_1036_weight_0_to_fp16_palettized, x = input_99_cast_fp16)[name = tensor("op_1036_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1036_cast_fp16)[name = tensor("inputs_33_cast_fp16")]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([1])]; + tensor var_1046_to_fp16 = const()[name = tensor("op_1046_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1046_to_fp16, x = inputs_33_cast_fp16)[name = tensor("out_33_cast_fp16")]; + tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66588160)))]; + tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66590272)))]; + tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_33_cast_fp16)[name = tensor("obj_15_cast_fp16")]; + tensor query_13_pad_type_0 = const()[name = tensor("query_13_pad_type_0"), val = tensor("valid")]; + tensor query_13_strides_0 = const()[name = tensor("query_13_strides_0"), val = tensor([1, 1])]; + tensor query_13_pad_0 = const()[name = tensor("query_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_13_dilations_0 = const()[name = tensor("query_13_dilations_0"), val = tensor([1, 1])]; + tensor query_13_groups_0 = const()[name = tensor("query_13_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66592384))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67378880))), name = tensor("layers_3_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_13_dilations_0, groups = query_13_groups_0, pad = query_13_pad_0, pad_type = query_13_pad_type_0, strides = query_13_strides_0, weight = layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("query_13_cast_fp16")]; + tensor key_7_pad_type_0 = const()[name = tensor("key_7_pad_type_0"), val = tensor("valid")]; + tensor key_7_strides_0 = const()[name = tensor("key_7_strides_0"), val = tensor([1, 1])]; + tensor key_7_pad_0 = const()[name = tensor("key_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_7_dilations_0 = const()[name = tensor("key_7_dilations_0"), val = tensor([1, 1])]; + tensor key_7_groups_0 = const()[name = tensor("key_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67379072))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68165568))), name = tensor("layers_3_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_7_cast_fp16 = conv(dilations = key_7_dilations_0, groups = key_7_groups_0, pad = key_7_pad_0, pad_type = key_7_pad_type_0, strides = key_7_strides_0, weight = layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("key_7_cast_fp16")]; + tensor value_7_pad_type_0 = const()[name = tensor("value_7_pad_type_0"), val = tensor("valid")]; + tensor value_7_strides_0 = const()[name = tensor("value_7_strides_0"), val = tensor([1, 1])]; + tensor value_7_pad_0 = const()[name = tensor("value_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_7_dilations_0 = const()[name = tensor("value_7_dilations_0"), val = tensor([1, 1])]; + tensor value_7_groups_0 = const()[name = tensor("value_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68165760))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68952256))), name = tensor("layers_3_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_7_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_7_dilations_0, groups = value_7_groups_0, pad = value_7_pad_0, pad_type = value_7_pad_type_0, strides = value_7_strides_0, weight = layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("value_7_cast_fp16")]; + tensor var_1084_to_fp16 = const()[name = tensor("op_1084_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68952448)))]; + tensor query_15_cast_fp16 = add(x = query_13_cast_fp16, y = var_1084_to_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_1087_to_fp16 = const()[name = tensor("op_1087_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68954560)))]; + tensor q_with_bias_v_7_cast_fp16 = add(x = query_13_cast_fp16, y = var_1087_to_fp16)[name = tensor("q_with_bias_v_7_cast_fp16")]; + tensor p_7_pad_type_0 = const()[name = tensor("p_7_pad_type_0"), val = tensor("valid")]; + tensor p_7_strides_0 = const()[name = tensor("p_7_strides_0"), val = tensor([1, 1])]; + tensor p_7_pad_0 = const()[name = tensor("p_7_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_7_dilations_0 = const()[name = tensor("p_7_dilations_0"), val = tensor([1, 1])]; + tensor p_7_groups_0 = const()[name = tensor("p_7_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68956672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69743168))), name = tensor("layers_3_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_7_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_7_dilations_0, groups = p_7_groups_0, pad = p_7_pad_0, pad_type = p_7_pad_type_0, strides = p_7_strides_0, weight = layers_3_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_7_cast_fp16")]; + tensor var_1098 = const()[name = tensor("op_1098"), val = tensor([1, 8, 128, 188])]; + tensor var_1099_cast_fp16 = reshape(shape = var_1098, x = q_with_bias_v_7_cast_fp16)[name = tensor("op_1099_cast_fp16")]; + tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1, 8, 128, -1])]; + tensor var_1101_cast_fp16 = reshape(shape = var_1100, x = p_7_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor matrix_bd_25_transpose_x_0 = const()[name = tensor("matrix_bd_25_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_25_transpose_y_0 = const()[name = tensor("matrix_bd_25_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_25_cast_fp16 = matmul(transpose_x = matrix_bd_25_transpose_x_0, transpose_y = matrix_bd_25_transpose_y_0, x = var_1099_cast_fp16, y = var_1101_cast_fp16)[name = tensor("matrix_bd_25_cast_fp16")]; + tensor matrix_bd_27_pad_0 = const()[name = tensor("matrix_bd_27_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_27_mode_0 = const()[name = tensor("matrix_bd_27_mode_0"), val = tensor("constant")]; + tensor const_43_to_fp16 = const()[name = tensor("const_43_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_27_cast_fp16 = pad(constant_val = const_43_to_fp16, mode = matrix_bd_27_mode_0, pad = matrix_bd_27_pad_0, x = matrix_bd_25_cast_fp16)[name = tensor("matrix_bd_27_cast_fp16")]; + tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_1110, x = matrix_bd_27_cast_fp16)[name = tensor("matrix_bd_29_cast_fp16")]; + tensor var_1114_begin_0 = const()[name = tensor("op_1114_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1114_end_0 = const()[name = tensor("op_1114_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1114_end_mask_0 = const()[name = tensor("op_1114_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1114_cast_fp16 = slice_by_index(begin = var_1114_begin_0, end = var_1114_end_0, end_mask = var_1114_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor("op_1114_cast_fp16")]; + tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_31_cast_fp16 = reshape(shape = var_1115, x = var_1114_cast_fp16)[name = tensor("matrix_bd_31_cast_fp16")]; + tensor var_1120_begin_0 = const()[name = tensor("op_1120_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1120_end_0 = const()[name = tensor("op_1120_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1120_end_mask_0 = const()[name = tensor("op_1120_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, x = matrix_bd_31_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1121_to_fp16 = const()[name = tensor("op_1121_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_7_cast_fp16 = mul(x = var_1120_cast_fp16, y = var_1121_to_fp16)[name = tensor("qk_mask_7_cast_fp16")]; + tensor var_1125 = const()[name = tensor("op_1125"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_1125, x = query_15_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_1127_to_fp16 = const()[name = tensor("op_1127_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1128_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_1127_to_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([1, 8, 128, 188])]; + tensor var_1132_cast_fp16 = reshape(shape = var_1131, x = key_7_cast_fp16)[name = tensor("op_1132_cast_fp16")]; + tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; + tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; + tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_1128_cast_fp16, y = var_1132_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; + tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = qk_mask_7_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; + tensor var_1136_cast_fp16 = softmax(axis = var_977, x = mh_w_15_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor var_1137 = const()[name = tensor("op_1137"), val = tensor([1, 8, 128, 188])]; + tensor var_1138_cast_fp16 = reshape(shape = var_1137, x = value_7_cast_fp16)[name = tensor("op_1138_cast_fp16")]; + tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; + tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; + tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_1138_cast_fp16, y = var_1136_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([1, 1024, 1, 188])]; + tensor input_101_cast_fp16 = reshape(shape = var_1141, x = attn_7_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor obj_17_pad_type_0 = const()[name = tensor("obj_17_pad_type_0"), val = tensor("valid")]; + tensor obj_17_strides_0 = const()[name = tensor("obj_17_strides_0"), val = tensor([1, 1])]; + tensor obj_17_pad_0 = const()[name = tensor("obj_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_17_dilations_0 = const()[name = tensor("obj_17_dilations_0"), val = tensor([1, 1])]; + tensor obj_17_groups_0 = const()[name = tensor("obj_17_groups_0"), val = tensor(1)]; + tensor layers_3_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69743360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70529856))), name = tensor("layers_3_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_17_dilations_0, groups = obj_17_groups_0, pad = obj_17_pad_0, pad_type = obj_17_pad_type_0, strides = obj_17_strides_0, weight = layers_3_self_attn_o_proj_weight_to_fp16_palettized, x = input_101_cast_fp16)[name = tensor("obj_17_cast_fp16")]; + tensor inputs_35_cast_fp16 = add(x = inputs_33_cast_fp16, y = obj_17_cast_fp16)[name = tensor("inputs_35_cast_fp16")]; + tensor out_35_axes_0 = const()[name = tensor("out_35_axes_0"), val = tensor([1])]; + tensor var_1159_to_fp16 = const()[name = tensor("op_1159_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1159_to_fp16, x = inputs_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_103_gamma_0_to_fp16 = const()[name = tensor("input_103_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70530048)))]; + tensor input_103_beta_0_to_fp16 = const()[name = tensor("input_103_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70532160)))]; + tensor input_103_epsilon_0_to_fp16 = const()[name = tensor("input_103_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_103_cast_fp16 = batch_norm(beta = input_103_beta_0_to_fp16, epsilon = input_103_epsilon_0_to_fp16, gamma = input_103_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_pad_type_0 = const()[name = tensor("input_105_pad_type_0"), val = tensor("valid")]; + tensor input_105_strides_0 = const()[name = tensor("input_105_strides_0"), val = tensor([1, 1])]; + tensor input_105_pad_0 = const()[name = tensor("input_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_105_dilations_0 = const()[name = tensor("input_105_dilations_0"), val = tensor([1, 1])]; + tensor input_105_groups_0 = const()[name = tensor("input_105_groups_0"), val = tensor(1)]; + tensor layers_3_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(70534272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72107200))), name = tensor("layers_3_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_105_cast_fp16 = conv(dilations = input_105_dilations_0, groups = input_105_groups_0, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = input_105_strides_0, weight = layers_3_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_split_num_splits_0 = const()[name = tensor("input_107_split_num_splits_0"), val = tensor(2)]; + tensor input_107_split_axis_0 = const()[name = tensor("input_107_split_axis_0"), val = tensor(1)]; + tensor input_107_split_cast_fp16_0, tensor input_107_split_cast_fp16_1 = split(axis = input_107_split_axis_0, num_splits = input_107_split_num_splits_0, x = input_105_cast_fp16)[name = tensor("input_107_split_cast_fp16")]; + tensor input_107_split_1_sigmoid_cast_fp16 = sigmoid(x = input_107_split_cast_fp16_1)[name = tensor("input_107_split_1_sigmoid_cast_fp16")]; + tensor input_107_cast_fp16 = mul(x = input_107_split_cast_fp16_0, y = input_107_split_1_sigmoid_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("custom")]; + tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1024)]; + tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; + tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; + tensor const_274_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72107392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72114368))), name = tensor("const_274_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_275_to_fp16 = const()[name = tensor("const_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72114560)))]; + tensor input_111_cast_fp16 = conv(bias = const_275_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = const_274_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_23_pad_type_0 = const()[name = tensor("x_23_pad_type_0"), val = tensor("valid")]; + tensor x_23_strides_0 = const()[name = tensor("x_23_strides_0"), val = tensor([1, 1])]; + tensor x_23_pad_0 = const()[name = tensor("x_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_23_dilations_0 = const()[name = tensor("x_23_dilations_0"), val = tensor([1, 1])]; + tensor x_23_groups_0 = const()[name = tensor("x_23_groups_0"), val = tensor(1)]; + tensor layers_3_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72116672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72903168))), name = tensor("layers_3_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_23_cast_fp16 = conv(dilations = x_23_dilations_0, groups = x_23_groups_0, pad = x_23_pad_0, pad_type = x_23_pad_type_0, strides = x_23_strides_0, weight = layers_3_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_113_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor inputs_37_cast_fp16 = add(x = inputs_35_cast_fp16, y = x_23_cast_fp16)[name = tensor("inputs_37_cast_fp16")]; + tensor out_37_axes_0 = const()[name = tensor("out_37_axes_0"), val = tensor([1])]; + tensor var_1207_to_fp16 = const()[name = tensor("op_1207_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1207_to_fp16, x = inputs_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor input_115_gamma_0_to_fp16 = const()[name = tensor("input_115_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72903360)))]; + tensor input_115_beta_0_to_fp16 = const()[name = tensor("input_115_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72905472)))]; + tensor input_115_epsilon_0_to_fp16 = const()[name = tensor("input_115_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_115_cast_fp16 = batch_norm(beta = input_115_beta_0_to_fp16, epsilon = input_115_epsilon_0_to_fp16, gamma = input_115_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_pad_type_0 = const()[name = tensor("input_117_pad_type_0"), val = tensor("valid")]; + tensor input_117_strides_0 = const()[name = tensor("input_117_strides_0"), val = tensor([1, 1])]; + tensor input_117_pad_0 = const()[name = tensor("input_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_117_dilations_0 = const()[name = tensor("input_117_dilations_0"), val = tensor([1, 1])]; + tensor input_117_groups_0 = const()[name = tensor("input_117_groups_0"), val = tensor(1)]; + tensor layers_3_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72907584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76053376))), name = tensor("layers_3_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_117_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_117_dilations_0, groups = input_117_groups_0, pad = input_117_pad_0, pad_type = input_117_pad_type_0, strides = input_117_strides_0, weight = layers_3_feed_forward2_fc1_weight_to_fp16_palettized, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_cast_fp16 = silu(x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor x_25_pad_type_0 = const()[name = tensor("x_25_pad_type_0"), val = tensor("valid")]; + tensor x_25_strides_0 = const()[name = tensor("x_25_strides_0"), val = tensor([1, 1])]; + tensor x_25_pad_0 = const()[name = tensor("x_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_25_dilations_0 = const()[name = tensor("x_25_dilations_0"), val = tensor([1, 1])]; + tensor x_25_groups_0 = const()[name = tensor("x_25_groups_0"), val = tensor(1)]; + tensor op_1235_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76053568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79199360))), name = tensor("op_1235_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_1235_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = op_1235_weight_0_to_fp16_palettized, x = input_119_cast_fp16)[name = tensor("op_1235_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_1235_cast_fp16)[name = tensor("inputs_39_cast_fp16")]; + tensor out_39_axes_0 = const()[name = tensor("out_39_axes_0"), val = tensor([1])]; + tensor var_1245_to_fp16 = const()[name = tensor("op_1245_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1245_to_fp16, x = inputs_39_cast_fp16)[name = tensor("out_39_cast_fp16")]; + tensor inputs_41_gamma_0_to_fp16 = const()[name = tensor("inputs_41_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79199552)))]; + tensor inputs_41_beta_0_to_fp16 = const()[name = tensor("inputs_41_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79201664)))]; + tensor inputs_41_epsilon_0_to_fp16 = const()[name = tensor("inputs_41_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_41_cast_fp16 = batch_norm(beta = inputs_41_beta_0_to_fp16, epsilon = inputs_41_epsilon_0_to_fp16, gamma = inputs_41_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor(3)]; + tensor out_41_axes_0 = const()[name = tensor("out_41_axes_0"), val = tensor([1])]; + tensor var_1290_to_fp16 = const()[name = tensor("op_1290_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1290_to_fp16, x = inputs_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_121_gamma_0_to_fp16 = const()[name = tensor("input_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79203776)))]; + tensor input_121_beta_0_to_fp16 = const()[name = tensor("input_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79205888)))]; + tensor input_121_epsilon_0_to_fp16 = const()[name = tensor("input_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_121_cast_fp16 = batch_norm(beta = input_121_beta_0_to_fp16, epsilon = input_121_epsilon_0_to_fp16, gamma = input_121_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_pad_type_0 = const()[name = tensor("input_123_pad_type_0"), val = tensor("valid")]; + tensor input_123_strides_0 = const()[name = tensor("input_123_strides_0"), val = tensor([1, 1])]; + tensor input_123_pad_0 = const()[name = tensor("input_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_123_dilations_0 = const()[name = tensor("input_123_dilations_0"), val = tensor([1, 1])]; + tensor input_123_groups_0 = const()[name = tensor("input_123_groups_0"), val = tensor(1)]; + tensor layers_4_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79208000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82353792))), name = tensor("layers_4_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_123_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_123_dilations_0, groups = input_123_groups_0, pad = input_123_pad_0, pad_type = input_123_pad_type_0, strides = input_123_strides_0, weight = layers_4_feed_forward1_fc1_weight_to_fp16_palettized, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_cast_fp16 = silu(x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor x_27_pad_type_0 = const()[name = tensor("x_27_pad_type_0"), val = tensor("valid")]; + tensor x_27_strides_0 = const()[name = tensor("x_27_strides_0"), val = tensor([1, 1])]; + tensor x_27_pad_0 = const()[name = tensor("x_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_27_dilations_0 = const()[name = tensor("x_27_dilations_0"), val = tensor([1, 1])]; + tensor x_27_groups_0 = const()[name = tensor("x_27_groups_0"), val = tensor(1)]; + tensor op_1318_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82353984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85499776))), name = tensor("op_1318_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_1318_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_27_dilations_0, groups = x_27_groups_0, pad = x_27_pad_0, pad_type = x_27_pad_type_0, strides = x_27_strides_0, weight = op_1318_weight_0_to_fp16_palettized, x = input_125_cast_fp16)[name = tensor("op_1318_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_1318_cast_fp16)[name = tensor("inputs_43_cast_fp16")]; + tensor out_43_axes_0 = const()[name = tensor("out_43_axes_0"), val = tensor([1])]; + tensor var_1328_to_fp16 = const()[name = tensor("op_1328_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1328_to_fp16, x = inputs_43_cast_fp16)[name = tensor("out_43_cast_fp16")]; + tensor obj_19_gamma_0_to_fp16 = const()[name = tensor("obj_19_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85499968)))]; + tensor obj_19_beta_0_to_fp16 = const()[name = tensor("obj_19_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85502080)))]; + tensor obj_19_epsilon_0_to_fp16 = const()[name = tensor("obj_19_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_19_cast_fp16 = batch_norm(beta = obj_19_beta_0_to_fp16, epsilon = obj_19_epsilon_0_to_fp16, gamma = obj_19_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_43_cast_fp16)[name = tensor("obj_19_cast_fp16")]; + tensor query_17_pad_type_0 = const()[name = tensor("query_17_pad_type_0"), val = tensor("valid")]; + tensor query_17_strides_0 = const()[name = tensor("query_17_strides_0"), val = tensor([1, 1])]; + tensor query_17_pad_0 = const()[name = tensor("query_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_17_dilations_0 = const()[name = tensor("query_17_dilations_0"), val = tensor([1, 1])]; + tensor query_17_groups_0 = const()[name = tensor("query_17_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85504192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86290688))), name = tensor("layers_4_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_17_dilations_0, groups = query_17_groups_0, pad = query_17_pad_0, pad_type = query_17_pad_type_0, strides = query_17_strides_0, weight = layers_4_self_attn_q_proj_weight_to_fp16_palettized, x = obj_19_cast_fp16)[name = tensor("query_17_cast_fp16")]; + tensor key_9_pad_type_0 = const()[name = tensor("key_9_pad_type_0"), val = tensor("valid")]; + tensor key_9_strides_0 = const()[name = tensor("key_9_strides_0"), val = tensor([1, 1])]; + tensor key_9_pad_0 = const()[name = tensor("key_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_9_dilations_0 = const()[name = tensor("key_9_dilations_0"), val = tensor([1, 1])]; + tensor key_9_groups_0 = const()[name = tensor("key_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(86290880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87077376))), name = tensor("layers_4_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_9_cast_fp16 = conv(dilations = key_9_dilations_0, groups = key_9_groups_0, pad = key_9_pad_0, pad_type = key_9_pad_type_0, strides = key_9_strides_0, weight = layers_4_self_attn_k_proj_weight_to_fp16_palettized, x = obj_19_cast_fp16)[name = tensor("key_9_cast_fp16")]; + tensor value_9_pad_type_0 = const()[name = tensor("value_9_pad_type_0"), val = tensor("valid")]; + tensor value_9_strides_0 = const()[name = tensor("value_9_strides_0"), val = tensor([1, 1])]; + tensor value_9_pad_0 = const()[name = tensor("value_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_9_dilations_0 = const()[name = tensor("value_9_dilations_0"), val = tensor([1, 1])]; + tensor value_9_groups_0 = const()[name = tensor("value_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87077568))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87864064))), name = tensor("layers_4_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_9_dilations_0, groups = value_9_groups_0, pad = value_9_pad_0, pad_type = value_9_pad_type_0, strides = value_9_strides_0, weight = layers_4_self_attn_v_proj_weight_to_fp16_palettized, x = obj_19_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_1366_to_fp16 = const()[name = tensor("op_1366_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87864256)))]; + tensor query_19_cast_fp16 = add(x = query_17_cast_fp16, y = var_1366_to_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_1369_to_fp16 = const()[name = tensor("op_1369_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87866368)))]; + tensor q_with_bias_v_9_cast_fp16 = add(x = query_17_cast_fp16, y = var_1369_to_fp16)[name = tensor("q_with_bias_v_9_cast_fp16")]; + tensor p_9_pad_type_0 = const()[name = tensor("p_9_pad_type_0"), val = tensor("valid")]; + tensor p_9_strides_0 = const()[name = tensor("p_9_strides_0"), val = tensor([1, 1])]; + tensor p_9_pad_0 = const()[name = tensor("p_9_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_9_dilations_0 = const()[name = tensor("p_9_dilations_0"), val = tensor([1, 1])]; + tensor p_9_groups_0 = const()[name = tensor("p_9_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87868480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88654976))), name = tensor("layers_4_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_9_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_9_dilations_0, groups = p_9_groups_0, pad = p_9_pad_0, pad_type = p_9_pad_type_0, strides = p_9_strides_0, weight = layers_4_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_9_cast_fp16")]; + tensor var_1380 = const()[name = tensor("op_1380"), val = tensor([1, 8, 128, 188])]; + tensor var_1381_cast_fp16 = reshape(shape = var_1380, x = q_with_bias_v_9_cast_fp16)[name = tensor("op_1381_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([1, 8, 128, -1])]; + tensor var_1383_cast_fp16 = reshape(shape = var_1382, x = p_9_cast_fp16)[name = tensor("op_1383_cast_fp16")]; + tensor matrix_bd_33_transpose_x_0 = const()[name = tensor("matrix_bd_33_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_33_transpose_y_0 = const()[name = tensor("matrix_bd_33_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_33_cast_fp16 = matmul(transpose_x = matrix_bd_33_transpose_x_0, transpose_y = matrix_bd_33_transpose_y_0, x = var_1381_cast_fp16, y = var_1383_cast_fp16)[name = tensor("matrix_bd_33_cast_fp16")]; + tensor matrix_bd_35_pad_0 = const()[name = tensor("matrix_bd_35_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_35_mode_0 = const()[name = tensor("matrix_bd_35_mode_0"), val = tensor("constant")]; + tensor const_54_to_fp16 = const()[name = tensor("const_54_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_35_cast_fp16 = pad(constant_val = const_54_to_fp16, mode = matrix_bd_35_mode_0, pad = matrix_bd_35_pad_0, x = matrix_bd_33_cast_fp16)[name = tensor("matrix_bd_35_cast_fp16")]; + tensor var_1392 = const()[name = tensor("op_1392"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_1392, x = matrix_bd_35_cast_fp16)[name = tensor("matrix_bd_37_cast_fp16")]; + tensor var_1396_begin_0 = const()[name = tensor("op_1396_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1396_end_0 = const()[name = tensor("op_1396_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1396_end_mask_0 = const()[name = tensor("op_1396_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1396_cast_fp16 = slice_by_index(begin = var_1396_begin_0, end = var_1396_end_0, end_mask = var_1396_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor("op_1396_cast_fp16")]; + tensor var_1397 = const()[name = tensor("op_1397"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_39_cast_fp16 = reshape(shape = var_1397, x = var_1396_cast_fp16)[name = tensor("matrix_bd_39_cast_fp16")]; + tensor var_1402_begin_0 = const()[name = tensor("op_1402_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1402_end_0 = const()[name = tensor("op_1402_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1402_end_mask_0 = const()[name = tensor("op_1402_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1402_cast_fp16 = slice_by_index(begin = var_1402_begin_0, end = var_1402_end_0, end_mask = var_1402_end_mask_0, x = matrix_bd_39_cast_fp16)[name = tensor("op_1402_cast_fp16")]; + tensor var_1403_to_fp16 = const()[name = tensor("op_1403_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_9_cast_fp16 = mul(x = var_1402_cast_fp16, y = var_1403_to_fp16)[name = tensor("qk_mask_9_cast_fp16")]; + tensor var_1407 = const()[name = tensor("op_1407"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_1407, x = query_19_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1410_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([1, 8, 128, 188])]; + tensor var_1414_cast_fp16 = reshape(shape = var_1413, x = key_9_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; + tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; + tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1410_cast_fp16, y = var_1414_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; + tensor mh_w_19_cast_fp16 = add(x = mh_w_17_cast_fp16, y = qk_mask_9_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; + tensor var_1418_cast_fp16 = softmax(axis = var_1259, x = mh_w_19_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 8, 128, 188])]; + tensor var_1420_cast_fp16 = reshape(shape = var_1419, x = value_9_cast_fp16)[name = tensor("op_1420_cast_fp16")]; + tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; + tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; + tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_1420_cast_fp16, y = var_1418_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1423 = const()[name = tensor("op_1423"), val = tensor([1, 1024, 1, 188])]; + tensor input_127_cast_fp16 = reshape(shape = var_1423, x = attn_9_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor obj_21_pad_type_0 = const()[name = tensor("obj_21_pad_type_0"), val = tensor("valid")]; + tensor obj_21_strides_0 = const()[name = tensor("obj_21_strides_0"), val = tensor([1, 1])]; + tensor obj_21_pad_0 = const()[name = tensor("obj_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_21_dilations_0 = const()[name = tensor("obj_21_dilations_0"), val = tensor([1, 1])]; + tensor obj_21_groups_0 = const()[name = tensor("obj_21_groups_0"), val = tensor(1)]; + tensor layers_4_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88655168))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89441664))), name = tensor("layers_4_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_21_dilations_0, groups = obj_21_groups_0, pad = obj_21_pad_0, pad_type = obj_21_pad_type_0, strides = obj_21_strides_0, weight = layers_4_self_attn_o_proj_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = tensor("obj_21_cast_fp16")]; + tensor inputs_45_cast_fp16 = add(x = inputs_43_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_45_cast_fp16")]; + tensor out_45_axes_0 = const()[name = tensor("out_45_axes_0"), val = tensor([1])]; + tensor var_1441_to_fp16 = const()[name = tensor("op_1441_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1441_to_fp16, x = inputs_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor input_129_gamma_0_to_fp16 = const()[name = tensor("input_129_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89441856)))]; + tensor input_129_beta_0_to_fp16 = const()[name = tensor("input_129_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89443968)))]; + tensor input_129_epsilon_0_to_fp16 = const()[name = tensor("input_129_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = batch_norm(beta = input_129_beta_0_to_fp16, epsilon = input_129_epsilon_0_to_fp16, gamma = input_129_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_pad_type_0 = const()[name = tensor("input_131_pad_type_0"), val = tensor("valid")]; + tensor input_131_strides_0 = const()[name = tensor("input_131_strides_0"), val = tensor([1, 1])]; + tensor input_131_pad_0 = const()[name = tensor("input_131_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_131_dilations_0 = const()[name = tensor("input_131_dilations_0"), val = tensor([1, 1])]; + tensor input_131_groups_0 = const()[name = tensor("input_131_groups_0"), val = tensor(1)]; + tensor layers_4_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(89446080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91019008))), name = tensor("layers_4_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_131_cast_fp16 = conv(dilations = input_131_dilations_0, groups = input_131_groups_0, pad = input_131_pad_0, pad_type = input_131_pad_type_0, strides = input_131_strides_0, weight = layers_4_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor input_133_split_num_splits_0 = const()[name = tensor("input_133_split_num_splits_0"), val = tensor(2)]; + tensor input_133_split_axis_0 = const()[name = tensor("input_133_split_axis_0"), val = tensor(1)]; + tensor input_133_split_cast_fp16_0, tensor input_133_split_cast_fp16_1 = split(axis = input_133_split_axis_0, num_splits = input_133_split_num_splits_0, x = input_131_cast_fp16)[name = tensor("input_133_split_cast_fp16")]; + tensor input_133_split_1_sigmoid_cast_fp16 = sigmoid(x = input_133_split_cast_fp16_1)[name = tensor("input_133_split_1_sigmoid_cast_fp16")]; + tensor input_133_cast_fp16 = mul(x = input_133_split_cast_fp16_0, y = input_133_split_1_sigmoid_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_pad_type_0 = const()[name = tensor("input_135_pad_type_0"), val = tensor("custom")]; + tensor input_135_pad_0 = const()[name = tensor("input_135_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_135_groups_0 = const()[name = tensor("input_135_groups_0"), val = tensor(1024)]; + tensor input_135_strides_0 = const()[name = tensor("input_135_strides_0"), val = tensor([1, 1])]; + tensor input_135_dilations_0 = const()[name = tensor("input_135_dilations_0"), val = tensor([1, 1])]; + tensor const_276_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91019200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91026176))), name = tensor("const_276_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_277_to_fp16 = const()[name = tensor("const_277_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91026368)))]; + tensor input_137_cast_fp16 = conv(bias = const_277_to_fp16, dilations = input_135_dilations_0, groups = input_135_groups_0, pad = input_135_pad_0, pad_type = input_135_pad_type_0, strides = input_135_strides_0, weight = const_276_to_fp16_palettized, x = input_133_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_cast_fp16 = silu(x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor x_29_pad_type_0 = const()[name = tensor("x_29_pad_type_0"), val = tensor("valid")]; + tensor x_29_strides_0 = const()[name = tensor("x_29_strides_0"), val = tensor([1, 1])]; + tensor x_29_pad_0 = const()[name = tensor("x_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_29_dilations_0 = const()[name = tensor("x_29_dilations_0"), val = tensor([1, 1])]; + tensor x_29_groups_0 = const()[name = tensor("x_29_groups_0"), val = tensor(1)]; + tensor layers_4_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91028480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91814976))), name = tensor("layers_4_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_29_cast_fp16 = conv(dilations = x_29_dilations_0, groups = x_29_groups_0, pad = x_29_pad_0, pad_type = x_29_pad_type_0, strides = x_29_strides_0, weight = layers_4_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_139_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor inputs_47_cast_fp16 = add(x = inputs_45_cast_fp16, y = x_29_cast_fp16)[name = tensor("inputs_47_cast_fp16")]; + tensor out_47_axes_0 = const()[name = tensor("out_47_axes_0"), val = tensor([1])]; + tensor var_1489_to_fp16 = const()[name = tensor("op_1489_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1489_to_fp16, x = inputs_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_141_gamma_0_to_fp16 = const()[name = tensor("input_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91815168)))]; + tensor input_141_beta_0_to_fp16 = const()[name = tensor("input_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91817280)))]; + tensor input_141_epsilon_0_to_fp16 = const()[name = tensor("input_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = batch_norm(beta = input_141_beta_0_to_fp16, epsilon = input_141_epsilon_0_to_fp16, gamma = input_141_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_pad_type_0 = const()[name = tensor("input_143_pad_type_0"), val = tensor("valid")]; + tensor input_143_strides_0 = const()[name = tensor("input_143_strides_0"), val = tensor([1, 1])]; + tensor input_143_pad_0 = const()[name = tensor("input_143_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_143_dilations_0 = const()[name = tensor("input_143_dilations_0"), val = tensor([1, 1])]; + tensor input_143_groups_0 = const()[name = tensor("input_143_groups_0"), val = tensor(1)]; + tensor layers_4_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91819392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94965184))), name = tensor("layers_4_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_143_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_143_dilations_0, groups = input_143_groups_0, pad = input_143_pad_0, pad_type = input_143_pad_type_0, strides = input_143_strides_0, weight = layers_4_feed_forward2_fc1_weight_to_fp16_palettized, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor input_145_cast_fp16 = silu(x = input_143_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor x_31_pad_type_0 = const()[name = tensor("x_31_pad_type_0"), val = tensor("valid")]; + tensor x_31_strides_0 = const()[name = tensor("x_31_strides_0"), val = tensor([1, 1])]; + tensor x_31_pad_0 = const()[name = tensor("x_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_31_dilations_0 = const()[name = tensor("x_31_dilations_0"), val = tensor([1, 1])]; + tensor x_31_groups_0 = const()[name = tensor("x_31_groups_0"), val = tensor(1)]; + tensor op_1517_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(94965376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98111168))), name = tensor("op_1517_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_1517_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_31_dilations_0, groups = x_31_groups_0, pad = x_31_pad_0, pad_type = x_31_pad_type_0, strides = x_31_strides_0, weight = op_1517_weight_0_to_fp16_palettized, x = input_145_cast_fp16)[name = tensor("op_1517_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_1517_cast_fp16)[name = tensor("inputs_49_cast_fp16")]; + tensor out_49_axes_0 = const()[name = tensor("out_49_axes_0"), val = tensor([1])]; + tensor var_1527_to_fp16 = const()[name = tensor("op_1527_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1527_to_fp16, x = inputs_49_cast_fp16)[name = tensor("out_49_cast_fp16")]; + tensor inputs_51_gamma_0_to_fp16 = const()[name = tensor("inputs_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98111360)))]; + tensor inputs_51_beta_0_to_fp16 = const()[name = tensor("inputs_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98113472)))]; + tensor inputs_51_epsilon_0_to_fp16 = const()[name = tensor("inputs_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_51_cast_fp16 = batch_norm(beta = inputs_51_beta_0_to_fp16, epsilon = inputs_51_epsilon_0_to_fp16, gamma = inputs_51_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor var_1541 = const()[name = tensor("op_1541"), val = tensor(3)]; + tensor out_51_axes_0 = const()[name = tensor("out_51_axes_0"), val = tensor([1])]; + tensor var_1572_to_fp16 = const()[name = tensor("op_1572_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1572_to_fp16, x = inputs_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor input_147_gamma_0_to_fp16 = const()[name = tensor("input_147_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98115584)))]; + tensor input_147_beta_0_to_fp16 = const()[name = tensor("input_147_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98117696)))]; + tensor input_147_epsilon_0_to_fp16 = const()[name = tensor("input_147_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = batch_norm(beta = input_147_beta_0_to_fp16, epsilon = input_147_epsilon_0_to_fp16, gamma = input_147_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; + tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1, 1])]; + tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1, 1])]; + tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; + tensor layers_5_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98119808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101265600))), name = tensor("layers_5_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_149_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = layers_5_feed_forward1_fc1_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor x_33_pad_type_0 = const()[name = tensor("x_33_pad_type_0"), val = tensor("valid")]; + tensor x_33_strides_0 = const()[name = tensor("x_33_strides_0"), val = tensor([1, 1])]; + tensor x_33_pad_0 = const()[name = tensor("x_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_33_dilations_0 = const()[name = tensor("x_33_dilations_0"), val = tensor([1, 1])]; + tensor x_33_groups_0 = const()[name = tensor("x_33_groups_0"), val = tensor(1)]; + tensor op_1600_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101265792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104411584))), name = tensor("op_1600_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_1600_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_33_dilations_0, groups = x_33_groups_0, pad = x_33_pad_0, pad_type = x_33_pad_type_0, strides = x_33_strides_0, weight = op_1600_weight_0_to_fp16_palettized, x = input_151_cast_fp16)[name = tensor("op_1600_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_1600_cast_fp16)[name = tensor("inputs_53_cast_fp16")]; + tensor out_53_axes_0 = const()[name = tensor("out_53_axes_0"), val = tensor([1])]; + tensor var_1610_to_fp16 = const()[name = tensor("op_1610_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1610_to_fp16, x = inputs_53_cast_fp16)[name = tensor("out_53_cast_fp16")]; + tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104411776)))]; + tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104413888)))]; + tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_53_cast_fp16)[name = tensor("obj_23_cast_fp16")]; + tensor query_21_pad_type_0 = const()[name = tensor("query_21_pad_type_0"), val = tensor("valid")]; + tensor query_21_strides_0 = const()[name = tensor("query_21_strides_0"), val = tensor([1, 1])]; + tensor query_21_pad_0 = const()[name = tensor("query_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_21_dilations_0 = const()[name = tensor("query_21_dilations_0"), val = tensor([1, 1])]; + tensor query_21_groups_0 = const()[name = tensor("query_21_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(104416000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105202496))), name = tensor("layers_5_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_21_dilations_0, groups = query_21_groups_0, pad = query_21_pad_0, pad_type = query_21_pad_type_0, strides = query_21_strides_0, weight = layers_5_self_attn_q_proj_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor("query_21_cast_fp16")]; + tensor key_11_pad_type_0 = const()[name = tensor("key_11_pad_type_0"), val = tensor("valid")]; + tensor key_11_strides_0 = const()[name = tensor("key_11_strides_0"), val = tensor([1, 1])]; + tensor key_11_pad_0 = const()[name = tensor("key_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_11_dilations_0 = const()[name = tensor("key_11_dilations_0"), val = tensor([1, 1])]; + tensor key_11_groups_0 = const()[name = tensor("key_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105202688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105989184))), name = tensor("layers_5_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_11_cast_fp16 = conv(dilations = key_11_dilations_0, groups = key_11_groups_0, pad = key_11_pad_0, pad_type = key_11_pad_type_0, strides = key_11_strides_0, weight = layers_5_self_attn_k_proj_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor("key_11_cast_fp16")]; + tensor value_11_pad_type_0 = const()[name = tensor("value_11_pad_type_0"), val = tensor("valid")]; + tensor value_11_strides_0 = const()[name = tensor("value_11_strides_0"), val = tensor([1, 1])]; + tensor value_11_pad_0 = const()[name = tensor("value_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_11_dilations_0 = const()[name = tensor("value_11_dilations_0"), val = tensor([1, 1])]; + tensor value_11_groups_0 = const()[name = tensor("value_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105989376))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106775872))), name = tensor("layers_5_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_11_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_11_dilations_0, groups = value_11_groups_0, pad = value_11_pad_0, pad_type = value_11_pad_type_0, strides = value_11_strides_0, weight = layers_5_self_attn_v_proj_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor("value_11_cast_fp16")]; + tensor var_1648_to_fp16 = const()[name = tensor("op_1648_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106776064)))]; + tensor query_23_cast_fp16 = add(x = query_21_cast_fp16, y = var_1648_to_fp16)[name = tensor("query_23_cast_fp16")]; + tensor var_1651_to_fp16 = const()[name = tensor("op_1651_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106778176)))]; + tensor q_with_bias_v_11_cast_fp16 = add(x = query_21_cast_fp16, y = var_1651_to_fp16)[name = tensor("q_with_bias_v_11_cast_fp16")]; + tensor p_11_pad_type_0 = const()[name = tensor("p_11_pad_type_0"), val = tensor("valid")]; + tensor p_11_strides_0 = const()[name = tensor("p_11_strides_0"), val = tensor([1, 1])]; + tensor p_11_pad_0 = const()[name = tensor("p_11_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_11_dilations_0 = const()[name = tensor("p_11_dilations_0"), val = tensor([1, 1])]; + tensor p_11_groups_0 = const()[name = tensor("p_11_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(106780288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107566784))), name = tensor("layers_5_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_11_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_11_dilations_0, groups = p_11_groups_0, pad = p_11_pad_0, pad_type = p_11_pad_type_0, strides = p_11_strides_0, weight = layers_5_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_11_cast_fp16")]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 8, 128, 188])]; + tensor var_1663_cast_fp16 = reshape(shape = var_1662, x = q_with_bias_v_11_cast_fp16)[name = tensor("op_1663_cast_fp16")]; + tensor var_1664 = const()[name = tensor("op_1664"), val = tensor([1, 8, 128, -1])]; + tensor var_1665_cast_fp16 = reshape(shape = var_1664, x = p_11_cast_fp16)[name = tensor("op_1665_cast_fp16")]; + tensor matrix_bd_41_transpose_x_0 = const()[name = tensor("matrix_bd_41_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_41_transpose_y_0 = const()[name = tensor("matrix_bd_41_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_41_cast_fp16 = matmul(transpose_x = matrix_bd_41_transpose_x_0, transpose_y = matrix_bd_41_transpose_y_0, x = var_1663_cast_fp16, y = var_1665_cast_fp16)[name = tensor("matrix_bd_41_cast_fp16")]; + tensor matrix_bd_43_pad_0 = const()[name = tensor("matrix_bd_43_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_43_mode_0 = const()[name = tensor("matrix_bd_43_mode_0"), val = tensor("constant")]; + tensor const_65_to_fp16 = const()[name = tensor("const_65_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_43_cast_fp16 = pad(constant_val = const_65_to_fp16, mode = matrix_bd_43_mode_0, pad = matrix_bd_43_pad_0, x = matrix_bd_41_cast_fp16)[name = tensor("matrix_bd_43_cast_fp16")]; + tensor var_1674 = const()[name = tensor("op_1674"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_1674, x = matrix_bd_43_cast_fp16)[name = tensor("matrix_bd_45_cast_fp16")]; + tensor var_1678_begin_0 = const()[name = tensor("op_1678_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1678_end_0 = const()[name = tensor("op_1678_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1678_end_mask_0 = const()[name = tensor("op_1678_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1678_cast_fp16 = slice_by_index(begin = var_1678_begin_0, end = var_1678_end_0, end_mask = var_1678_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor("op_1678_cast_fp16")]; + tensor var_1679 = const()[name = tensor("op_1679"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_47_cast_fp16 = reshape(shape = var_1679, x = var_1678_cast_fp16)[name = tensor("matrix_bd_47_cast_fp16")]; + tensor var_1684_begin_0 = const()[name = tensor("op_1684_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1684_end_0 = const()[name = tensor("op_1684_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1684_end_mask_0 = const()[name = tensor("op_1684_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1684_cast_fp16 = slice_by_index(begin = var_1684_begin_0, end = var_1684_end_0, end_mask = var_1684_end_mask_0, x = matrix_bd_47_cast_fp16)[name = tensor("op_1684_cast_fp16")]; + tensor var_1685_to_fp16 = const()[name = tensor("op_1685_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_11_cast_fp16 = mul(x = var_1684_cast_fp16, y = var_1685_to_fp16)[name = tensor("qk_mask_11_cast_fp16")]; + tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_1689, x = query_23_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_1691_to_fp16 = const()[name = tensor("op_1691_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1692_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1691_to_fp16)[name = tensor("op_1692_cast_fp16")]; + tensor var_1695 = const()[name = tensor("op_1695"), val = tensor([1, 8, 128, 188])]; + tensor var_1696_cast_fp16 = reshape(shape = var_1695, x = key_11_cast_fp16)[name = tensor("op_1696_cast_fp16")]; + tensor mh_w_21_transpose_x_0 = const()[name = tensor("mh_w_21_transpose_x_0"), val = tensor(true)]; + tensor mh_w_21_transpose_y_0 = const()[name = tensor("mh_w_21_transpose_y_0"), val = tensor(false)]; + tensor mh_w_21_cast_fp16 = matmul(transpose_x = mh_w_21_transpose_x_0, transpose_y = mh_w_21_transpose_y_0, x = var_1692_cast_fp16, y = var_1696_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; + tensor mh_w_23_cast_fp16 = add(x = mh_w_21_cast_fp16, y = qk_mask_11_cast_fp16)[name = tensor("mh_w_23_cast_fp16")]; + tensor var_1700_cast_fp16 = softmax(axis = var_1541, x = mh_w_23_cast_fp16)[name = tensor("op_1700_cast_fp16")]; + tensor var_1701 = const()[name = tensor("op_1701"), val = tensor([1, 8, 128, 188])]; + tensor var_1702_cast_fp16 = reshape(shape = var_1701, x = value_11_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; + tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; + tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1702_cast_fp16, y = var_1700_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1705 = const()[name = tensor("op_1705"), val = tensor([1, 1024, 1, 188])]; + tensor input_153_cast_fp16 = reshape(shape = var_1705, x = attn_11_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor obj_25_pad_type_0 = const()[name = tensor("obj_25_pad_type_0"), val = tensor("valid")]; + tensor obj_25_strides_0 = const()[name = tensor("obj_25_strides_0"), val = tensor([1, 1])]; + tensor obj_25_pad_0 = const()[name = tensor("obj_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_25_dilations_0 = const()[name = tensor("obj_25_dilations_0"), val = tensor([1, 1])]; + tensor obj_25_groups_0 = const()[name = tensor("obj_25_groups_0"), val = tensor(1)]; + tensor layers_5_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(107566976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108353472))), name = tensor("layers_5_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_25_dilations_0, groups = obj_25_groups_0, pad = obj_25_pad_0, pad_type = obj_25_pad_type_0, strides = obj_25_strides_0, weight = layers_5_self_attn_o_proj_weight_to_fp16_palettized, x = input_153_cast_fp16)[name = tensor("obj_25_cast_fp16")]; + tensor inputs_55_cast_fp16 = add(x = inputs_53_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_55_cast_fp16")]; + tensor out_55_axes_0 = const()[name = tensor("out_55_axes_0"), val = tensor([1])]; + tensor var_1723_to_fp16 = const()[name = tensor("op_1723_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_1723_to_fp16, x = inputs_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor input_155_gamma_0_to_fp16 = const()[name = tensor("input_155_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108353664)))]; + tensor input_155_beta_0_to_fp16 = const()[name = tensor("input_155_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108355776)))]; + tensor input_155_epsilon_0_to_fp16 = const()[name = tensor("input_155_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = batch_norm(beta = input_155_beta_0_to_fp16, epsilon = input_155_epsilon_0_to_fp16, gamma = input_155_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor input_157_pad_type_0 = const()[name = tensor("input_157_pad_type_0"), val = tensor("valid")]; + tensor input_157_strides_0 = const()[name = tensor("input_157_strides_0"), val = tensor([1, 1])]; + tensor input_157_pad_0 = const()[name = tensor("input_157_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_157_dilations_0 = const()[name = tensor("input_157_dilations_0"), val = tensor([1, 1])]; + tensor input_157_groups_0 = const()[name = tensor("input_157_groups_0"), val = tensor(1)]; + tensor layers_5_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108357888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109930816))), name = tensor("layers_5_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_157_cast_fp16 = conv(dilations = input_157_dilations_0, groups = input_157_groups_0, pad = input_157_pad_0, pad_type = input_157_pad_type_0, strides = input_157_strides_0, weight = layers_5_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_155_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_split_num_splits_0 = const()[name = tensor("input_159_split_num_splits_0"), val = tensor(2)]; + tensor input_159_split_axis_0 = const()[name = tensor("input_159_split_axis_0"), val = tensor(1)]; + tensor input_159_split_cast_fp16_0, tensor input_159_split_cast_fp16_1 = split(axis = input_159_split_axis_0, num_splits = input_159_split_num_splits_0, x = input_157_cast_fp16)[name = tensor("input_159_split_cast_fp16")]; + tensor input_159_split_1_sigmoid_cast_fp16 = sigmoid(x = input_159_split_cast_fp16_1)[name = tensor("input_159_split_1_sigmoid_cast_fp16")]; + tensor input_159_cast_fp16 = mul(x = input_159_split_cast_fp16_0, y = input_159_split_1_sigmoid_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_pad_type_0 = const()[name = tensor("input_161_pad_type_0"), val = tensor("custom")]; + tensor input_161_pad_0 = const()[name = tensor("input_161_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_161_groups_0 = const()[name = tensor("input_161_groups_0"), val = tensor(1024)]; + tensor input_161_strides_0 = const()[name = tensor("input_161_strides_0"), val = tensor([1, 1])]; + tensor input_161_dilations_0 = const()[name = tensor("input_161_dilations_0"), val = tensor([1, 1])]; + tensor const_278_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109931008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109937984))), name = tensor("const_278_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_279_to_fp16 = const()[name = tensor("const_279_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109938176)))]; + tensor input_163_cast_fp16 = conv(bias = const_279_to_fp16, dilations = input_161_dilations_0, groups = input_161_groups_0, pad = input_161_pad_0, pad_type = input_161_pad_type_0, strides = input_161_strides_0, weight = const_278_to_fp16_palettized, x = input_159_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor x_35_pad_type_0 = const()[name = tensor("x_35_pad_type_0"), val = tensor("valid")]; + tensor x_35_strides_0 = const()[name = tensor("x_35_strides_0"), val = tensor([1, 1])]; + tensor x_35_pad_0 = const()[name = tensor("x_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_35_dilations_0 = const()[name = tensor("x_35_dilations_0"), val = tensor([1, 1])]; + tensor x_35_groups_0 = const()[name = tensor("x_35_groups_0"), val = tensor(1)]; + tensor layers_5_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109940288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110726784))), name = tensor("layers_5_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_35_cast_fp16 = conv(dilations = x_35_dilations_0, groups = x_35_groups_0, pad = x_35_pad_0, pad_type = x_35_pad_type_0, strides = x_35_strides_0, weight = layers_5_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_165_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor inputs_57_cast_fp16 = add(x = inputs_55_cast_fp16, y = x_35_cast_fp16)[name = tensor("inputs_57_cast_fp16")]; + tensor out_57_axes_0 = const()[name = tensor("out_57_axes_0"), val = tensor([1])]; + tensor var_1771_to_fp16 = const()[name = tensor("op_1771_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_1771_to_fp16, x = inputs_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor input_167_gamma_0_to_fp16 = const()[name = tensor("input_167_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110726976)))]; + tensor input_167_beta_0_to_fp16 = const()[name = tensor("input_167_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110729088)))]; + tensor input_167_epsilon_0_to_fp16 = const()[name = tensor("input_167_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = batch_norm(beta = input_167_beta_0_to_fp16, epsilon = input_167_epsilon_0_to_fp16, gamma = input_167_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_pad_type_0 = const()[name = tensor("input_169_pad_type_0"), val = tensor("valid")]; + tensor input_169_strides_0 = const()[name = tensor("input_169_strides_0"), val = tensor([1, 1])]; + tensor input_169_pad_0 = const()[name = tensor("input_169_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_169_dilations_0 = const()[name = tensor("input_169_dilations_0"), val = tensor([1, 1])]; + tensor input_169_groups_0 = const()[name = tensor("input_169_groups_0"), val = tensor(1)]; + tensor layers_5_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110731200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113876992))), name = tensor("layers_5_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_169_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_169_dilations_0, groups = input_169_groups_0, pad = input_169_pad_0, pad_type = input_169_pad_type_0, strides = input_169_strides_0, weight = layers_5_feed_forward2_fc1_weight_to_fp16_palettized, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_cast_fp16 = silu(x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor x_37_pad_type_0 = const()[name = tensor("x_37_pad_type_0"), val = tensor("valid")]; + tensor x_37_strides_0 = const()[name = tensor("x_37_strides_0"), val = tensor([1, 1])]; + tensor x_37_pad_0 = const()[name = tensor("x_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_37_dilations_0 = const()[name = tensor("x_37_dilations_0"), val = tensor([1, 1])]; + tensor x_37_groups_0 = const()[name = tensor("x_37_groups_0"), val = tensor(1)]; + tensor op_1799_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113877184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117022976))), name = tensor("op_1799_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_1799_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_37_dilations_0, groups = x_37_groups_0, pad = x_37_pad_0, pad_type = x_37_pad_type_0, strides = x_37_strides_0, weight = op_1799_weight_0_to_fp16_palettized, x = input_171_cast_fp16)[name = tensor("op_1799_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_1799_cast_fp16)[name = tensor("inputs_59_cast_fp16")]; + tensor out_59_axes_0 = const()[name = tensor("out_59_axes_0"), val = tensor([1])]; + tensor var_1809_to_fp16 = const()[name = tensor("op_1809_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_1809_to_fp16, x = inputs_59_cast_fp16)[name = tensor("out_59_cast_fp16")]; + tensor inputs_61_gamma_0_to_fp16 = const()[name = tensor("inputs_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117023168)))]; + tensor inputs_61_beta_0_to_fp16 = const()[name = tensor("inputs_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117025280)))]; + tensor inputs_61_epsilon_0_to_fp16 = const()[name = tensor("inputs_61_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_61_cast_fp16 = batch_norm(beta = inputs_61_beta_0_to_fp16, epsilon = inputs_61_epsilon_0_to_fp16, gamma = inputs_61_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_1823 = const()[name = tensor("op_1823"), val = tensor(3)]; + tensor out_61_axes_0 = const()[name = tensor("out_61_axes_0"), val = tensor([1])]; + tensor var_1854_to_fp16 = const()[name = tensor("op_1854_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_1854_to_fp16, x = inputs_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor input_173_gamma_0_to_fp16 = const()[name = tensor("input_173_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117027392)))]; + tensor input_173_beta_0_to_fp16 = const()[name = tensor("input_173_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117029504)))]; + tensor input_173_epsilon_0_to_fp16 = const()[name = tensor("input_173_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = batch_norm(beta = input_173_beta_0_to_fp16, epsilon = input_173_epsilon_0_to_fp16, gamma = input_173_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor input_175_pad_type_0 = const()[name = tensor("input_175_pad_type_0"), val = tensor("valid")]; + tensor input_175_strides_0 = const()[name = tensor("input_175_strides_0"), val = tensor([1, 1])]; + tensor input_175_pad_0 = const()[name = tensor("input_175_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_175_dilations_0 = const()[name = tensor("input_175_dilations_0"), val = tensor([1, 1])]; + tensor input_175_groups_0 = const()[name = tensor("input_175_groups_0"), val = tensor(1)]; + tensor layers_6_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117031616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120177408))), name = tensor("layers_6_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_175_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_175_dilations_0, groups = input_175_groups_0, pad = input_175_pad_0, pad_type = input_175_pad_type_0, strides = input_175_strides_0, weight = layers_6_feed_forward1_fc1_weight_to_fp16_palettized, x = input_173_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_cast_fp16 = silu(x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor x_39_pad_type_0 = const()[name = tensor("x_39_pad_type_0"), val = tensor("valid")]; + tensor x_39_strides_0 = const()[name = tensor("x_39_strides_0"), val = tensor([1, 1])]; + tensor x_39_pad_0 = const()[name = tensor("x_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_39_dilations_0 = const()[name = tensor("x_39_dilations_0"), val = tensor([1, 1])]; + tensor x_39_groups_0 = const()[name = tensor("x_39_groups_0"), val = tensor(1)]; + tensor op_1882_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(120177600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123323392))), name = tensor("op_1882_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_1882_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_39_dilations_0, groups = x_39_groups_0, pad = x_39_pad_0, pad_type = x_39_pad_type_0, strides = x_39_strides_0, weight = op_1882_weight_0_to_fp16_palettized, x = input_177_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_1882_cast_fp16)[name = tensor("inputs_63_cast_fp16")]; + tensor out_63_axes_0 = const()[name = tensor("out_63_axes_0"), val = tensor([1])]; + tensor var_1892_to_fp16 = const()[name = tensor("op_1892_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_1892_to_fp16, x = inputs_63_cast_fp16)[name = tensor("out_63_cast_fp16")]; + tensor obj_27_gamma_0_to_fp16 = const()[name = tensor("obj_27_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123323584)))]; + tensor obj_27_beta_0_to_fp16 = const()[name = tensor("obj_27_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123325696)))]; + tensor obj_27_epsilon_0_to_fp16 = const()[name = tensor("obj_27_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_27_cast_fp16 = batch_norm(beta = obj_27_beta_0_to_fp16, epsilon = obj_27_epsilon_0_to_fp16, gamma = obj_27_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_63_cast_fp16)[name = tensor("obj_27_cast_fp16")]; + tensor query_25_pad_type_0 = const()[name = tensor("query_25_pad_type_0"), val = tensor("valid")]; + tensor query_25_strides_0 = const()[name = tensor("query_25_strides_0"), val = tensor([1, 1])]; + tensor query_25_pad_0 = const()[name = tensor("query_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_25_dilations_0 = const()[name = tensor("query_25_dilations_0"), val = tensor([1, 1])]; + tensor query_25_groups_0 = const()[name = tensor("query_25_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(123327808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124114304))), name = tensor("layers_6_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_25_dilations_0, groups = query_25_groups_0, pad = query_25_pad_0, pad_type = query_25_pad_type_0, strides = query_25_strides_0, weight = layers_6_self_attn_q_proj_weight_to_fp16_palettized, x = obj_27_cast_fp16)[name = tensor("query_25_cast_fp16")]; + tensor key_13_pad_type_0 = const()[name = tensor("key_13_pad_type_0"), val = tensor("valid")]; + tensor key_13_strides_0 = const()[name = tensor("key_13_strides_0"), val = tensor([1, 1])]; + tensor key_13_pad_0 = const()[name = tensor("key_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_13_dilations_0 = const()[name = tensor("key_13_dilations_0"), val = tensor([1, 1])]; + tensor key_13_groups_0 = const()[name = tensor("key_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124114496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124900992))), name = tensor("layers_6_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_13_cast_fp16 = conv(dilations = key_13_dilations_0, groups = key_13_groups_0, pad = key_13_pad_0, pad_type = key_13_pad_type_0, strides = key_13_strides_0, weight = layers_6_self_attn_k_proj_weight_to_fp16_palettized, x = obj_27_cast_fp16)[name = tensor("key_13_cast_fp16")]; + tensor value_13_pad_type_0 = const()[name = tensor("value_13_pad_type_0"), val = tensor("valid")]; + tensor value_13_strides_0 = const()[name = tensor("value_13_strides_0"), val = tensor([1, 1])]; + tensor value_13_pad_0 = const()[name = tensor("value_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_13_dilations_0 = const()[name = tensor("value_13_dilations_0"), val = tensor([1, 1])]; + tensor value_13_groups_0 = const()[name = tensor("value_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(124901184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125687680))), name = tensor("layers_6_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_13_dilations_0, groups = value_13_groups_0, pad = value_13_pad_0, pad_type = value_13_pad_type_0, strides = value_13_strides_0, weight = layers_6_self_attn_v_proj_weight_to_fp16_palettized, x = obj_27_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_1930_to_fp16 = const()[name = tensor("op_1930_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125687872)))]; + tensor query_27_cast_fp16 = add(x = query_25_cast_fp16, y = var_1930_to_fp16)[name = tensor("query_27_cast_fp16")]; + tensor var_1933_to_fp16 = const()[name = tensor("op_1933_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125689984)))]; + tensor q_with_bias_v_13_cast_fp16 = add(x = query_25_cast_fp16, y = var_1933_to_fp16)[name = tensor("q_with_bias_v_13_cast_fp16")]; + tensor p_13_pad_type_0 = const()[name = tensor("p_13_pad_type_0"), val = tensor("valid")]; + tensor p_13_strides_0 = const()[name = tensor("p_13_strides_0"), val = tensor([1, 1])]; + tensor p_13_pad_0 = const()[name = tensor("p_13_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_13_dilations_0 = const()[name = tensor("p_13_dilations_0"), val = tensor([1, 1])]; + tensor p_13_groups_0 = const()[name = tensor("p_13_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(125692096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126478592))), name = tensor("layers_6_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_13_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_13_dilations_0, groups = p_13_groups_0, pad = p_13_pad_0, pad_type = p_13_pad_type_0, strides = p_13_strides_0, weight = layers_6_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_13_cast_fp16")]; + tensor var_1944 = const()[name = tensor("op_1944"), val = tensor([1, 8, 128, 188])]; + tensor var_1945_cast_fp16 = reshape(shape = var_1944, x = q_with_bias_v_13_cast_fp16)[name = tensor("op_1945_cast_fp16")]; + tensor var_1946 = const()[name = tensor("op_1946"), val = tensor([1, 8, 128, -1])]; + tensor var_1947_cast_fp16 = reshape(shape = var_1946, x = p_13_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor matrix_bd_49_transpose_x_0 = const()[name = tensor("matrix_bd_49_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_49_transpose_y_0 = const()[name = tensor("matrix_bd_49_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_49_cast_fp16 = matmul(transpose_x = matrix_bd_49_transpose_x_0, transpose_y = matrix_bd_49_transpose_y_0, x = var_1945_cast_fp16, y = var_1947_cast_fp16)[name = tensor("matrix_bd_49_cast_fp16")]; + tensor matrix_bd_51_pad_0 = const()[name = tensor("matrix_bd_51_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_51_mode_0 = const()[name = tensor("matrix_bd_51_mode_0"), val = tensor("constant")]; + tensor const_76_to_fp16 = const()[name = tensor("const_76_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_51_cast_fp16 = pad(constant_val = const_76_to_fp16, mode = matrix_bd_51_mode_0, pad = matrix_bd_51_pad_0, x = matrix_bd_49_cast_fp16)[name = tensor("matrix_bd_51_cast_fp16")]; + tensor var_1956 = const()[name = tensor("op_1956"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_1956, x = matrix_bd_51_cast_fp16)[name = tensor("matrix_bd_53_cast_fp16")]; + tensor var_1960_begin_0 = const()[name = tensor("op_1960_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1960_end_0 = const()[name = tensor("op_1960_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1960_end_mask_0 = const()[name = tensor("op_1960_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1960_cast_fp16 = slice_by_index(begin = var_1960_begin_0, end = var_1960_end_0, end_mask = var_1960_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor var_1961 = const()[name = tensor("op_1961"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_55_cast_fp16 = reshape(shape = var_1961, x = var_1960_cast_fp16)[name = tensor("matrix_bd_55_cast_fp16")]; + tensor var_1966_begin_0 = const()[name = tensor("op_1966_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1966_end_0 = const()[name = tensor("op_1966_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1966_end_mask_0 = const()[name = tensor("op_1966_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1966_cast_fp16 = slice_by_index(begin = var_1966_begin_0, end = var_1966_end_0, end_mask = var_1966_end_mask_0, x = matrix_bd_55_cast_fp16)[name = tensor("op_1966_cast_fp16")]; + tensor var_1967_to_fp16 = const()[name = tensor("op_1967_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_13_cast_fp16 = mul(x = var_1966_cast_fp16, y = var_1967_to_fp16)[name = tensor("qk_mask_13_cast_fp16")]; + tensor var_1971 = const()[name = tensor("op_1971"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_1971, x = query_27_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1974_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1973_to_fp16)[name = tensor("op_1974_cast_fp16")]; + tensor var_1977 = const()[name = tensor("op_1977"), val = tensor([1, 8, 128, 188])]; + tensor var_1978_cast_fp16 = reshape(shape = var_1977, x = key_13_cast_fp16)[name = tensor("op_1978_cast_fp16")]; + tensor mh_w_25_transpose_x_0 = const()[name = tensor("mh_w_25_transpose_x_0"), val = tensor(true)]; + tensor mh_w_25_transpose_y_0 = const()[name = tensor("mh_w_25_transpose_y_0"), val = tensor(false)]; + tensor mh_w_25_cast_fp16 = matmul(transpose_x = mh_w_25_transpose_x_0, transpose_y = mh_w_25_transpose_y_0, x = var_1974_cast_fp16, y = var_1978_cast_fp16)[name = tensor("mh_w_25_cast_fp16")]; + tensor mh_w_27_cast_fp16 = add(x = mh_w_25_cast_fp16, y = qk_mask_13_cast_fp16)[name = tensor("mh_w_27_cast_fp16")]; + tensor var_1982_cast_fp16 = softmax(axis = var_1823, x = mh_w_27_cast_fp16)[name = tensor("op_1982_cast_fp16")]; + tensor var_1983 = const()[name = tensor("op_1983"), val = tensor([1, 8, 128, 188])]; + tensor var_1984_cast_fp16 = reshape(shape = var_1983, x = value_13_cast_fp16)[name = tensor("op_1984_cast_fp16")]; + tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; + tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; + tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1984_cast_fp16, y = var_1982_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1987 = const()[name = tensor("op_1987"), val = tensor([1, 1024, 1, 188])]; + tensor input_179_cast_fp16 = reshape(shape = var_1987, x = attn_13_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor obj_29_pad_type_0 = const()[name = tensor("obj_29_pad_type_0"), val = tensor("valid")]; + tensor obj_29_strides_0 = const()[name = tensor("obj_29_strides_0"), val = tensor([1, 1])]; + tensor obj_29_pad_0 = const()[name = tensor("obj_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_29_dilations_0 = const()[name = tensor("obj_29_dilations_0"), val = tensor([1, 1])]; + tensor obj_29_groups_0 = const()[name = tensor("obj_29_groups_0"), val = tensor(1)]; + tensor layers_6_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126478784))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127265280))), name = tensor("layers_6_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_29_dilations_0, groups = obj_29_groups_0, pad = obj_29_pad_0, pad_type = obj_29_pad_type_0, strides = obj_29_strides_0, weight = layers_6_self_attn_o_proj_weight_to_fp16_palettized, x = input_179_cast_fp16)[name = tensor("obj_29_cast_fp16")]; + tensor inputs_65_cast_fp16 = add(x = inputs_63_cast_fp16, y = obj_29_cast_fp16)[name = tensor("inputs_65_cast_fp16")]; + tensor out_65_axes_0 = const()[name = tensor("out_65_axes_0"), val = tensor([1])]; + tensor var_2005_to_fp16 = const()[name = tensor("op_2005_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2005_to_fp16, x = inputs_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_181_gamma_0_to_fp16 = const()[name = tensor("input_181_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127265472)))]; + tensor input_181_beta_0_to_fp16 = const()[name = tensor("input_181_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127267584)))]; + tensor input_181_epsilon_0_to_fp16 = const()[name = tensor("input_181_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_181_cast_fp16 = batch_norm(beta = input_181_beta_0_to_fp16, epsilon = input_181_epsilon_0_to_fp16, gamma = input_181_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_pad_type_0 = const()[name = tensor("input_183_pad_type_0"), val = tensor("valid")]; + tensor input_183_strides_0 = const()[name = tensor("input_183_strides_0"), val = tensor([1, 1])]; + tensor input_183_pad_0 = const()[name = tensor("input_183_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_183_dilations_0 = const()[name = tensor("input_183_dilations_0"), val = tensor([1, 1])]; + tensor input_183_groups_0 = const()[name = tensor("input_183_groups_0"), val = tensor(1)]; + tensor layers_6_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127269696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128842624))), name = tensor("layers_6_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_183_cast_fp16 = conv(dilations = input_183_dilations_0, groups = input_183_groups_0, pad = input_183_pad_0, pad_type = input_183_pad_type_0, strides = input_183_strides_0, weight = layers_6_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_split_num_splits_0 = const()[name = tensor("input_185_split_num_splits_0"), val = tensor(2)]; + tensor input_185_split_axis_0 = const()[name = tensor("input_185_split_axis_0"), val = tensor(1)]; + tensor input_185_split_cast_fp16_0, tensor input_185_split_cast_fp16_1 = split(axis = input_185_split_axis_0, num_splits = input_185_split_num_splits_0, x = input_183_cast_fp16)[name = tensor("input_185_split_cast_fp16")]; + tensor input_185_split_1_sigmoid_cast_fp16 = sigmoid(x = input_185_split_cast_fp16_1)[name = tensor("input_185_split_1_sigmoid_cast_fp16")]; + tensor input_185_cast_fp16 = mul(x = input_185_split_cast_fp16_0, y = input_185_split_1_sigmoid_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor input_187_pad_type_0 = const()[name = tensor("input_187_pad_type_0"), val = tensor("custom")]; + tensor input_187_pad_0 = const()[name = tensor("input_187_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_187_groups_0 = const()[name = tensor("input_187_groups_0"), val = tensor(1024)]; + tensor input_187_strides_0 = const()[name = tensor("input_187_strides_0"), val = tensor([1, 1])]; + tensor input_187_dilations_0 = const()[name = tensor("input_187_dilations_0"), val = tensor([1, 1])]; + tensor const_280_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128842816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128849792))), name = tensor("const_280_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_281_to_fp16 = const()[name = tensor("const_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128849984)))]; + tensor input_189_cast_fp16 = conv(bias = const_281_to_fp16, dilations = input_187_dilations_0, groups = input_187_groups_0, pad = input_187_pad_0, pad_type = input_187_pad_type_0, strides = input_187_strides_0, weight = const_280_to_fp16_palettized, x = input_185_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_cast_fp16 = silu(x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor x_41_pad_type_0 = const()[name = tensor("x_41_pad_type_0"), val = tensor("valid")]; + tensor x_41_strides_0 = const()[name = tensor("x_41_strides_0"), val = tensor([1, 1])]; + tensor x_41_pad_0 = const()[name = tensor("x_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_41_dilations_0 = const()[name = tensor("x_41_dilations_0"), val = tensor([1, 1])]; + tensor x_41_groups_0 = const()[name = tensor("x_41_groups_0"), val = tensor(1)]; + tensor layers_6_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(128852096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129638592))), name = tensor("layers_6_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_41_cast_fp16 = conv(dilations = x_41_dilations_0, groups = x_41_groups_0, pad = x_41_pad_0, pad_type = x_41_pad_type_0, strides = x_41_strides_0, weight = layers_6_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_191_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor inputs_67_cast_fp16 = add(x = inputs_65_cast_fp16, y = x_41_cast_fp16)[name = tensor("inputs_67_cast_fp16")]; + tensor out_67_axes_0 = const()[name = tensor("out_67_axes_0"), val = tensor([1])]; + tensor var_2053_to_fp16 = const()[name = tensor("op_2053_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2053_to_fp16, x = inputs_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor input_193_gamma_0_to_fp16 = const()[name = tensor("input_193_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129638784)))]; + tensor input_193_beta_0_to_fp16 = const()[name = tensor("input_193_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129640896)))]; + tensor input_193_epsilon_0_to_fp16 = const()[name = tensor("input_193_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_193_cast_fp16 = batch_norm(beta = input_193_beta_0_to_fp16, epsilon = input_193_epsilon_0_to_fp16, gamma = input_193_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_pad_type_0 = const()[name = tensor("input_195_pad_type_0"), val = tensor("valid")]; + tensor input_195_strides_0 = const()[name = tensor("input_195_strides_0"), val = tensor([1, 1])]; + tensor input_195_pad_0 = const()[name = tensor("input_195_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_195_dilations_0 = const()[name = tensor("input_195_dilations_0"), val = tensor([1, 1])]; + tensor input_195_groups_0 = const()[name = tensor("input_195_groups_0"), val = tensor(1)]; + tensor layers_6_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129643008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132788800))), name = tensor("layers_6_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_195_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_195_dilations_0, groups = input_195_groups_0, pad = input_195_pad_0, pad_type = input_195_pad_type_0, strides = input_195_strides_0, weight = layers_6_feed_forward2_fc1_weight_to_fp16_palettized, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor x_43_pad_type_0 = const()[name = tensor("x_43_pad_type_0"), val = tensor("valid")]; + tensor x_43_strides_0 = const()[name = tensor("x_43_strides_0"), val = tensor([1, 1])]; + tensor x_43_pad_0 = const()[name = tensor("x_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_43_dilations_0 = const()[name = tensor("x_43_dilations_0"), val = tensor([1, 1])]; + tensor x_43_groups_0 = const()[name = tensor("x_43_groups_0"), val = tensor(1)]; + tensor op_2081_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132788992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135934784))), name = tensor("op_2081_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_2081_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, 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 = op_2081_weight_0_to_fp16_palettized, x = input_197_cast_fp16)[name = tensor("op_2081_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_2081_cast_fp16)[name = tensor("inputs_69_cast_fp16")]; + tensor out_69_axes_0 = const()[name = tensor("out_69_axes_0"), val = tensor([1])]; + tensor var_2091_to_fp16 = const()[name = tensor("op_2091_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2091_to_fp16, x = inputs_69_cast_fp16)[name = tensor("out_69_cast_fp16")]; + tensor inputs_71_gamma_0_to_fp16 = const()[name = tensor("inputs_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135934976)))]; + tensor inputs_71_beta_0_to_fp16 = const()[name = tensor("inputs_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135937088)))]; + tensor inputs_71_epsilon_0_to_fp16 = const()[name = tensor("inputs_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_71_cast_fp16 = batch_norm(beta = inputs_71_beta_0_to_fp16, epsilon = inputs_71_epsilon_0_to_fp16, gamma = inputs_71_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor var_2105 = const()[name = tensor("op_2105"), val = tensor(3)]; + tensor out_71_axes_0 = const()[name = tensor("out_71_axes_0"), val = tensor([1])]; + tensor var_2136_to_fp16 = const()[name = tensor("op_2136_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2136_to_fp16, x = inputs_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_199_gamma_0_to_fp16 = const()[name = tensor("input_199_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135939200)))]; + tensor input_199_beta_0_to_fp16 = const()[name = tensor("input_199_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135941312)))]; + tensor input_199_epsilon_0_to_fp16 = const()[name = tensor("input_199_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_199_cast_fp16 = batch_norm(beta = input_199_beta_0_to_fp16, epsilon = input_199_epsilon_0_to_fp16, gamma = input_199_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_pad_type_0 = const()[name = tensor("input_201_pad_type_0"), val = tensor("valid")]; + tensor input_201_strides_0 = const()[name = tensor("input_201_strides_0"), val = tensor([1, 1])]; + tensor input_201_pad_0 = const()[name = tensor("input_201_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_201_dilations_0 = const()[name = tensor("input_201_dilations_0"), val = tensor([1, 1])]; + tensor input_201_groups_0 = const()[name = tensor("input_201_groups_0"), val = tensor(1)]; + tensor layers_7_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135943424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139089216))), name = tensor("layers_7_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_201_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_201_dilations_0, groups = input_201_groups_0, pad = input_201_pad_0, pad_type = input_201_pad_type_0, strides = input_201_strides_0, weight = layers_7_feed_forward1_fc1_weight_to_fp16_palettized, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_cast_fp16 = silu(x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor x_45_pad_type_0 = const()[name = tensor("x_45_pad_type_0"), val = tensor("valid")]; + tensor x_45_strides_0 = const()[name = tensor("x_45_strides_0"), val = tensor([1, 1])]; + tensor x_45_pad_0 = const()[name = tensor("x_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_45_dilations_0 = const()[name = tensor("x_45_dilations_0"), val = tensor([1, 1])]; + tensor x_45_groups_0 = const()[name = tensor("x_45_groups_0"), val = tensor(1)]; + tensor op_2164_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139089408))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142235200))), name = tensor("op_2164_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_2164_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_45_dilations_0, groups = x_45_groups_0, pad = x_45_pad_0, pad_type = x_45_pad_type_0, strides = x_45_strides_0, weight = op_2164_weight_0_to_fp16_palettized, x = input_203_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_2164_cast_fp16)[name = tensor("inputs_73_cast_fp16")]; + tensor out_73_axes_0 = const()[name = tensor("out_73_axes_0"), val = tensor([1])]; + tensor var_2174_to_fp16 = const()[name = tensor("op_2174_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2174_to_fp16, x = inputs_73_cast_fp16)[name = tensor("out_73_cast_fp16")]; + tensor obj_31_gamma_0_to_fp16 = const()[name = tensor("obj_31_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142235392)))]; + tensor obj_31_beta_0_to_fp16 = const()[name = tensor("obj_31_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142237504)))]; + tensor obj_31_epsilon_0_to_fp16 = const()[name = tensor("obj_31_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_31_cast_fp16 = batch_norm(beta = obj_31_beta_0_to_fp16, epsilon = obj_31_epsilon_0_to_fp16, gamma = obj_31_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_73_cast_fp16)[name = tensor("obj_31_cast_fp16")]; + tensor query_29_pad_type_0 = const()[name = tensor("query_29_pad_type_0"), val = tensor("valid")]; + tensor query_29_strides_0 = const()[name = tensor("query_29_strides_0"), val = tensor([1, 1])]; + tensor query_29_pad_0 = const()[name = tensor("query_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_29_dilations_0 = const()[name = tensor("query_29_dilations_0"), val = tensor([1, 1])]; + tensor query_29_groups_0 = const()[name = tensor("query_29_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142239616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143026112))), name = tensor("layers_7_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_29_dilations_0, groups = query_29_groups_0, pad = query_29_pad_0, pad_type = query_29_pad_type_0, strides = query_29_strides_0, weight = layers_7_self_attn_q_proj_weight_to_fp16_palettized, x = obj_31_cast_fp16)[name = tensor("query_29_cast_fp16")]; + tensor key_15_pad_type_0 = const()[name = tensor("key_15_pad_type_0"), val = tensor("valid")]; + tensor key_15_strides_0 = const()[name = tensor("key_15_strides_0"), val = tensor([1, 1])]; + tensor key_15_pad_0 = const()[name = tensor("key_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_15_dilations_0 = const()[name = tensor("key_15_dilations_0"), val = tensor([1, 1])]; + tensor key_15_groups_0 = const()[name = tensor("key_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143026304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143812800))), name = tensor("layers_7_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_15_cast_fp16 = conv(dilations = key_15_dilations_0, groups = key_15_groups_0, pad = key_15_pad_0, pad_type = key_15_pad_type_0, strides = key_15_strides_0, weight = layers_7_self_attn_k_proj_weight_to_fp16_palettized, x = obj_31_cast_fp16)[name = tensor("key_15_cast_fp16")]; + tensor value_15_pad_type_0 = const()[name = tensor("value_15_pad_type_0"), val = tensor("valid")]; + tensor value_15_strides_0 = const()[name = tensor("value_15_strides_0"), val = tensor([1, 1])]; + tensor value_15_pad_0 = const()[name = tensor("value_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_15_dilations_0 = const()[name = tensor("value_15_dilations_0"), val = tensor([1, 1])]; + tensor value_15_groups_0 = const()[name = tensor("value_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143812992))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144599488))), name = tensor("layers_7_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_15_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_15_dilations_0, groups = value_15_groups_0, pad = value_15_pad_0, pad_type = value_15_pad_type_0, strides = value_15_strides_0, weight = layers_7_self_attn_v_proj_weight_to_fp16_palettized, x = obj_31_cast_fp16)[name = tensor("value_15_cast_fp16")]; + tensor var_2212_to_fp16 = const()[name = tensor("op_2212_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144599680)))]; + tensor query_31_cast_fp16 = add(x = query_29_cast_fp16, y = var_2212_to_fp16)[name = tensor("query_31_cast_fp16")]; + tensor var_2215_to_fp16 = const()[name = tensor("op_2215_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144601792)))]; + tensor q_with_bias_v_15_cast_fp16 = add(x = query_29_cast_fp16, y = var_2215_to_fp16)[name = tensor("q_with_bias_v_15_cast_fp16")]; + tensor p_15_pad_type_0 = const()[name = tensor("p_15_pad_type_0"), val = tensor("valid")]; + tensor p_15_strides_0 = const()[name = tensor("p_15_strides_0"), val = tensor([1, 1])]; + tensor p_15_pad_0 = const()[name = tensor("p_15_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_15_dilations_0 = const()[name = tensor("p_15_dilations_0"), val = tensor([1, 1])]; + tensor p_15_groups_0 = const()[name = tensor("p_15_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144603904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145390400))), name = tensor("layers_7_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_15_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_15_dilations_0, groups = p_15_groups_0, pad = p_15_pad_0, pad_type = p_15_pad_type_0, strides = p_15_strides_0, weight = layers_7_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_15_cast_fp16")]; + tensor var_2226 = const()[name = tensor("op_2226"), val = tensor([1, 8, 128, 188])]; + tensor var_2227_cast_fp16 = reshape(shape = var_2226, x = q_with_bias_v_15_cast_fp16)[name = tensor("op_2227_cast_fp16")]; + tensor var_2228 = const()[name = tensor("op_2228"), val = tensor([1, 8, 128, -1])]; + tensor var_2229_cast_fp16 = reshape(shape = var_2228, x = p_15_cast_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor matrix_bd_57_transpose_x_0 = const()[name = tensor("matrix_bd_57_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_57_transpose_y_0 = const()[name = tensor("matrix_bd_57_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_57_cast_fp16 = matmul(transpose_x = matrix_bd_57_transpose_x_0, transpose_y = matrix_bd_57_transpose_y_0, x = var_2227_cast_fp16, y = var_2229_cast_fp16)[name = tensor("matrix_bd_57_cast_fp16")]; + tensor matrix_bd_59_pad_0 = const()[name = tensor("matrix_bd_59_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_59_mode_0 = const()[name = tensor("matrix_bd_59_mode_0"), val = tensor("constant")]; + tensor const_87_to_fp16 = const()[name = tensor("const_87_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_59_cast_fp16 = pad(constant_val = const_87_to_fp16, mode = matrix_bd_59_mode_0, pad = matrix_bd_59_pad_0, x = matrix_bd_57_cast_fp16)[name = tensor("matrix_bd_59_cast_fp16")]; + tensor var_2238 = const()[name = tensor("op_2238"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_2238, x = matrix_bd_59_cast_fp16)[name = tensor("matrix_bd_61_cast_fp16")]; + tensor var_2242_begin_0 = const()[name = tensor("op_2242_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2242_end_0 = const()[name = tensor("op_2242_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_2242_end_mask_0 = const()[name = tensor("op_2242_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2242_cast_fp16 = slice_by_index(begin = var_2242_begin_0, end = var_2242_end_0, end_mask = var_2242_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor var_2243 = const()[name = tensor("op_2243"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_63_cast_fp16 = reshape(shape = var_2243, x = var_2242_cast_fp16)[name = tensor("matrix_bd_63_cast_fp16")]; + tensor var_2248_begin_0 = const()[name = tensor("op_2248_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2248_end_0 = const()[name = tensor("op_2248_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_2248_end_mask_0 = const()[name = tensor("op_2248_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, x = matrix_bd_63_cast_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor var_2249_to_fp16 = const()[name = tensor("op_2249_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_15_cast_fp16 = mul(x = var_2248_cast_fp16, y = var_2249_to_fp16)[name = tensor("qk_mask_15_cast_fp16")]; + tensor var_2253 = const()[name = tensor("op_2253"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_2253, x = query_31_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_2255_to_fp16 = const()[name = tensor("op_2255_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_2256_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_2255_to_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor var_2259 = const()[name = tensor("op_2259"), val = tensor([1, 8, 128, 188])]; + tensor var_2260_cast_fp16 = reshape(shape = var_2259, x = key_15_cast_fp16)[name = tensor("op_2260_cast_fp16")]; + tensor mh_w_29_transpose_x_0 = const()[name = tensor("mh_w_29_transpose_x_0"), val = tensor(true)]; + tensor mh_w_29_transpose_y_0 = const()[name = tensor("mh_w_29_transpose_y_0"), val = tensor(false)]; + tensor mh_w_29_cast_fp16 = matmul(transpose_x = mh_w_29_transpose_x_0, transpose_y = mh_w_29_transpose_y_0, x = var_2256_cast_fp16, y = var_2260_cast_fp16)[name = tensor("mh_w_29_cast_fp16")]; + tensor mh_w_31_cast_fp16 = add(x = mh_w_29_cast_fp16, y = qk_mask_15_cast_fp16)[name = tensor("mh_w_31_cast_fp16")]; + tensor var_2264_cast_fp16 = softmax(axis = var_2105, x = mh_w_31_cast_fp16)[name = tensor("op_2264_cast_fp16")]; + tensor var_2265 = const()[name = tensor("op_2265"), val = tensor([1, 8, 128, 188])]; + tensor var_2266_cast_fp16 = reshape(shape = var_2265, x = value_15_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor attn_15_transpose_x_0 = const()[name = tensor("attn_15_transpose_x_0"), val = tensor(false)]; + tensor attn_15_transpose_y_0 = const()[name = tensor("attn_15_transpose_y_0"), val = tensor(true)]; + tensor attn_15_cast_fp16 = matmul(transpose_x = attn_15_transpose_x_0, transpose_y = attn_15_transpose_y_0, x = var_2266_cast_fp16, y = var_2264_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2269 = const()[name = tensor("op_2269"), val = tensor([1, 1024, 1, 188])]; + tensor input_205_cast_fp16 = reshape(shape = var_2269, x = attn_15_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor obj_33_pad_type_0 = const()[name = tensor("obj_33_pad_type_0"), val = tensor("valid")]; + tensor obj_33_strides_0 = const()[name = tensor("obj_33_strides_0"), val = tensor([1, 1])]; + tensor obj_33_pad_0 = const()[name = tensor("obj_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_33_dilations_0 = const()[name = tensor("obj_33_dilations_0"), val = tensor([1, 1])]; + tensor obj_33_groups_0 = const()[name = tensor("obj_33_groups_0"), val = tensor(1)]; + tensor layers_7_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(145390592))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146177088))), name = tensor("layers_7_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_33_dilations_0, groups = obj_33_groups_0, pad = obj_33_pad_0, pad_type = obj_33_pad_type_0, strides = obj_33_strides_0, weight = layers_7_self_attn_o_proj_weight_to_fp16_palettized, x = input_205_cast_fp16)[name = tensor("obj_33_cast_fp16")]; + tensor inputs_75_cast_fp16 = add(x = inputs_73_cast_fp16, y = obj_33_cast_fp16)[name = tensor("inputs_75_cast_fp16")]; + tensor out_75_axes_0 = const()[name = tensor("out_75_axes_0"), val = tensor([1])]; + tensor var_2287_to_fp16 = const()[name = tensor("op_2287_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2287_to_fp16, x = inputs_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor input_207_gamma_0_to_fp16 = const()[name = tensor("input_207_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146177280)))]; + tensor input_207_beta_0_to_fp16 = const()[name = tensor("input_207_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146179392)))]; + tensor input_207_epsilon_0_to_fp16 = const()[name = tensor("input_207_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = batch_norm(beta = input_207_beta_0_to_fp16, epsilon = input_207_epsilon_0_to_fp16, gamma = input_207_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_pad_type_0 = const()[name = tensor("input_209_pad_type_0"), val = tensor("valid")]; + tensor input_209_strides_0 = const()[name = tensor("input_209_strides_0"), val = tensor([1, 1])]; + tensor input_209_pad_0 = const()[name = tensor("input_209_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_209_dilations_0 = const()[name = tensor("input_209_dilations_0"), val = tensor([1, 1])]; + tensor input_209_groups_0 = const()[name = tensor("input_209_groups_0"), val = tensor(1)]; + tensor layers_7_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(146181504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147754432))), name = tensor("layers_7_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_209_cast_fp16 = conv(dilations = input_209_dilations_0, groups = input_209_groups_0, pad = input_209_pad_0, pad_type = input_209_pad_type_0, strides = input_209_strides_0, weight = layers_7_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor input_211_split_num_splits_0 = const()[name = tensor("input_211_split_num_splits_0"), val = tensor(2)]; + tensor input_211_split_axis_0 = const()[name = tensor("input_211_split_axis_0"), val = tensor(1)]; + tensor input_211_split_cast_fp16_0, tensor input_211_split_cast_fp16_1 = split(axis = input_211_split_axis_0, num_splits = input_211_split_num_splits_0, x = input_209_cast_fp16)[name = tensor("input_211_split_cast_fp16")]; + tensor input_211_split_1_sigmoid_cast_fp16 = sigmoid(x = input_211_split_cast_fp16_1)[name = tensor("input_211_split_1_sigmoid_cast_fp16")]; + tensor input_211_cast_fp16 = mul(x = input_211_split_cast_fp16_0, y = input_211_split_1_sigmoid_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_pad_type_0 = const()[name = tensor("input_213_pad_type_0"), val = tensor("custom")]; + tensor input_213_pad_0 = const()[name = tensor("input_213_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_213_groups_0 = const()[name = tensor("input_213_groups_0"), val = tensor(1024)]; + tensor input_213_strides_0 = const()[name = tensor("input_213_strides_0"), val = tensor([1, 1])]; + tensor input_213_dilations_0 = const()[name = tensor("input_213_dilations_0"), val = tensor([1, 1])]; + tensor const_282_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147754624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147761600))), name = tensor("const_282_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_283_to_fp16 = const()[name = tensor("const_283_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147761792)))]; + tensor input_215_cast_fp16 = conv(bias = const_283_to_fp16, dilations = input_213_dilations_0, groups = input_213_groups_0, pad = input_213_pad_0, pad_type = input_213_pad_type_0, strides = input_213_strides_0, weight = const_282_to_fp16_palettized, x = input_211_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor input_217_cast_fp16 = silu(x = input_215_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor x_47_pad_type_0 = const()[name = tensor("x_47_pad_type_0"), val = tensor("valid")]; + tensor x_47_strides_0 = const()[name = tensor("x_47_strides_0"), val = tensor([1, 1])]; + tensor x_47_pad_0 = const()[name = tensor("x_47_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_47_dilations_0 = const()[name = tensor("x_47_dilations_0"), val = tensor([1, 1])]; + tensor x_47_groups_0 = const()[name = tensor("x_47_groups_0"), val = tensor(1)]; + tensor layers_7_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147763904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148550400))), name = tensor("layers_7_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_47_cast_fp16 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = layers_7_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_217_cast_fp16)[name = tensor("x_47_cast_fp16")]; + tensor inputs_77_cast_fp16 = add(x = inputs_75_cast_fp16, y = x_47_cast_fp16)[name = tensor("inputs_77_cast_fp16")]; + tensor out_77_axes_0 = const()[name = tensor("out_77_axes_0"), val = tensor([1])]; + tensor var_2335_to_fp16 = const()[name = tensor("op_2335_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2335_to_fp16, x = inputs_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_219_gamma_0_to_fp16 = const()[name = tensor("input_219_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148550592)))]; + tensor input_219_beta_0_to_fp16 = const()[name = tensor("input_219_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148552704)))]; + tensor input_219_epsilon_0_to_fp16 = const()[name = tensor("input_219_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = batch_norm(beta = input_219_beta_0_to_fp16, epsilon = input_219_epsilon_0_to_fp16, gamma = input_219_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_pad_type_0 = const()[name = tensor("input_221_pad_type_0"), val = tensor("valid")]; + tensor input_221_strides_0 = const()[name = tensor("input_221_strides_0"), val = tensor([1, 1])]; + tensor input_221_pad_0 = const()[name = tensor("input_221_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_221_dilations_0 = const()[name = tensor("input_221_dilations_0"), val = tensor([1, 1])]; + tensor input_221_groups_0 = const()[name = tensor("input_221_groups_0"), val = tensor(1)]; + tensor layers_7_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148554816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151700608))), name = tensor("layers_7_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_221_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_221_dilations_0, groups = input_221_groups_0, pad = input_221_pad_0, pad_type = input_221_pad_type_0, strides = input_221_strides_0, weight = layers_7_feed_forward2_fc1_weight_to_fp16_palettized, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor input_223_cast_fp16 = silu(x = input_221_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor x_49_pad_type_0 = const()[name = tensor("x_49_pad_type_0"), val = tensor("valid")]; + tensor x_49_strides_0 = const()[name = tensor("x_49_strides_0"), val = tensor([1, 1])]; + tensor x_49_pad_0 = const()[name = tensor("x_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_49_dilations_0 = const()[name = tensor("x_49_dilations_0"), val = tensor([1, 1])]; + tensor x_49_groups_0 = const()[name = tensor("x_49_groups_0"), val = tensor(1)]; + tensor op_2363_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151700800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154846592))), name = tensor("op_2363_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_2363_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_49_dilations_0, groups = x_49_groups_0, pad = x_49_pad_0, pad_type = x_49_pad_type_0, strides = x_49_strides_0, weight = op_2363_weight_0_to_fp16_palettized, x = input_223_cast_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_2363_cast_fp16)[name = tensor("inputs_79_cast_fp16")]; + tensor out_79_axes_0 = const()[name = tensor("out_79_axes_0"), val = tensor([1])]; + tensor var_2373_to_fp16 = const()[name = tensor("op_2373_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2373_to_fp16, x = inputs_79_cast_fp16)[name = tensor("out_79_cast_fp16")]; + tensor inputs_81_gamma_0_to_fp16 = const()[name = tensor("inputs_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154846784)))]; + tensor inputs_81_beta_0_to_fp16 = const()[name = tensor("inputs_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154848896)))]; + tensor inputs_81_epsilon_0_to_fp16 = const()[name = tensor("inputs_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_81_cast_fp16 = batch_norm(beta = inputs_81_beta_0_to_fp16, epsilon = inputs_81_epsilon_0_to_fp16, gamma = inputs_81_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_2387 = const()[name = tensor("op_2387"), val = tensor(3)]; + tensor out_81_axes_0 = const()[name = tensor("out_81_axes_0"), val = tensor([1])]; + tensor var_2418_to_fp16 = const()[name = tensor("op_2418_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2418_to_fp16, x = inputs_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor input_225_gamma_0_to_fp16 = const()[name = tensor("input_225_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154851008)))]; + tensor input_225_beta_0_to_fp16 = const()[name = tensor("input_225_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154853120)))]; + tensor input_225_epsilon_0_to_fp16 = const()[name = tensor("input_225_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = batch_norm(beta = input_225_beta_0_to_fp16, epsilon = input_225_epsilon_0_to_fp16, gamma = input_225_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_pad_type_0 = const()[name = tensor("input_227_pad_type_0"), val = tensor("valid")]; + tensor input_227_strides_0 = const()[name = tensor("input_227_strides_0"), val = tensor([1, 1])]; + tensor input_227_pad_0 = const()[name = tensor("input_227_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_227_dilations_0 = const()[name = tensor("input_227_dilations_0"), val = tensor([1, 1])]; + tensor input_227_groups_0 = const()[name = tensor("input_227_groups_0"), val = tensor(1)]; + tensor layers_8_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154855232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158001024))), name = tensor("layers_8_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_227_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_227_dilations_0, groups = input_227_groups_0, pad = input_227_pad_0, pad_type = input_227_pad_type_0, strides = input_227_strides_0, weight = layers_8_feed_forward1_fc1_weight_to_fp16_palettized, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_cast_fp16 = silu(x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor x_51_pad_type_0 = const()[name = tensor("x_51_pad_type_0"), val = tensor("valid")]; + tensor x_51_strides_0 = const()[name = tensor("x_51_strides_0"), val = tensor([1, 1])]; + tensor x_51_pad_0 = const()[name = tensor("x_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_51_dilations_0 = const()[name = tensor("x_51_dilations_0"), val = tensor([1, 1])]; + tensor x_51_groups_0 = const()[name = tensor("x_51_groups_0"), val = tensor(1)]; + tensor op_2446_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158001216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161147008))), name = tensor("op_2446_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_2446_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = op_2446_weight_0_to_fp16_palettized, x = input_229_cast_fp16)[name = tensor("op_2446_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_2446_cast_fp16)[name = tensor("inputs_83_cast_fp16")]; + tensor out_83_axes_0 = const()[name = tensor("out_83_axes_0"), val = tensor([1])]; + tensor var_2456_to_fp16 = const()[name = tensor("op_2456_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_2456_to_fp16, x = inputs_83_cast_fp16)[name = tensor("out_83_cast_fp16")]; + tensor obj_35_gamma_0_to_fp16 = const()[name = tensor("obj_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161147200)))]; + tensor obj_35_beta_0_to_fp16 = const()[name = tensor("obj_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161149312)))]; + tensor obj_35_epsilon_0_to_fp16 = const()[name = tensor("obj_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_35_cast_fp16 = batch_norm(beta = obj_35_beta_0_to_fp16, epsilon = obj_35_epsilon_0_to_fp16, gamma = obj_35_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_83_cast_fp16)[name = tensor("obj_35_cast_fp16")]; + tensor query_33_pad_type_0 = const()[name = tensor("query_33_pad_type_0"), val = tensor("valid")]; + tensor query_33_strides_0 = const()[name = tensor("query_33_strides_0"), val = tensor([1, 1])]; + tensor query_33_pad_0 = const()[name = tensor("query_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_33_dilations_0 = const()[name = tensor("query_33_dilations_0"), val = tensor([1, 1])]; + tensor query_33_groups_0 = const()[name = tensor("query_33_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161151424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161937920))), name = tensor("layers_8_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_33_dilations_0, groups = query_33_groups_0, pad = query_33_pad_0, pad_type = query_33_pad_type_0, strides = query_33_strides_0, weight = layers_8_self_attn_q_proj_weight_to_fp16_palettized, x = obj_35_cast_fp16)[name = tensor("query_33_cast_fp16")]; + tensor key_17_pad_type_0 = const()[name = tensor("key_17_pad_type_0"), val = tensor("valid")]; + tensor key_17_strides_0 = const()[name = tensor("key_17_strides_0"), val = tensor([1, 1])]; + tensor key_17_pad_0 = const()[name = tensor("key_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_17_dilations_0 = const()[name = tensor("key_17_dilations_0"), val = tensor([1, 1])]; + tensor key_17_groups_0 = const()[name = tensor("key_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(161938112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162724608))), name = tensor("layers_8_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_17_cast_fp16 = conv(dilations = key_17_dilations_0, groups = key_17_groups_0, pad = key_17_pad_0, pad_type = key_17_pad_type_0, strides = key_17_strides_0, weight = layers_8_self_attn_k_proj_weight_to_fp16_palettized, x = obj_35_cast_fp16)[name = tensor("key_17_cast_fp16")]; + tensor value_17_pad_type_0 = const()[name = tensor("value_17_pad_type_0"), val = tensor("valid")]; + tensor value_17_strides_0 = const()[name = tensor("value_17_strides_0"), val = tensor([1, 1])]; + tensor value_17_pad_0 = const()[name = tensor("value_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_17_dilations_0 = const()[name = tensor("value_17_dilations_0"), val = tensor([1, 1])]; + tensor value_17_groups_0 = const()[name = tensor("value_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162724800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163511296))), name = tensor("layers_8_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_17_dilations_0, groups = value_17_groups_0, pad = value_17_pad_0, pad_type = value_17_pad_type_0, strides = value_17_strides_0, weight = layers_8_self_attn_v_proj_weight_to_fp16_palettized, x = obj_35_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_2494_to_fp16 = const()[name = tensor("op_2494_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163511488)))]; + tensor query_35_cast_fp16 = add(x = query_33_cast_fp16, y = var_2494_to_fp16)[name = tensor("query_35_cast_fp16")]; + tensor var_2497_to_fp16 = const()[name = tensor("op_2497_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163513600)))]; + tensor q_with_bias_v_17_cast_fp16 = add(x = query_33_cast_fp16, y = var_2497_to_fp16)[name = tensor("q_with_bias_v_17_cast_fp16")]; + tensor p_17_pad_type_0 = const()[name = tensor("p_17_pad_type_0"), val = tensor("valid")]; + tensor p_17_strides_0 = const()[name = tensor("p_17_strides_0"), val = tensor([1, 1])]; + tensor p_17_pad_0 = const()[name = tensor("p_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_17_dilations_0 = const()[name = tensor("p_17_dilations_0"), val = tensor([1, 1])]; + tensor p_17_groups_0 = const()[name = tensor("p_17_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163515712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164302208))), name = tensor("layers_8_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_17_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_17_dilations_0, groups = p_17_groups_0, pad = p_17_pad_0, pad_type = p_17_pad_type_0, strides = p_17_strides_0, weight = layers_8_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_17_cast_fp16")]; + tensor var_2508 = const()[name = tensor("op_2508"), val = tensor([1, 8, 128, 188])]; + tensor var_2509_cast_fp16 = reshape(shape = var_2508, x = q_with_bias_v_17_cast_fp16)[name = tensor("op_2509_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 8, 128, -1])]; + tensor var_2511_cast_fp16 = reshape(shape = var_2510, x = p_17_cast_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor matrix_bd_65_transpose_x_0 = const()[name = tensor("matrix_bd_65_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_65_transpose_y_0 = const()[name = tensor("matrix_bd_65_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_65_cast_fp16 = matmul(transpose_x = matrix_bd_65_transpose_x_0, transpose_y = matrix_bd_65_transpose_y_0, x = var_2509_cast_fp16, y = var_2511_cast_fp16)[name = tensor("matrix_bd_65_cast_fp16")]; + tensor matrix_bd_67_pad_0 = const()[name = tensor("matrix_bd_67_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_67_mode_0 = const()[name = tensor("matrix_bd_67_mode_0"), val = tensor("constant")]; + tensor const_98_to_fp16 = const()[name = tensor("const_98_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_67_cast_fp16 = pad(constant_val = const_98_to_fp16, mode = matrix_bd_67_mode_0, pad = matrix_bd_67_pad_0, x = matrix_bd_65_cast_fp16)[name = tensor("matrix_bd_67_cast_fp16")]; + tensor var_2520 = const()[name = tensor("op_2520"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_2520, x = matrix_bd_67_cast_fp16)[name = tensor("matrix_bd_69_cast_fp16")]; + tensor var_2524_begin_0 = const()[name = tensor("op_2524_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2524_end_0 = const()[name = tensor("op_2524_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_2524_end_mask_0 = const()[name = tensor("op_2524_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2524_cast_fp16 = slice_by_index(begin = var_2524_begin_0, end = var_2524_end_0, end_mask = var_2524_end_mask_0, x = matrix_bd_69_cast_fp16)[name = tensor("op_2524_cast_fp16")]; + tensor var_2525 = const()[name = tensor("op_2525"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_71_cast_fp16 = reshape(shape = var_2525, x = var_2524_cast_fp16)[name = tensor("matrix_bd_71_cast_fp16")]; + tensor var_2530_begin_0 = const()[name = tensor("op_2530_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2530_end_0 = const()[name = tensor("op_2530_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_2530_end_mask_0 = const()[name = tensor("op_2530_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_2530_cast_fp16 = slice_by_index(begin = var_2530_begin_0, end = var_2530_end_0, end_mask = var_2530_end_mask_0, x = matrix_bd_71_cast_fp16)[name = tensor("op_2530_cast_fp16")]; + tensor var_2531_to_fp16 = const()[name = tensor("op_2531_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_17_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2531_to_fp16)[name = tensor("qk_mask_17_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_2535, x = query_35_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_2537_to_fp16 = const()[name = tensor("op_2537_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_2538_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_2537_to_fp16)[name = tensor("op_2538_cast_fp16")]; + tensor var_2541 = const()[name = tensor("op_2541"), val = tensor([1, 8, 128, 188])]; + tensor var_2542_cast_fp16 = reshape(shape = var_2541, x = key_17_cast_fp16)[name = tensor("op_2542_cast_fp16")]; + tensor mh_w_33_transpose_x_0 = const()[name = tensor("mh_w_33_transpose_x_0"), val = tensor(true)]; + tensor mh_w_33_transpose_y_0 = const()[name = tensor("mh_w_33_transpose_y_0"), val = tensor(false)]; + tensor mh_w_33_cast_fp16 = matmul(transpose_x = mh_w_33_transpose_x_0, transpose_y = mh_w_33_transpose_y_0, x = var_2538_cast_fp16, y = var_2542_cast_fp16)[name = tensor("mh_w_33_cast_fp16")]; + tensor mh_w_35_cast_fp16 = add(x = mh_w_33_cast_fp16, y = qk_mask_17_cast_fp16)[name = tensor("mh_w_35_cast_fp16")]; + tensor var_2546_cast_fp16 = softmax(axis = var_2387, x = mh_w_35_cast_fp16)[name = tensor("op_2546_cast_fp16")]; + tensor var_2547 = const()[name = tensor("op_2547"), val = tensor([1, 8, 128, 188])]; + tensor var_2548_cast_fp16 = reshape(shape = var_2547, x = value_17_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor attn_17_transpose_x_0 = const()[name = tensor("attn_17_transpose_x_0"), val = tensor(false)]; + tensor attn_17_transpose_y_0 = const()[name = tensor("attn_17_transpose_y_0"), val = tensor(true)]; + tensor attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = var_2548_cast_fp16, y = var_2546_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([1, 1024, 1, 188])]; + tensor input_231_cast_fp16 = reshape(shape = var_2551, x = attn_17_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor obj_37_pad_type_0 = const()[name = tensor("obj_37_pad_type_0"), val = tensor("valid")]; + tensor obj_37_strides_0 = const()[name = tensor("obj_37_strides_0"), val = tensor([1, 1])]; + tensor obj_37_pad_0 = const()[name = tensor("obj_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_37_dilations_0 = const()[name = tensor("obj_37_dilations_0"), val = tensor([1, 1])]; + tensor obj_37_groups_0 = const()[name = tensor("obj_37_groups_0"), val = tensor(1)]; + tensor layers_8_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164302400))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165088896))), name = tensor("layers_8_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_37_dilations_0, groups = obj_37_groups_0, pad = obj_37_pad_0, pad_type = obj_37_pad_type_0, strides = obj_37_strides_0, weight = layers_8_self_attn_o_proj_weight_to_fp16_palettized, x = input_231_cast_fp16)[name = tensor("obj_37_cast_fp16")]; + tensor inputs_85_cast_fp16 = add(x = inputs_83_cast_fp16, y = obj_37_cast_fp16)[name = tensor("inputs_85_cast_fp16")]; + tensor out_85_axes_0 = const()[name = tensor("out_85_axes_0"), val = tensor([1])]; + tensor var_2569_to_fp16 = const()[name = tensor("op_2569_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_2569_to_fp16, x = inputs_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor input_233_gamma_0_to_fp16 = const()[name = tensor("input_233_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165089088)))]; + tensor input_233_beta_0_to_fp16 = const()[name = tensor("input_233_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165091200)))]; + tensor input_233_epsilon_0_to_fp16 = const()[name = tensor("input_233_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = batch_norm(beta = input_233_beta_0_to_fp16, epsilon = input_233_epsilon_0_to_fp16, gamma = input_233_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor input_235_pad_type_0 = const()[name = tensor("input_235_pad_type_0"), val = tensor("valid")]; + tensor input_235_strides_0 = const()[name = tensor("input_235_strides_0"), val = tensor([1, 1])]; + tensor input_235_pad_0 = const()[name = tensor("input_235_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_235_dilations_0 = const()[name = tensor("input_235_dilations_0"), val = tensor([1, 1])]; + tensor input_235_groups_0 = const()[name = tensor("input_235_groups_0"), val = tensor(1)]; + tensor layers_8_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165093312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166666240))), name = tensor("layers_8_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_235_cast_fp16 = conv(dilations = input_235_dilations_0, groups = input_235_groups_0, pad = input_235_pad_0, pad_type = input_235_pad_type_0, strides = input_235_strides_0, weight = layers_8_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_233_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_split_num_splits_0 = const()[name = tensor("input_237_split_num_splits_0"), val = tensor(2)]; + tensor input_237_split_axis_0 = const()[name = tensor("input_237_split_axis_0"), val = tensor(1)]; + tensor input_237_split_cast_fp16_0, tensor input_237_split_cast_fp16_1 = split(axis = input_237_split_axis_0, num_splits = input_237_split_num_splits_0, x = input_235_cast_fp16)[name = tensor("input_237_split_cast_fp16")]; + tensor input_237_split_1_sigmoid_cast_fp16 = sigmoid(x = input_237_split_cast_fp16_1)[name = tensor("input_237_split_1_sigmoid_cast_fp16")]; + tensor input_237_cast_fp16 = mul(x = input_237_split_cast_fp16_0, y = input_237_split_1_sigmoid_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_pad_type_0 = const()[name = tensor("input_239_pad_type_0"), val = tensor("custom")]; + tensor input_239_pad_0 = const()[name = tensor("input_239_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_239_groups_0 = const()[name = tensor("input_239_groups_0"), val = tensor(1024)]; + tensor input_239_strides_0 = const()[name = tensor("input_239_strides_0"), val = tensor([1, 1])]; + tensor input_239_dilations_0 = const()[name = tensor("input_239_dilations_0"), val = tensor([1, 1])]; + tensor const_284_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166666432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166673408))), name = tensor("const_284_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_285_to_fp16 = const()[name = tensor("const_285_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166673600)))]; + tensor input_241_cast_fp16 = conv(bias = const_285_to_fp16, dilations = input_239_dilations_0, groups = input_239_groups_0, pad = input_239_pad_0, pad_type = input_239_pad_type_0, strides = input_239_strides_0, weight = const_284_to_fp16_palettized, x = input_237_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_cast_fp16 = silu(x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor x_53_pad_type_0 = const()[name = tensor("x_53_pad_type_0"), val = tensor("valid")]; + tensor x_53_strides_0 = const()[name = tensor("x_53_strides_0"), val = tensor([1, 1])]; + tensor x_53_pad_0 = const()[name = tensor("x_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_53_dilations_0 = const()[name = tensor("x_53_dilations_0"), val = tensor([1, 1])]; + tensor x_53_groups_0 = const()[name = tensor("x_53_groups_0"), val = tensor(1)]; + tensor layers_8_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166675712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167462208))), name = tensor("layers_8_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_53_cast_fp16 = conv(dilations = x_53_dilations_0, groups = x_53_groups_0, pad = x_53_pad_0, pad_type = x_53_pad_type_0, strides = x_53_strides_0, weight = layers_8_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_243_cast_fp16)[name = tensor("x_53_cast_fp16")]; + tensor inputs_87_cast_fp16 = add(x = inputs_85_cast_fp16, y = x_53_cast_fp16)[name = tensor("inputs_87_cast_fp16")]; + tensor out_87_axes_0 = const()[name = tensor("out_87_axes_0"), val = tensor([1])]; + tensor var_2617_to_fp16 = const()[name = tensor("op_2617_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_2617_to_fp16, x = inputs_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor input_245_gamma_0_to_fp16 = const()[name = tensor("input_245_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167462400)))]; + tensor input_245_beta_0_to_fp16 = const()[name = tensor("input_245_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167464512)))]; + tensor input_245_epsilon_0_to_fp16 = const()[name = tensor("input_245_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = batch_norm(beta = input_245_beta_0_to_fp16, epsilon = input_245_epsilon_0_to_fp16, gamma = input_245_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor input_247_pad_type_0 = const()[name = tensor("input_247_pad_type_0"), val = tensor("valid")]; + tensor input_247_strides_0 = const()[name = tensor("input_247_strides_0"), val = tensor([1, 1])]; + tensor input_247_pad_0 = const()[name = tensor("input_247_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_247_dilations_0 = const()[name = tensor("input_247_dilations_0"), val = tensor([1, 1])]; + tensor input_247_groups_0 = const()[name = tensor("input_247_groups_0"), val = tensor(1)]; + tensor layers_8_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167466624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170612416))), name = tensor("layers_8_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_247_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_247_dilations_0, groups = input_247_groups_0, pad = input_247_pad_0, pad_type = input_247_pad_type_0, strides = input_247_strides_0, weight = layers_8_feed_forward2_fc1_weight_to_fp16_palettized, x = input_245_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor x_55_pad_type_0 = const()[name = tensor("x_55_pad_type_0"), val = tensor("valid")]; + tensor x_55_strides_0 = const()[name = tensor("x_55_strides_0"), val = tensor([1, 1])]; + tensor x_55_pad_0 = const()[name = tensor("x_55_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_55_dilations_0 = const()[name = tensor("x_55_dilations_0"), val = tensor([1, 1])]; + tensor x_55_groups_0 = const()[name = tensor("x_55_groups_0"), val = tensor(1)]; + tensor op_2645_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(170612608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173758400))), name = tensor("op_2645_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_2645_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_55_dilations_0, groups = x_55_groups_0, pad = x_55_pad_0, pad_type = x_55_pad_type_0, strides = x_55_strides_0, weight = op_2645_weight_0_to_fp16_palettized, x = input_249_cast_fp16)[name = tensor("op_2645_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_2645_cast_fp16)[name = tensor("inputs_89_cast_fp16")]; + tensor out_89_axes_0 = const()[name = tensor("out_89_axes_0"), val = tensor([1])]; + tensor var_2655_to_fp16 = const()[name = tensor("op_2655_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_2655_to_fp16, x = inputs_89_cast_fp16)[name = tensor("out_89_cast_fp16")]; + tensor inputs_91_gamma_0_to_fp16 = const()[name = tensor("inputs_91_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173758592)))]; + tensor inputs_91_beta_0_to_fp16 = const()[name = tensor("inputs_91_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173760704)))]; + tensor inputs_91_epsilon_0_to_fp16 = const()[name = tensor("inputs_91_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_91_cast_fp16 = batch_norm(beta = inputs_91_beta_0_to_fp16, epsilon = inputs_91_epsilon_0_to_fp16, gamma = inputs_91_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_2669 = const()[name = tensor("op_2669"), val = tensor(3)]; + tensor out_91_axes_0 = const()[name = tensor("out_91_axes_0"), val = tensor([1])]; + tensor var_2700_to_fp16 = const()[name = tensor("op_2700_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_2700_to_fp16, x = inputs_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor input_251_gamma_0_to_fp16 = const()[name = tensor("input_251_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173762816)))]; + tensor input_251_beta_0_to_fp16 = const()[name = tensor("input_251_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173764928)))]; + tensor input_251_epsilon_0_to_fp16 = const()[name = tensor("input_251_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = batch_norm(beta = input_251_beta_0_to_fp16, epsilon = input_251_epsilon_0_to_fp16, gamma = input_251_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor input_253_pad_type_0 = const()[name = tensor("input_253_pad_type_0"), val = tensor("valid")]; + tensor input_253_strides_0 = const()[name = tensor("input_253_strides_0"), val = tensor([1, 1])]; + tensor input_253_pad_0 = const()[name = tensor("input_253_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_253_dilations_0 = const()[name = tensor("input_253_dilations_0"), val = tensor([1, 1])]; + tensor input_253_groups_0 = const()[name = tensor("input_253_groups_0"), val = tensor(1)]; + tensor layers_9_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173767040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176912832))), name = tensor("layers_9_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_253_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_253_dilations_0, groups = input_253_groups_0, pad = input_253_pad_0, pad_type = input_253_pad_type_0, strides = input_253_strides_0, weight = layers_9_feed_forward1_fc1_weight_to_fp16_palettized, x = input_251_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_cast_fp16 = silu(x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor x_57_pad_type_0 = const()[name = tensor("x_57_pad_type_0"), val = tensor("valid")]; + tensor x_57_strides_0 = const()[name = tensor("x_57_strides_0"), val = tensor([1, 1])]; + tensor x_57_pad_0 = const()[name = tensor("x_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_57_dilations_0 = const()[name = tensor("x_57_dilations_0"), val = tensor([1, 1])]; + tensor x_57_groups_0 = const()[name = tensor("x_57_groups_0"), val = tensor(1)]; + tensor op_2728_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176913024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180058816))), name = tensor("op_2728_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_2728_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_57_dilations_0, groups = x_57_groups_0, pad = x_57_pad_0, pad_type = x_57_pad_type_0, strides = x_57_strides_0, weight = op_2728_weight_0_to_fp16_palettized, x = input_255_cast_fp16)[name = tensor("op_2728_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_2728_cast_fp16)[name = tensor("inputs_93_cast_fp16")]; + tensor out_93_axes_0 = const()[name = tensor("out_93_axes_0"), val = tensor([1])]; + tensor var_2738_to_fp16 = const()[name = tensor("op_2738_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_2738_to_fp16, x = inputs_93_cast_fp16)[name = tensor("out_93_cast_fp16")]; + tensor obj_39_gamma_0_to_fp16 = const()[name = tensor("obj_39_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180059008)))]; + tensor obj_39_beta_0_to_fp16 = const()[name = tensor("obj_39_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180061120)))]; + tensor obj_39_epsilon_0_to_fp16 = const()[name = tensor("obj_39_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_39_cast_fp16 = batch_norm(beta = obj_39_beta_0_to_fp16, epsilon = obj_39_epsilon_0_to_fp16, gamma = obj_39_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_93_cast_fp16)[name = tensor("obj_39_cast_fp16")]; + tensor query_37_pad_type_0 = const()[name = tensor("query_37_pad_type_0"), val = tensor("valid")]; + tensor query_37_strides_0 = const()[name = tensor("query_37_strides_0"), val = tensor([1, 1])]; + tensor query_37_pad_0 = const()[name = tensor("query_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_37_dilations_0 = const()[name = tensor("query_37_dilations_0"), val = tensor([1, 1])]; + tensor query_37_groups_0 = const()[name = tensor("query_37_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180063232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180849728))), name = tensor("layers_9_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_37_dilations_0, groups = query_37_groups_0, pad = query_37_pad_0, pad_type = query_37_pad_type_0, strides = query_37_strides_0, weight = layers_9_self_attn_q_proj_weight_to_fp16_palettized, x = obj_39_cast_fp16)[name = tensor("query_37_cast_fp16")]; + tensor key_19_pad_type_0 = const()[name = tensor("key_19_pad_type_0"), val = tensor("valid")]; + tensor key_19_strides_0 = const()[name = tensor("key_19_strides_0"), val = tensor([1, 1])]; + tensor key_19_pad_0 = const()[name = tensor("key_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_19_dilations_0 = const()[name = tensor("key_19_dilations_0"), val = tensor([1, 1])]; + tensor key_19_groups_0 = const()[name = tensor("key_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180849920))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181636416))), name = tensor("layers_9_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_19_cast_fp16 = conv(dilations = key_19_dilations_0, groups = key_19_groups_0, pad = key_19_pad_0, pad_type = key_19_pad_type_0, strides = key_19_strides_0, weight = layers_9_self_attn_k_proj_weight_to_fp16_palettized, x = obj_39_cast_fp16)[name = tensor("key_19_cast_fp16")]; + tensor value_19_pad_type_0 = const()[name = tensor("value_19_pad_type_0"), val = tensor("valid")]; + tensor value_19_strides_0 = const()[name = tensor("value_19_strides_0"), val = tensor([1, 1])]; + tensor value_19_pad_0 = const()[name = tensor("value_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_19_dilations_0 = const()[name = tensor("value_19_dilations_0"), val = tensor([1, 1])]; + tensor value_19_groups_0 = const()[name = tensor("value_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181636608))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182423104))), name = tensor("layers_9_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_19_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_19_dilations_0, groups = value_19_groups_0, pad = value_19_pad_0, pad_type = value_19_pad_type_0, strides = value_19_strides_0, weight = layers_9_self_attn_v_proj_weight_to_fp16_palettized, x = obj_39_cast_fp16)[name = tensor("value_19_cast_fp16")]; + tensor var_2776_to_fp16 = const()[name = tensor("op_2776_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182423296)))]; + tensor query_39_cast_fp16 = add(x = query_37_cast_fp16, y = var_2776_to_fp16)[name = tensor("query_39_cast_fp16")]; + tensor var_2779_to_fp16 = const()[name = tensor("op_2779_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182425408)))]; + tensor q_with_bias_v_19_cast_fp16 = add(x = query_37_cast_fp16, y = var_2779_to_fp16)[name = tensor("q_with_bias_v_19_cast_fp16")]; + tensor p_19_pad_type_0 = const()[name = tensor("p_19_pad_type_0"), val = tensor("valid")]; + tensor p_19_strides_0 = const()[name = tensor("p_19_strides_0"), val = tensor([1, 1])]; + tensor p_19_pad_0 = const()[name = tensor("p_19_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_19_dilations_0 = const()[name = tensor("p_19_dilations_0"), val = tensor([1, 1])]; + tensor p_19_groups_0 = const()[name = tensor("p_19_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182427520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183214016))), name = tensor("layers_9_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_19_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_19_dilations_0, groups = p_19_groups_0, pad = p_19_pad_0, pad_type = p_19_pad_type_0, strides = p_19_strides_0, weight = layers_9_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_19_cast_fp16")]; + tensor var_2790 = const()[name = tensor("op_2790"), val = tensor([1, 8, 128, 188])]; + tensor var_2791_cast_fp16 = reshape(shape = var_2790, x = q_with_bias_v_19_cast_fp16)[name = tensor("op_2791_cast_fp16")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([1, 8, 128, -1])]; + tensor var_2793_cast_fp16 = reshape(shape = var_2792, x = p_19_cast_fp16)[name = tensor("op_2793_cast_fp16")]; + tensor matrix_bd_73_transpose_x_0 = const()[name = tensor("matrix_bd_73_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_73_transpose_y_0 = const()[name = tensor("matrix_bd_73_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_73_cast_fp16 = matmul(transpose_x = matrix_bd_73_transpose_x_0, transpose_y = matrix_bd_73_transpose_y_0, x = var_2791_cast_fp16, y = var_2793_cast_fp16)[name = tensor("matrix_bd_73_cast_fp16")]; + tensor matrix_bd_75_pad_0 = const()[name = tensor("matrix_bd_75_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_75_mode_0 = const()[name = tensor("matrix_bd_75_mode_0"), val = tensor("constant")]; + tensor const_109_to_fp16 = const()[name = tensor("const_109_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_75_cast_fp16 = pad(constant_val = const_109_to_fp16, mode = matrix_bd_75_mode_0, pad = matrix_bd_75_pad_0, x = matrix_bd_73_cast_fp16)[name = tensor("matrix_bd_75_cast_fp16")]; + tensor var_2802 = const()[name = tensor("op_2802"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_2802, x = matrix_bd_75_cast_fp16)[name = tensor("matrix_bd_77_cast_fp16")]; + tensor var_2806_begin_0 = const()[name = tensor("op_2806_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2806_end_0 = const()[name = tensor("op_2806_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_2806_end_mask_0 = const()[name = tensor("op_2806_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2806_cast_fp16 = slice_by_index(begin = var_2806_begin_0, end = var_2806_end_0, end_mask = var_2806_end_mask_0, x = matrix_bd_77_cast_fp16)[name = tensor("op_2806_cast_fp16")]; + tensor var_2807 = const()[name = tensor("op_2807"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_79_cast_fp16 = reshape(shape = var_2807, x = var_2806_cast_fp16)[name = tensor("matrix_bd_79_cast_fp16")]; + tensor var_2812_begin_0 = const()[name = tensor("op_2812_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2812_end_0 = const()[name = tensor("op_2812_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_2812_end_mask_0 = const()[name = tensor("op_2812_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_2812_cast_fp16 = slice_by_index(begin = var_2812_begin_0, end = var_2812_end_0, end_mask = var_2812_end_mask_0, x = matrix_bd_79_cast_fp16)[name = tensor("op_2812_cast_fp16")]; + tensor var_2813_to_fp16 = const()[name = tensor("op_2813_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_19_cast_fp16 = mul(x = var_2812_cast_fp16, y = var_2813_to_fp16)[name = tensor("qk_mask_19_cast_fp16")]; + tensor var_2817 = const()[name = tensor("op_2817"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_2817, x = query_39_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_2819_to_fp16 = const()[name = tensor("op_2819_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_2820_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_2819_to_fp16)[name = tensor("op_2820_cast_fp16")]; + tensor var_2823 = const()[name = tensor("op_2823"), val = tensor([1, 8, 128, 188])]; + tensor var_2824_cast_fp16 = reshape(shape = var_2823, x = key_19_cast_fp16)[name = tensor("op_2824_cast_fp16")]; + tensor mh_w_37_transpose_x_0 = const()[name = tensor("mh_w_37_transpose_x_0"), val = tensor(true)]; + tensor mh_w_37_transpose_y_0 = const()[name = tensor("mh_w_37_transpose_y_0"), val = tensor(false)]; + tensor mh_w_37_cast_fp16 = matmul(transpose_x = mh_w_37_transpose_x_0, transpose_y = mh_w_37_transpose_y_0, x = var_2820_cast_fp16, y = var_2824_cast_fp16)[name = tensor("mh_w_37_cast_fp16")]; + tensor mh_w_39_cast_fp16 = add(x = mh_w_37_cast_fp16, y = qk_mask_19_cast_fp16)[name = tensor("mh_w_39_cast_fp16")]; + tensor var_2828_cast_fp16 = softmax(axis = var_2669, x = mh_w_39_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829 = const()[name = tensor("op_2829"), val = tensor([1, 8, 128, 188])]; + tensor var_2830_cast_fp16 = reshape(shape = var_2829, x = value_19_cast_fp16)[name = tensor("op_2830_cast_fp16")]; + tensor attn_19_transpose_x_0 = const()[name = tensor("attn_19_transpose_x_0"), val = tensor(false)]; + tensor attn_19_transpose_y_0 = const()[name = tensor("attn_19_transpose_y_0"), val = tensor(true)]; + tensor attn_19_cast_fp16 = matmul(transpose_x = attn_19_transpose_x_0, transpose_y = attn_19_transpose_y_0, x = var_2830_cast_fp16, y = var_2828_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2833 = const()[name = tensor("op_2833"), val = tensor([1, 1024, 1, 188])]; + tensor input_257_cast_fp16 = reshape(shape = var_2833, x = attn_19_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor obj_41_pad_type_0 = const()[name = tensor("obj_41_pad_type_0"), val = tensor("valid")]; + tensor obj_41_strides_0 = const()[name = tensor("obj_41_strides_0"), val = tensor([1, 1])]; + tensor obj_41_pad_0 = const()[name = tensor("obj_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_41_dilations_0 = const()[name = tensor("obj_41_dilations_0"), val = tensor([1, 1])]; + tensor obj_41_groups_0 = const()[name = tensor("obj_41_groups_0"), val = tensor(1)]; + tensor layers_9_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183214208))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184000704))), name = tensor("layers_9_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_41_dilations_0, groups = obj_41_groups_0, pad = obj_41_pad_0, pad_type = obj_41_pad_type_0, strides = obj_41_strides_0, weight = layers_9_self_attn_o_proj_weight_to_fp16_palettized, x = input_257_cast_fp16)[name = tensor("obj_41_cast_fp16")]; + tensor inputs_95_cast_fp16 = add(x = inputs_93_cast_fp16, y = obj_41_cast_fp16)[name = tensor("inputs_95_cast_fp16")]; + tensor out_95_axes_0 = const()[name = tensor("out_95_axes_0"), val = tensor([1])]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_2851_to_fp16, x = inputs_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_259_gamma_0_to_fp16 = const()[name = tensor("input_259_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184000896)))]; + tensor input_259_beta_0_to_fp16 = const()[name = tensor("input_259_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184003008)))]; + tensor input_259_epsilon_0_to_fp16 = const()[name = tensor("input_259_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_259_cast_fp16 = batch_norm(beta = input_259_beta_0_to_fp16, epsilon = input_259_epsilon_0_to_fp16, gamma = input_259_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_pad_type_0 = const()[name = tensor("input_261_pad_type_0"), val = tensor("valid")]; + tensor input_261_strides_0 = const()[name = tensor("input_261_strides_0"), val = tensor([1, 1])]; + tensor input_261_pad_0 = const()[name = tensor("input_261_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_261_dilations_0 = const()[name = tensor("input_261_dilations_0"), val = tensor([1, 1])]; + tensor input_261_groups_0 = const()[name = tensor("input_261_groups_0"), val = tensor(1)]; + tensor layers_9_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184005120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185578048))), name = tensor("layers_9_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_261_cast_fp16 = conv(dilations = input_261_dilations_0, groups = input_261_groups_0, pad = input_261_pad_0, pad_type = input_261_pad_type_0, strides = input_261_strides_0, weight = layers_9_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_split_num_splits_0 = const()[name = tensor("input_263_split_num_splits_0"), val = tensor(2)]; + tensor input_263_split_axis_0 = const()[name = tensor("input_263_split_axis_0"), val = tensor(1)]; + tensor input_263_split_cast_fp16_0, tensor input_263_split_cast_fp16_1 = split(axis = input_263_split_axis_0, num_splits = input_263_split_num_splits_0, x = input_261_cast_fp16)[name = tensor("input_263_split_cast_fp16")]; + tensor input_263_split_1_sigmoid_cast_fp16 = sigmoid(x = input_263_split_cast_fp16_1)[name = tensor("input_263_split_1_sigmoid_cast_fp16")]; + tensor input_263_cast_fp16 = mul(x = input_263_split_cast_fp16_0, y = input_263_split_1_sigmoid_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor input_265_pad_type_0 = const()[name = tensor("input_265_pad_type_0"), val = tensor("custom")]; + tensor input_265_pad_0 = const()[name = tensor("input_265_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_265_groups_0 = const()[name = tensor("input_265_groups_0"), val = tensor(1024)]; + tensor input_265_strides_0 = const()[name = tensor("input_265_strides_0"), val = tensor([1, 1])]; + tensor input_265_dilations_0 = const()[name = tensor("input_265_dilations_0"), val = tensor([1, 1])]; + tensor const_286_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185578240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185585216))), name = tensor("const_286_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_287_to_fp16 = const()[name = tensor("const_287_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185585408)))]; + tensor input_267_cast_fp16 = conv(bias = const_287_to_fp16, dilations = input_265_dilations_0, groups = input_265_groups_0, pad = input_265_pad_0, pad_type = input_265_pad_type_0, strides = input_265_strides_0, weight = const_286_to_fp16_palettized, x = input_263_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_cast_fp16 = silu(x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor x_59_pad_type_0 = const()[name = tensor("x_59_pad_type_0"), val = tensor("valid")]; + tensor x_59_strides_0 = const()[name = tensor("x_59_strides_0"), val = tensor([1, 1])]; + tensor x_59_pad_0 = const()[name = tensor("x_59_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_59_dilations_0 = const()[name = tensor("x_59_dilations_0"), val = tensor([1, 1])]; + tensor x_59_groups_0 = const()[name = tensor("x_59_groups_0"), val = tensor(1)]; + tensor layers_9_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185587520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186374016))), name = tensor("layers_9_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_59_cast_fp16 = conv(dilations = x_59_dilations_0, groups = x_59_groups_0, pad = x_59_pad_0, pad_type = x_59_pad_type_0, strides = x_59_strides_0, weight = layers_9_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_269_cast_fp16)[name = tensor("x_59_cast_fp16")]; + tensor inputs_97_cast_fp16 = add(x = inputs_95_cast_fp16, y = x_59_cast_fp16)[name = tensor("inputs_97_cast_fp16")]; + tensor out_97_axes_0 = const()[name = tensor("out_97_axes_0"), val = tensor([1])]; + tensor var_2899_to_fp16 = const()[name = tensor("op_2899_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_2899_to_fp16, x = inputs_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor input_271_gamma_0_to_fp16 = const()[name = tensor("input_271_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186374208)))]; + tensor input_271_beta_0_to_fp16 = const()[name = tensor("input_271_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186376320)))]; + tensor input_271_epsilon_0_to_fp16 = const()[name = tensor("input_271_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_271_cast_fp16 = batch_norm(beta = input_271_beta_0_to_fp16, epsilon = input_271_epsilon_0_to_fp16, gamma = input_271_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_pad_type_0 = const()[name = tensor("input_273_pad_type_0"), val = tensor("valid")]; + tensor input_273_strides_0 = const()[name = tensor("input_273_strides_0"), val = tensor([1, 1])]; + tensor input_273_pad_0 = const()[name = tensor("input_273_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_273_dilations_0 = const()[name = tensor("input_273_dilations_0"), val = tensor([1, 1])]; + tensor input_273_groups_0 = const()[name = tensor("input_273_groups_0"), val = tensor(1)]; + tensor layers_9_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(186378432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189524224))), name = tensor("layers_9_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_273_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_273_dilations_0, groups = input_273_groups_0, pad = input_273_pad_0, pad_type = input_273_pad_type_0, strides = input_273_strides_0, weight = layers_9_feed_forward2_fc1_weight_to_fp16_palettized, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_cast_fp16 = silu(x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor x_61_pad_type_0 = const()[name = tensor("x_61_pad_type_0"), val = tensor("valid")]; + tensor x_61_strides_0 = const()[name = tensor("x_61_strides_0"), val = tensor([1, 1])]; + tensor x_61_pad_0 = const()[name = tensor("x_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_61_dilations_0 = const()[name = tensor("x_61_dilations_0"), val = tensor([1, 1])]; + tensor x_61_groups_0 = const()[name = tensor("x_61_groups_0"), val = tensor(1)]; + tensor op_2927_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189524416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192670208))), name = tensor("op_2927_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_2927_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_61_dilations_0, groups = x_61_groups_0, pad = x_61_pad_0, pad_type = x_61_pad_type_0, strides = x_61_strides_0, weight = op_2927_weight_0_to_fp16_palettized, x = input_275_cast_fp16)[name = tensor("op_2927_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = var_2927_cast_fp16)[name = tensor("inputs_99_cast_fp16")]; + tensor out_99_axes_0 = const()[name = tensor("out_99_axes_0"), val = tensor([1])]; + tensor var_2937_to_fp16 = const()[name = tensor("op_2937_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_2937_to_fp16, x = inputs_99_cast_fp16)[name = tensor("out_99_cast_fp16")]; + tensor inputs_101_gamma_0_to_fp16 = const()[name = tensor("inputs_101_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192670400)))]; + tensor inputs_101_beta_0_to_fp16 = const()[name = tensor("inputs_101_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192672512)))]; + tensor inputs_101_epsilon_0_to_fp16 = const()[name = tensor("inputs_101_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_101_cast_fp16 = batch_norm(beta = inputs_101_beta_0_to_fp16, epsilon = inputs_101_epsilon_0_to_fp16, gamma = inputs_101_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_2951 = const()[name = tensor("op_2951"), val = tensor(3)]; + tensor out_101_axes_0 = const()[name = tensor("out_101_axes_0"), val = tensor([1])]; + tensor var_2982_to_fp16 = const()[name = tensor("op_2982_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_2982_to_fp16, x = inputs_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_277_gamma_0_to_fp16 = const()[name = tensor("input_277_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192674624)))]; + tensor input_277_beta_0_to_fp16 = const()[name = tensor("input_277_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192676736)))]; + tensor input_277_epsilon_0_to_fp16 = const()[name = tensor("input_277_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_277_cast_fp16 = batch_norm(beta = input_277_beta_0_to_fp16, epsilon = input_277_epsilon_0_to_fp16, gamma = input_277_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_pad_type_0 = const()[name = tensor("input_279_pad_type_0"), val = tensor("valid")]; + tensor input_279_strides_0 = const()[name = tensor("input_279_strides_0"), val = tensor([1, 1])]; + tensor input_279_pad_0 = const()[name = tensor("input_279_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_279_dilations_0 = const()[name = tensor("input_279_dilations_0"), val = tensor([1, 1])]; + tensor input_279_groups_0 = const()[name = tensor("input_279_groups_0"), val = tensor(1)]; + tensor layers_10_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192678848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195824640))), name = tensor("layers_10_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_279_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_279_dilations_0, groups = input_279_groups_0, pad = input_279_pad_0, pad_type = input_279_pad_type_0, strides = input_279_strides_0, weight = layers_10_feed_forward1_fc1_weight_to_fp16_palettized, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_cast_fp16 = silu(x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor x_63_pad_type_0 = const()[name = tensor("x_63_pad_type_0"), val = tensor("valid")]; + tensor x_63_strides_0 = const()[name = tensor("x_63_strides_0"), val = tensor([1, 1])]; + tensor x_63_pad_0 = const()[name = tensor("x_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_63_dilations_0 = const()[name = tensor("x_63_dilations_0"), val = tensor([1, 1])]; + tensor x_63_groups_0 = const()[name = tensor("x_63_groups_0"), val = tensor(1)]; + tensor op_3010_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(195824832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198970624))), name = tensor("op_3010_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_3010_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_63_dilations_0, groups = x_63_groups_0, pad = x_63_pad_0, pad_type = x_63_pad_type_0, strides = x_63_strides_0, weight = op_3010_weight_0_to_fp16_palettized, x = input_281_cast_fp16)[name = tensor("op_3010_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = var_3010_cast_fp16)[name = tensor("inputs_103_cast_fp16")]; + tensor out_103_axes_0 = const()[name = tensor("out_103_axes_0"), val = tensor([1])]; + tensor var_3020_to_fp16 = const()[name = tensor("op_3020_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3020_to_fp16, x = inputs_103_cast_fp16)[name = tensor("out_103_cast_fp16")]; + tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198970816)))]; + tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198972928)))]; + tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_103_cast_fp16)[name = tensor("obj_43_cast_fp16")]; + tensor query_41_pad_type_0 = const()[name = tensor("query_41_pad_type_0"), val = tensor("valid")]; + tensor query_41_strides_0 = const()[name = tensor("query_41_strides_0"), val = tensor([1, 1])]; + tensor query_41_pad_0 = const()[name = tensor("query_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_41_dilations_0 = const()[name = tensor("query_41_dilations_0"), val = tensor([1, 1])]; + tensor query_41_groups_0 = const()[name = tensor("query_41_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(198975040))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199761536))), name = tensor("layers_10_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_41_dilations_0, groups = query_41_groups_0, pad = query_41_pad_0, pad_type = query_41_pad_type_0, strides = query_41_strides_0, weight = layers_10_self_attn_q_proj_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("query_41_cast_fp16")]; + tensor key_21_pad_type_0 = const()[name = tensor("key_21_pad_type_0"), val = tensor("valid")]; + tensor key_21_strides_0 = const()[name = tensor("key_21_strides_0"), val = tensor([1, 1])]; + tensor key_21_pad_0 = const()[name = tensor("key_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_21_dilations_0 = const()[name = tensor("key_21_dilations_0"), val = tensor([1, 1])]; + tensor key_21_groups_0 = const()[name = tensor("key_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(199761728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200548224))), name = tensor("layers_10_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_21_cast_fp16 = conv(dilations = key_21_dilations_0, groups = key_21_groups_0, pad = key_21_pad_0, pad_type = key_21_pad_type_0, strides = key_21_strides_0, weight = layers_10_self_attn_k_proj_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("key_21_cast_fp16")]; + tensor value_21_pad_type_0 = const()[name = tensor("value_21_pad_type_0"), val = tensor("valid")]; + tensor value_21_strides_0 = const()[name = tensor("value_21_strides_0"), val = tensor([1, 1])]; + tensor value_21_pad_0 = const()[name = tensor("value_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_21_dilations_0 = const()[name = tensor("value_21_dilations_0"), val = tensor([1, 1])]; + tensor value_21_groups_0 = const()[name = tensor("value_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(200548416))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201334912))), name = tensor("layers_10_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_21_dilations_0, groups = value_21_groups_0, pad = value_21_pad_0, pad_type = value_21_pad_type_0, strides = value_21_strides_0, weight = layers_10_self_attn_v_proj_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_3058_to_fp16 = const()[name = tensor("op_3058_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201335104)))]; + tensor query_43_cast_fp16 = add(x = query_41_cast_fp16, y = var_3058_to_fp16)[name = tensor("query_43_cast_fp16")]; + tensor var_3061_to_fp16 = const()[name = tensor("op_3061_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201337216)))]; + tensor q_with_bias_v_21_cast_fp16 = add(x = query_41_cast_fp16, y = var_3061_to_fp16)[name = tensor("q_with_bias_v_21_cast_fp16")]; + tensor p_21_pad_type_0 = const()[name = tensor("p_21_pad_type_0"), val = tensor("valid")]; + tensor p_21_strides_0 = const()[name = tensor("p_21_strides_0"), val = tensor([1, 1])]; + tensor p_21_pad_0 = const()[name = tensor("p_21_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_21_dilations_0 = const()[name = tensor("p_21_dilations_0"), val = tensor([1, 1])]; + tensor p_21_groups_0 = const()[name = tensor("p_21_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201339328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202125824))), name = tensor("layers_10_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_21_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_21_dilations_0, groups = p_21_groups_0, pad = p_21_pad_0, pad_type = p_21_pad_type_0, strides = p_21_strides_0, weight = layers_10_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_21_cast_fp16")]; + tensor var_3072 = const()[name = tensor("op_3072"), val = tensor([1, 8, 128, 188])]; + tensor var_3073_cast_fp16 = reshape(shape = var_3072, x = q_with_bias_v_21_cast_fp16)[name = tensor("op_3073_cast_fp16")]; + tensor var_3074 = const()[name = tensor("op_3074"), val = tensor([1, 8, 128, -1])]; + tensor var_3075_cast_fp16 = reshape(shape = var_3074, x = p_21_cast_fp16)[name = tensor("op_3075_cast_fp16")]; + tensor matrix_bd_81_transpose_x_0 = const()[name = tensor("matrix_bd_81_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_81_transpose_y_0 = const()[name = tensor("matrix_bd_81_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_81_cast_fp16 = matmul(transpose_x = matrix_bd_81_transpose_x_0, transpose_y = matrix_bd_81_transpose_y_0, x = var_3073_cast_fp16, y = var_3075_cast_fp16)[name = tensor("matrix_bd_81_cast_fp16")]; + tensor matrix_bd_83_pad_0 = const()[name = tensor("matrix_bd_83_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_83_mode_0 = const()[name = tensor("matrix_bd_83_mode_0"), val = tensor("constant")]; + tensor const_120_to_fp16 = const()[name = tensor("const_120_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_83_cast_fp16 = pad(constant_val = const_120_to_fp16, mode = matrix_bd_83_mode_0, pad = matrix_bd_83_pad_0, x = matrix_bd_81_cast_fp16)[name = tensor("matrix_bd_83_cast_fp16")]; + tensor var_3084 = const()[name = tensor("op_3084"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_3084, x = matrix_bd_83_cast_fp16)[name = tensor("matrix_bd_85_cast_fp16")]; + tensor var_3088_begin_0 = const()[name = tensor("op_3088_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3088_end_0 = const()[name = tensor("op_3088_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3088_end_mask_0 = const()[name = tensor("op_3088_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3088_cast_fp16 = slice_by_index(begin = var_3088_begin_0, end = var_3088_end_0, end_mask = var_3088_end_mask_0, x = matrix_bd_85_cast_fp16)[name = tensor("op_3088_cast_fp16")]; + tensor var_3089 = const()[name = tensor("op_3089"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_87_cast_fp16 = reshape(shape = var_3089, x = var_3088_cast_fp16)[name = tensor("matrix_bd_87_cast_fp16")]; + tensor var_3094_begin_0 = const()[name = tensor("op_3094_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3094_end_0 = const()[name = tensor("op_3094_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3094_end_mask_0 = const()[name = tensor("op_3094_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3094_cast_fp16 = slice_by_index(begin = var_3094_begin_0, end = var_3094_end_0, end_mask = var_3094_end_mask_0, x = matrix_bd_87_cast_fp16)[name = tensor("op_3094_cast_fp16")]; + tensor var_3095_to_fp16 = const()[name = tensor("op_3095_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_21_cast_fp16 = mul(x = var_3094_cast_fp16, y = var_3095_to_fp16)[name = tensor("qk_mask_21_cast_fp16")]; + tensor var_3099 = const()[name = tensor("op_3099"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_3099, x = query_43_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_3101_to_fp16 = const()[name = tensor("op_3101_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3102_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_3101_to_fp16)[name = tensor("op_3102_cast_fp16")]; + tensor var_3105 = const()[name = tensor("op_3105"), val = tensor([1, 8, 128, 188])]; + tensor var_3106_cast_fp16 = reshape(shape = var_3105, x = key_21_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mh_w_41_transpose_x_0 = const()[name = tensor("mh_w_41_transpose_x_0"), val = tensor(true)]; + tensor mh_w_41_transpose_y_0 = const()[name = tensor("mh_w_41_transpose_y_0"), val = tensor(false)]; + tensor mh_w_41_cast_fp16 = matmul(transpose_x = mh_w_41_transpose_x_0, transpose_y = mh_w_41_transpose_y_0, x = var_3102_cast_fp16, y = var_3106_cast_fp16)[name = tensor("mh_w_41_cast_fp16")]; + tensor mh_w_43_cast_fp16 = add(x = mh_w_41_cast_fp16, y = qk_mask_21_cast_fp16)[name = tensor("mh_w_43_cast_fp16")]; + tensor var_3110_cast_fp16 = softmax(axis = var_2951, x = mh_w_43_cast_fp16)[name = tensor("op_3110_cast_fp16")]; + tensor var_3111 = const()[name = tensor("op_3111"), val = tensor([1, 8, 128, 188])]; + tensor var_3112_cast_fp16 = reshape(shape = var_3111, x = value_21_cast_fp16)[name = tensor("op_3112_cast_fp16")]; + tensor attn_21_transpose_x_0 = const()[name = tensor("attn_21_transpose_x_0"), val = tensor(false)]; + tensor attn_21_transpose_y_0 = const()[name = tensor("attn_21_transpose_y_0"), val = tensor(true)]; + tensor attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = var_3112_cast_fp16, y = var_3110_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_3115 = const()[name = tensor("op_3115"), val = tensor([1, 1024, 1, 188])]; + tensor input_283_cast_fp16 = reshape(shape = var_3115, x = attn_21_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor obj_45_pad_type_0 = const()[name = tensor("obj_45_pad_type_0"), val = tensor("valid")]; + tensor obj_45_strides_0 = const()[name = tensor("obj_45_strides_0"), val = tensor([1, 1])]; + tensor obj_45_pad_0 = const()[name = tensor("obj_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_45_dilations_0 = const()[name = tensor("obj_45_dilations_0"), val = tensor([1, 1])]; + tensor obj_45_groups_0 = const()[name = tensor("obj_45_groups_0"), val = tensor(1)]; + tensor layers_10_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202126016))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202912512))), name = tensor("layers_10_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_45_dilations_0, groups = obj_45_groups_0, pad = obj_45_pad_0, pad_type = obj_45_pad_type_0, strides = obj_45_strides_0, weight = layers_10_self_attn_o_proj_weight_to_fp16_palettized, x = input_283_cast_fp16)[name = tensor("obj_45_cast_fp16")]; + tensor inputs_105_cast_fp16 = add(x = inputs_103_cast_fp16, y = obj_45_cast_fp16)[name = tensor("inputs_105_cast_fp16")]; + tensor out_105_axes_0 = const()[name = tensor("out_105_axes_0"), val = tensor([1])]; + tensor var_3133_to_fp16 = const()[name = tensor("op_3133_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3133_to_fp16, x = inputs_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor input_285_gamma_0_to_fp16 = const()[name = tensor("input_285_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202912704)))]; + tensor input_285_beta_0_to_fp16 = const()[name = tensor("input_285_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202914816)))]; + tensor input_285_epsilon_0_to_fp16 = const()[name = tensor("input_285_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_285_cast_fp16 = batch_norm(beta = input_285_beta_0_to_fp16, epsilon = input_285_epsilon_0_to_fp16, gamma = input_285_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_pad_type_0 = const()[name = tensor("input_287_pad_type_0"), val = tensor("valid")]; + tensor input_287_strides_0 = const()[name = tensor("input_287_strides_0"), val = tensor([1, 1])]; + tensor input_287_pad_0 = const()[name = tensor("input_287_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_287_dilations_0 = const()[name = tensor("input_287_dilations_0"), val = tensor([1, 1])]; + tensor input_287_groups_0 = const()[name = tensor("input_287_groups_0"), val = tensor(1)]; + tensor layers_10_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202916928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204489856))), name = tensor("layers_10_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_287_cast_fp16 = conv(dilations = input_287_dilations_0, groups = input_287_groups_0, pad = input_287_pad_0, pad_type = input_287_pad_type_0, strides = input_287_strides_0, weight = layers_10_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor input_289_split_num_splits_0 = const()[name = tensor("input_289_split_num_splits_0"), val = tensor(2)]; + tensor input_289_split_axis_0 = const()[name = tensor("input_289_split_axis_0"), val = tensor(1)]; + tensor input_289_split_cast_fp16_0, tensor input_289_split_cast_fp16_1 = split(axis = input_289_split_axis_0, num_splits = input_289_split_num_splits_0, x = input_287_cast_fp16)[name = tensor("input_289_split_cast_fp16")]; + tensor input_289_split_1_sigmoid_cast_fp16 = sigmoid(x = input_289_split_cast_fp16_1)[name = tensor("input_289_split_1_sigmoid_cast_fp16")]; + tensor input_289_cast_fp16 = mul(x = input_289_split_cast_fp16_0, y = input_289_split_1_sigmoid_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor input_291_pad_type_0 = const()[name = tensor("input_291_pad_type_0"), val = tensor("custom")]; + tensor input_291_pad_0 = const()[name = tensor("input_291_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_291_groups_0 = const()[name = tensor("input_291_groups_0"), val = tensor(1024)]; + tensor input_291_strides_0 = const()[name = tensor("input_291_strides_0"), val = tensor([1, 1])]; + tensor input_291_dilations_0 = const()[name = tensor("input_291_dilations_0"), val = tensor([1, 1])]; + tensor const_288_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204490048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204497024))), name = tensor("const_288_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_289_to_fp16 = const()[name = tensor("const_289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204497216)))]; + tensor input_293_cast_fp16 = conv(bias = const_289_to_fp16, dilations = input_291_dilations_0, groups = input_291_groups_0, pad = input_291_pad_0, pad_type = input_291_pad_type_0, strides = input_291_strides_0, weight = const_288_to_fp16_palettized, x = input_289_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor input_295_cast_fp16 = silu(x = input_293_cast_fp16)[name = tensor("input_295_cast_fp16")]; + tensor x_65_pad_type_0 = const()[name = tensor("x_65_pad_type_0"), val = tensor("valid")]; + tensor x_65_strides_0 = const()[name = tensor("x_65_strides_0"), val = tensor([1, 1])]; + tensor x_65_pad_0 = const()[name = tensor("x_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_65_dilations_0 = const()[name = tensor("x_65_dilations_0"), val = tensor([1, 1])]; + tensor x_65_groups_0 = const()[name = tensor("x_65_groups_0"), val = tensor(1)]; + tensor layers_10_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204499328))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205285824))), name = tensor("layers_10_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor 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 = layers_10_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_295_cast_fp16)[name = tensor("x_65_cast_fp16")]; + tensor inputs_107_cast_fp16 = add(x = inputs_105_cast_fp16, y = x_65_cast_fp16)[name = tensor("inputs_107_cast_fp16")]; + tensor out_107_axes_0 = const()[name = tensor("out_107_axes_0"), val = tensor([1])]; + tensor var_3181_to_fp16 = const()[name = tensor("op_3181_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3181_to_fp16, x = inputs_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_297_gamma_0_to_fp16 = const()[name = tensor("input_297_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205286016)))]; + tensor input_297_beta_0_to_fp16 = const()[name = tensor("input_297_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205288128)))]; + tensor input_297_epsilon_0_to_fp16 = const()[name = tensor("input_297_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_297_cast_fp16 = batch_norm(beta = input_297_beta_0_to_fp16, epsilon = input_297_epsilon_0_to_fp16, gamma = input_297_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor input_299_pad_type_0 = const()[name = tensor("input_299_pad_type_0"), val = tensor("valid")]; + tensor input_299_strides_0 = const()[name = tensor("input_299_strides_0"), val = tensor([1, 1])]; + tensor input_299_pad_0 = const()[name = tensor("input_299_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_299_dilations_0 = const()[name = tensor("input_299_dilations_0"), val = tensor([1, 1])]; + tensor input_299_groups_0 = const()[name = tensor("input_299_groups_0"), val = tensor(1)]; + tensor layers_10_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205290240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208436032))), name = tensor("layers_10_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_299_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_299_dilations_0, groups = input_299_groups_0, pad = input_299_pad_0, pad_type = input_299_pad_type_0, strides = input_299_strides_0, weight = layers_10_feed_forward2_fc1_weight_to_fp16_palettized, x = input_297_cast_fp16)[name = tensor("input_299_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = input_299_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor x_67_pad_type_0 = const()[name = tensor("x_67_pad_type_0"), val = tensor("valid")]; + tensor x_67_strides_0 = const()[name = tensor("x_67_strides_0"), val = tensor([1, 1])]; + tensor x_67_pad_0 = const()[name = tensor("x_67_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_67_dilations_0 = const()[name = tensor("x_67_dilations_0"), val = tensor([1, 1])]; + tensor x_67_groups_0 = const()[name = tensor("x_67_groups_0"), val = tensor(1)]; + tensor op_3209_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208436224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211582016))), name = tensor("op_3209_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_3209_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_67_dilations_0, groups = x_67_groups_0, pad = x_67_pad_0, pad_type = x_67_pad_type_0, strides = x_67_strides_0, weight = op_3209_weight_0_to_fp16_palettized, x = input_301_cast_fp16)[name = tensor("op_3209_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = var_3209_cast_fp16)[name = tensor("inputs_109_cast_fp16")]; + tensor out_109_axes_0 = const()[name = tensor("out_109_axes_0"), val = tensor([1])]; + tensor var_3219_to_fp16 = const()[name = tensor("op_3219_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3219_to_fp16, x = inputs_109_cast_fp16)[name = tensor("out_109_cast_fp16")]; + tensor inputs_111_gamma_0_to_fp16 = const()[name = tensor("inputs_111_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211582208)))]; + tensor inputs_111_beta_0_to_fp16 = const()[name = tensor("inputs_111_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211584320)))]; + tensor inputs_111_epsilon_0_to_fp16 = const()[name = tensor("inputs_111_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_111_cast_fp16 = batch_norm(beta = inputs_111_beta_0_to_fp16, epsilon = inputs_111_epsilon_0_to_fp16, gamma = inputs_111_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor var_3233 = const()[name = tensor("op_3233"), val = tensor(3)]; + tensor out_111_axes_0 = const()[name = tensor("out_111_axes_0"), val = tensor([1])]; + tensor var_3264_to_fp16 = const()[name = tensor("op_3264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_3264_to_fp16, x = inputs_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor input_303_gamma_0_to_fp16 = const()[name = tensor("input_303_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211586432)))]; + tensor input_303_beta_0_to_fp16 = const()[name = tensor("input_303_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211588544)))]; + tensor input_303_epsilon_0_to_fp16 = const()[name = tensor("input_303_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_303_cast_fp16 = batch_norm(beta = input_303_beta_0_to_fp16, epsilon = input_303_epsilon_0_to_fp16, gamma = input_303_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor input_305_pad_type_0 = const()[name = tensor("input_305_pad_type_0"), val = tensor("valid")]; + tensor input_305_strides_0 = const()[name = tensor("input_305_strides_0"), val = tensor([1, 1])]; + tensor input_305_pad_0 = const()[name = tensor("input_305_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_305_dilations_0 = const()[name = tensor("input_305_dilations_0"), val = tensor([1, 1])]; + tensor input_305_groups_0 = const()[name = tensor("input_305_groups_0"), val = tensor(1)]; + tensor layers_11_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(211590656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214736448))), name = tensor("layers_11_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_305_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_305_dilations_0, groups = input_305_groups_0, pad = input_305_pad_0, pad_type = input_305_pad_type_0, strides = input_305_strides_0, weight = layers_11_feed_forward1_fc1_weight_to_fp16_palettized, x = input_303_cast_fp16)[name = tensor("input_305_cast_fp16")]; + tensor input_307_cast_fp16 = silu(x = input_305_cast_fp16)[name = tensor("input_307_cast_fp16")]; + tensor x_69_pad_type_0 = const()[name = tensor("x_69_pad_type_0"), val = tensor("valid")]; + tensor x_69_strides_0 = const()[name = tensor("x_69_strides_0"), val = tensor([1, 1])]; + tensor x_69_pad_0 = const()[name = tensor("x_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_69_dilations_0 = const()[name = tensor("x_69_dilations_0"), val = tensor([1, 1])]; + tensor x_69_groups_0 = const()[name = tensor("x_69_groups_0"), val = tensor(1)]; + tensor op_3292_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214736640))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217882432))), name = tensor("op_3292_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_3292_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_69_dilations_0, groups = x_69_groups_0, pad = x_69_pad_0, pad_type = x_69_pad_type_0, strides = x_69_strides_0, weight = op_3292_weight_0_to_fp16_palettized, x = input_307_cast_fp16)[name = tensor("op_3292_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = var_3292_cast_fp16)[name = tensor("inputs_113_cast_fp16")]; + tensor out_113_axes_0 = const()[name = tensor("out_113_axes_0"), val = tensor([1])]; + tensor var_3302_to_fp16 = const()[name = tensor("op_3302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_3302_to_fp16, x = inputs_113_cast_fp16)[name = tensor("out_113_cast_fp16")]; + tensor obj_47_gamma_0_to_fp16 = const()[name = tensor("obj_47_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217882624)))]; + tensor obj_47_beta_0_to_fp16 = const()[name = tensor("obj_47_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217884736)))]; + tensor obj_47_epsilon_0_to_fp16 = const()[name = tensor("obj_47_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_47_cast_fp16 = batch_norm(beta = obj_47_beta_0_to_fp16, epsilon = obj_47_epsilon_0_to_fp16, gamma = obj_47_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_113_cast_fp16)[name = tensor("obj_47_cast_fp16")]; + tensor query_45_pad_type_0 = const()[name = tensor("query_45_pad_type_0"), val = tensor("valid")]; + tensor query_45_strides_0 = const()[name = tensor("query_45_strides_0"), val = tensor([1, 1])]; + tensor query_45_pad_0 = const()[name = tensor("query_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_45_dilations_0 = const()[name = tensor("query_45_dilations_0"), val = tensor([1, 1])]; + tensor query_45_groups_0 = const()[name = tensor("query_45_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217886848))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218673344))), name = tensor("layers_11_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_45_dilations_0, groups = query_45_groups_0, pad = query_45_pad_0, pad_type = query_45_pad_type_0, strides = query_45_strides_0, weight = layers_11_self_attn_q_proj_weight_to_fp16_palettized, x = obj_47_cast_fp16)[name = tensor("query_45_cast_fp16")]; + tensor key_23_pad_type_0 = const()[name = tensor("key_23_pad_type_0"), val = tensor("valid")]; + tensor key_23_strides_0 = const()[name = tensor("key_23_strides_0"), val = tensor([1, 1])]; + tensor key_23_pad_0 = const()[name = tensor("key_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_23_dilations_0 = const()[name = tensor("key_23_dilations_0"), val = tensor([1, 1])]; + tensor key_23_groups_0 = const()[name = tensor("key_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218673536))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219460032))), name = tensor("layers_11_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_23_cast_fp16 = conv(dilations = key_23_dilations_0, groups = key_23_groups_0, pad = key_23_pad_0, pad_type = key_23_pad_type_0, strides = key_23_strides_0, weight = layers_11_self_attn_k_proj_weight_to_fp16_palettized, x = obj_47_cast_fp16)[name = tensor("key_23_cast_fp16")]; + tensor value_23_pad_type_0 = const()[name = tensor("value_23_pad_type_0"), val = tensor("valid")]; + tensor value_23_strides_0 = const()[name = tensor("value_23_strides_0"), val = tensor([1, 1])]; + tensor value_23_pad_0 = const()[name = tensor("value_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_23_dilations_0 = const()[name = tensor("value_23_dilations_0"), val = tensor([1, 1])]; + tensor value_23_groups_0 = const()[name = tensor("value_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219460224))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220246720))), name = tensor("layers_11_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_23_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_23_dilations_0, groups = value_23_groups_0, pad = value_23_pad_0, pad_type = value_23_pad_type_0, strides = value_23_strides_0, weight = layers_11_self_attn_v_proj_weight_to_fp16_palettized, x = obj_47_cast_fp16)[name = tensor("value_23_cast_fp16")]; + tensor var_3340_to_fp16 = const()[name = tensor("op_3340_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220246912)))]; + tensor query_47_cast_fp16 = add(x = query_45_cast_fp16, y = var_3340_to_fp16)[name = tensor("query_47_cast_fp16")]; + tensor var_3343_to_fp16 = const()[name = tensor("op_3343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220249024)))]; + tensor q_with_bias_v_23_cast_fp16 = add(x = query_45_cast_fp16, y = var_3343_to_fp16)[name = tensor("q_with_bias_v_23_cast_fp16")]; + tensor p_23_pad_type_0 = const()[name = tensor("p_23_pad_type_0"), val = tensor("valid")]; + tensor p_23_strides_0 = const()[name = tensor("p_23_strides_0"), val = tensor([1, 1])]; + tensor p_23_pad_0 = const()[name = tensor("p_23_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_23_dilations_0 = const()[name = tensor("p_23_dilations_0"), val = tensor([1, 1])]; + tensor p_23_groups_0 = const()[name = tensor("p_23_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(220251136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221037632))), name = tensor("layers_11_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_23_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_23_dilations_0, groups = p_23_groups_0, pad = p_23_pad_0, pad_type = p_23_pad_type_0, strides = p_23_strides_0, weight = layers_11_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_23_cast_fp16")]; + tensor var_3354 = const()[name = tensor("op_3354"), val = tensor([1, 8, 128, 188])]; + tensor var_3355_cast_fp16 = reshape(shape = var_3354, x = q_with_bias_v_23_cast_fp16)[name = tensor("op_3355_cast_fp16")]; + tensor var_3356 = const()[name = tensor("op_3356"), val = tensor([1, 8, 128, -1])]; + tensor var_3357_cast_fp16 = reshape(shape = var_3356, x = p_23_cast_fp16)[name = tensor("op_3357_cast_fp16")]; + tensor matrix_bd_89_transpose_x_0 = const()[name = tensor("matrix_bd_89_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_89_transpose_y_0 = const()[name = tensor("matrix_bd_89_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_89_cast_fp16 = matmul(transpose_x = matrix_bd_89_transpose_x_0, transpose_y = matrix_bd_89_transpose_y_0, x = var_3355_cast_fp16, y = var_3357_cast_fp16)[name = tensor("matrix_bd_89_cast_fp16")]; + tensor matrix_bd_91_pad_0 = const()[name = tensor("matrix_bd_91_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_91_mode_0 = const()[name = tensor("matrix_bd_91_mode_0"), val = tensor("constant")]; + tensor const_131_to_fp16 = const()[name = tensor("const_131_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_91_cast_fp16 = pad(constant_val = const_131_to_fp16, mode = matrix_bd_91_mode_0, pad = matrix_bd_91_pad_0, x = matrix_bd_89_cast_fp16)[name = tensor("matrix_bd_91_cast_fp16")]; + tensor var_3366 = const()[name = tensor("op_3366"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_3366, x = matrix_bd_91_cast_fp16)[name = tensor("matrix_bd_93_cast_fp16")]; + tensor var_3370_begin_0 = const()[name = tensor("op_3370_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3370_end_0 = const()[name = tensor("op_3370_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3370_end_mask_0 = const()[name = tensor("op_3370_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3370_cast_fp16 = slice_by_index(begin = var_3370_begin_0, end = var_3370_end_0, end_mask = var_3370_end_mask_0, x = matrix_bd_93_cast_fp16)[name = tensor("op_3370_cast_fp16")]; + tensor var_3371 = const()[name = tensor("op_3371"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_95_cast_fp16 = reshape(shape = var_3371, x = var_3370_cast_fp16)[name = tensor("matrix_bd_95_cast_fp16")]; + tensor var_3376_begin_0 = const()[name = tensor("op_3376_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3376_end_0 = const()[name = tensor("op_3376_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3376_end_mask_0 = const()[name = tensor("op_3376_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3376_cast_fp16 = slice_by_index(begin = var_3376_begin_0, end = var_3376_end_0, end_mask = var_3376_end_mask_0, x = matrix_bd_95_cast_fp16)[name = tensor("op_3376_cast_fp16")]; + tensor var_3377_to_fp16 = const()[name = tensor("op_3377_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_23_cast_fp16 = mul(x = var_3376_cast_fp16, y = var_3377_to_fp16)[name = tensor("qk_mask_23_cast_fp16")]; + tensor var_3381 = const()[name = tensor("op_3381"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_3381, x = query_47_cast_fp16)[name = tensor("mh_q_23_cast_fp16")]; + tensor var_3383_to_fp16 = const()[name = tensor("op_3383_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3384_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_3383_to_fp16)[name = tensor("op_3384_cast_fp16")]; + tensor var_3387 = const()[name = tensor("op_3387"), val = tensor([1, 8, 128, 188])]; + tensor var_3388_cast_fp16 = reshape(shape = var_3387, x = key_23_cast_fp16)[name = tensor("op_3388_cast_fp16")]; + tensor mh_w_45_transpose_x_0 = const()[name = tensor("mh_w_45_transpose_x_0"), val = tensor(true)]; + tensor mh_w_45_transpose_y_0 = const()[name = tensor("mh_w_45_transpose_y_0"), val = tensor(false)]; + tensor mh_w_45_cast_fp16 = matmul(transpose_x = mh_w_45_transpose_x_0, transpose_y = mh_w_45_transpose_y_0, x = var_3384_cast_fp16, y = var_3388_cast_fp16)[name = tensor("mh_w_45_cast_fp16")]; + tensor mh_w_47_cast_fp16 = add(x = mh_w_45_cast_fp16, y = qk_mask_23_cast_fp16)[name = tensor("mh_w_47_cast_fp16")]; + tensor var_3392_cast_fp16 = softmax(axis = var_3233, x = mh_w_47_cast_fp16)[name = tensor("op_3392_cast_fp16")]; + tensor var_3393 = const()[name = tensor("op_3393"), val = tensor([1, 8, 128, 188])]; + tensor var_3394_cast_fp16 = reshape(shape = var_3393, x = value_23_cast_fp16)[name = tensor("op_3394_cast_fp16")]; + tensor attn_23_transpose_x_0 = const()[name = tensor("attn_23_transpose_x_0"), val = tensor(false)]; + tensor attn_23_transpose_y_0 = const()[name = tensor("attn_23_transpose_y_0"), val = tensor(true)]; + tensor attn_23_cast_fp16 = matmul(transpose_x = attn_23_transpose_x_0, transpose_y = attn_23_transpose_y_0, x = var_3394_cast_fp16, y = var_3392_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3397 = const()[name = tensor("op_3397"), val = tensor([1, 1024, 1, 188])]; + tensor input_309_cast_fp16 = reshape(shape = var_3397, x = attn_23_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor obj_49_pad_type_0 = const()[name = tensor("obj_49_pad_type_0"), val = tensor("valid")]; + tensor obj_49_strides_0 = const()[name = tensor("obj_49_strides_0"), val = tensor([1, 1])]; + tensor obj_49_pad_0 = const()[name = tensor("obj_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_49_dilations_0 = const()[name = tensor("obj_49_dilations_0"), val = tensor([1, 1])]; + tensor obj_49_groups_0 = const()[name = tensor("obj_49_groups_0"), val = tensor(1)]; + tensor layers_11_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221037824))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221824320))), name = tensor("layers_11_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_49_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_49_dilations_0, groups = obj_49_groups_0, pad = obj_49_pad_0, pad_type = obj_49_pad_type_0, strides = obj_49_strides_0, weight = layers_11_self_attn_o_proj_weight_to_fp16_palettized, x = input_309_cast_fp16)[name = tensor("obj_49_cast_fp16")]; + tensor inputs_115_cast_fp16 = add(x = inputs_113_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_115_cast_fp16")]; + tensor out_115_axes_0 = const()[name = tensor("out_115_axes_0"), val = tensor([1])]; + tensor var_3415_to_fp16 = const()[name = tensor("op_3415_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_3415_to_fp16, x = inputs_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor input_311_gamma_0_to_fp16 = const()[name = tensor("input_311_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221824512)))]; + tensor input_311_beta_0_to_fp16 = const()[name = tensor("input_311_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221826624)))]; + tensor input_311_epsilon_0_to_fp16 = const()[name = tensor("input_311_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_311_cast_fp16 = batch_norm(beta = input_311_beta_0_to_fp16, epsilon = input_311_epsilon_0_to_fp16, gamma = input_311_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor input_313_pad_type_0 = const()[name = tensor("input_313_pad_type_0"), val = tensor("valid")]; + tensor input_313_strides_0 = const()[name = tensor("input_313_strides_0"), val = tensor([1, 1])]; + tensor input_313_pad_0 = const()[name = tensor("input_313_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_313_dilations_0 = const()[name = tensor("input_313_dilations_0"), val = tensor([1, 1])]; + tensor input_313_groups_0 = const()[name = tensor("input_313_groups_0"), val = tensor(1)]; + tensor layers_11_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(221828736))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223401664))), name = tensor("layers_11_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_313_cast_fp16 = conv(dilations = input_313_dilations_0, groups = input_313_groups_0, pad = input_313_pad_0, pad_type = input_313_pad_type_0, strides = input_313_strides_0, weight = layers_11_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_311_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor input_315_split_num_splits_0 = const()[name = tensor("input_315_split_num_splits_0"), val = tensor(2)]; + tensor input_315_split_axis_0 = const()[name = tensor("input_315_split_axis_0"), val = tensor(1)]; + tensor input_315_split_cast_fp16_0, tensor input_315_split_cast_fp16_1 = split(axis = input_315_split_axis_0, num_splits = input_315_split_num_splits_0, x = input_313_cast_fp16)[name = tensor("input_315_split_cast_fp16")]; + tensor input_315_split_1_sigmoid_cast_fp16 = sigmoid(x = input_315_split_cast_fp16_1)[name = tensor("input_315_split_1_sigmoid_cast_fp16")]; + tensor input_315_cast_fp16 = mul(x = input_315_split_cast_fp16_0, y = input_315_split_1_sigmoid_cast_fp16)[name = tensor("input_315_cast_fp16")]; + tensor input_317_pad_type_0 = const()[name = tensor("input_317_pad_type_0"), val = tensor("custom")]; + tensor input_317_pad_0 = const()[name = tensor("input_317_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_317_groups_0 = const()[name = tensor("input_317_groups_0"), val = tensor(1024)]; + tensor input_317_strides_0 = const()[name = tensor("input_317_strides_0"), val = tensor([1, 1])]; + tensor input_317_dilations_0 = const()[name = tensor("input_317_dilations_0"), val = tensor([1, 1])]; + tensor const_290_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223401856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223408832))), name = tensor("const_290_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_291_to_fp16 = const()[name = tensor("const_291_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223409024)))]; + tensor input_319_cast_fp16 = conv(bias = const_291_to_fp16, dilations = input_317_dilations_0, groups = input_317_groups_0, pad = input_317_pad_0, pad_type = input_317_pad_type_0, strides = input_317_strides_0, weight = const_290_to_fp16_palettized, x = input_315_cast_fp16)[name = tensor("input_319_cast_fp16")]; + tensor input_321_cast_fp16 = silu(x = input_319_cast_fp16)[name = tensor("input_321_cast_fp16")]; + tensor x_71_pad_type_0 = const()[name = tensor("x_71_pad_type_0"), val = tensor("valid")]; + tensor x_71_strides_0 = const()[name = tensor("x_71_strides_0"), val = tensor([1, 1])]; + tensor x_71_pad_0 = const()[name = tensor("x_71_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_71_dilations_0 = const()[name = tensor("x_71_dilations_0"), val = tensor([1, 1])]; + tensor x_71_groups_0 = const()[name = tensor("x_71_groups_0"), val = tensor(1)]; + tensor layers_11_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(223411136))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224197632))), name = tensor("layers_11_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_71_cast_fp16 = conv(dilations = x_71_dilations_0, groups = x_71_groups_0, pad = x_71_pad_0, pad_type = x_71_pad_type_0, strides = x_71_strides_0, weight = layers_11_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_321_cast_fp16)[name = tensor("x_71_cast_fp16")]; + tensor inputs_117_cast_fp16 = add(x = inputs_115_cast_fp16, y = x_71_cast_fp16)[name = tensor("inputs_117_cast_fp16")]; + tensor out_117_axes_0 = const()[name = tensor("out_117_axes_0"), val = tensor([1])]; + tensor var_3463_to_fp16 = const()[name = tensor("op_3463_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_3463_to_fp16, x = inputs_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor input_323_gamma_0_to_fp16 = const()[name = tensor("input_323_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224197824)))]; + tensor input_323_beta_0_to_fp16 = const()[name = tensor("input_323_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224199936)))]; + tensor input_323_epsilon_0_to_fp16 = const()[name = tensor("input_323_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_323_cast_fp16 = batch_norm(beta = input_323_beta_0_to_fp16, epsilon = input_323_epsilon_0_to_fp16, gamma = input_323_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("input_323_cast_fp16")]; + tensor input_325_pad_type_0 = const()[name = tensor("input_325_pad_type_0"), val = tensor("valid")]; + tensor input_325_strides_0 = const()[name = tensor("input_325_strides_0"), val = tensor([1, 1])]; + tensor input_325_pad_0 = const()[name = tensor("input_325_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_325_dilations_0 = const()[name = tensor("input_325_dilations_0"), val = tensor([1, 1])]; + tensor input_325_groups_0 = const()[name = tensor("input_325_groups_0"), val = tensor(1)]; + tensor layers_11_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(224202048))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227347840))), name = tensor("layers_11_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_325_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_325_dilations_0, groups = input_325_groups_0, pad = input_325_pad_0, pad_type = input_325_pad_type_0, strides = input_325_strides_0, weight = layers_11_feed_forward2_fc1_weight_to_fp16_palettized, x = input_323_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor input_327_cast_fp16 = silu(x = input_325_cast_fp16)[name = tensor("input_327_cast_fp16")]; + tensor x_73_pad_type_0 = const()[name = tensor("x_73_pad_type_0"), val = tensor("valid")]; + tensor x_73_strides_0 = const()[name = tensor("x_73_strides_0"), val = tensor([1, 1])]; + tensor x_73_pad_0 = const()[name = tensor("x_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_73_dilations_0 = const()[name = tensor("x_73_dilations_0"), val = tensor([1, 1])]; + tensor x_73_groups_0 = const()[name = tensor("x_73_groups_0"), val = tensor(1)]; + tensor op_3491_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227348032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230493824))), name = tensor("op_3491_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_3491_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = op_3491_weight_0_to_fp16_palettized, x = input_327_cast_fp16)[name = tensor("op_3491_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = var_3491_cast_fp16)[name = tensor("inputs_119_cast_fp16")]; + tensor out_119_axes_0 = const()[name = tensor("out_119_axes_0"), val = tensor([1])]; + tensor var_3501_to_fp16 = const()[name = tensor("op_3501_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_3501_to_fp16, x = inputs_119_cast_fp16)[name = tensor("out_119_cast_fp16")]; + tensor inputs_121_gamma_0_to_fp16 = const()[name = tensor("inputs_121_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230494016)))]; + tensor inputs_121_beta_0_to_fp16 = const()[name = tensor("inputs_121_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230496128)))]; + tensor inputs_121_epsilon_0_to_fp16 = const()[name = tensor("inputs_121_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_121_cast_fp16 = batch_norm(beta = inputs_121_beta_0_to_fp16, epsilon = inputs_121_epsilon_0_to_fp16, gamma = inputs_121_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_3515 = const()[name = tensor("op_3515"), val = tensor(3)]; + tensor out_121_axes_0 = const()[name = tensor("out_121_axes_0"), val = tensor([1])]; + tensor var_3546_to_fp16 = const()[name = tensor("op_3546_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_3546_to_fp16, x = inputs_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor input_329_gamma_0_to_fp16 = const()[name = tensor("input_329_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230498240)))]; + tensor input_329_beta_0_to_fp16 = const()[name = tensor("input_329_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230500352)))]; + tensor input_329_epsilon_0_to_fp16 = const()[name = tensor("input_329_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_329_cast_fp16 = batch_norm(beta = input_329_beta_0_to_fp16, epsilon = input_329_epsilon_0_to_fp16, gamma = input_329_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("input_329_cast_fp16")]; + tensor input_331_pad_type_0 = const()[name = tensor("input_331_pad_type_0"), val = tensor("valid")]; + tensor input_331_strides_0 = const()[name = tensor("input_331_strides_0"), val = tensor([1, 1])]; + tensor input_331_pad_0 = const()[name = tensor("input_331_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_331_dilations_0 = const()[name = tensor("input_331_dilations_0"), val = tensor([1, 1])]; + tensor input_331_groups_0 = const()[name = tensor("input_331_groups_0"), val = tensor(1)]; + tensor layers_12_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(230502464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233648256))), name = tensor("layers_12_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_331_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_331_dilations_0, groups = input_331_groups_0, pad = input_331_pad_0, pad_type = input_331_pad_type_0, strides = input_331_strides_0, weight = layers_12_feed_forward1_fc1_weight_to_fp16_palettized, x = input_329_cast_fp16)[name = tensor("input_331_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = input_331_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor x_75_pad_type_0 = const()[name = tensor("x_75_pad_type_0"), val = tensor("valid")]; + tensor x_75_strides_0 = const()[name = tensor("x_75_strides_0"), val = tensor([1, 1])]; + tensor x_75_pad_0 = const()[name = tensor("x_75_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_75_dilations_0 = const()[name = tensor("x_75_dilations_0"), val = tensor([1, 1])]; + tensor x_75_groups_0 = const()[name = tensor("x_75_groups_0"), val = tensor(1)]; + tensor op_3574_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233648448))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236794240))), name = tensor("op_3574_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_3574_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_75_dilations_0, groups = x_75_groups_0, pad = x_75_pad_0, pad_type = x_75_pad_type_0, strides = x_75_strides_0, weight = op_3574_weight_0_to_fp16_palettized, x = input_333_cast_fp16)[name = tensor("op_3574_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = var_3574_cast_fp16)[name = tensor("inputs_123_cast_fp16")]; + tensor out_123_axes_0 = const()[name = tensor("out_123_axes_0"), val = tensor([1])]; + tensor var_3584_to_fp16 = const()[name = tensor("op_3584_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_3584_to_fp16, x = inputs_123_cast_fp16)[name = tensor("out_123_cast_fp16")]; + tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236794432)))]; + tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236796544)))]; + tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_123_cast_fp16)[name = tensor("obj_51_cast_fp16")]; + tensor query_49_pad_type_0 = const()[name = tensor("query_49_pad_type_0"), val = tensor("valid")]; + tensor query_49_strides_0 = const()[name = tensor("query_49_strides_0"), val = tensor([1, 1])]; + tensor query_49_pad_0 = const()[name = tensor("query_49_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_49_dilations_0 = const()[name = tensor("query_49_dilations_0"), val = tensor([1, 1])]; + tensor query_49_groups_0 = const()[name = tensor("query_49_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236798656))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237585152))), name = tensor("layers_12_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_49_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_49_dilations_0, groups = query_49_groups_0, pad = query_49_pad_0, pad_type = query_49_pad_type_0, strides = query_49_strides_0, weight = layers_12_self_attn_q_proj_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor("query_49_cast_fp16")]; + tensor key_25_pad_type_0 = const()[name = tensor("key_25_pad_type_0"), val = tensor("valid")]; + tensor key_25_strides_0 = const()[name = tensor("key_25_strides_0"), val = tensor([1, 1])]; + tensor key_25_pad_0 = const()[name = tensor("key_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_25_dilations_0 = const()[name = tensor("key_25_dilations_0"), val = tensor([1, 1])]; + tensor key_25_groups_0 = const()[name = tensor("key_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(237585344))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238371840))), name = tensor("layers_12_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_25_cast_fp16 = conv(dilations = key_25_dilations_0, groups = key_25_groups_0, pad = key_25_pad_0, pad_type = key_25_pad_type_0, strides = key_25_strides_0, weight = layers_12_self_attn_k_proj_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor("key_25_cast_fp16")]; + tensor value_25_pad_type_0 = const()[name = tensor("value_25_pad_type_0"), val = tensor("valid")]; + tensor value_25_strides_0 = const()[name = tensor("value_25_strides_0"), val = tensor([1, 1])]; + tensor value_25_pad_0 = const()[name = tensor("value_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_25_dilations_0 = const()[name = tensor("value_25_dilations_0"), val = tensor([1, 1])]; + tensor value_25_groups_0 = const()[name = tensor("value_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238372032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239158528))), name = tensor("layers_12_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_25_dilations_0, groups = value_25_groups_0, pad = value_25_pad_0, pad_type = value_25_pad_type_0, strides = value_25_strides_0, weight = layers_12_self_attn_v_proj_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_3622_to_fp16 = const()[name = tensor("op_3622_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239158720)))]; + tensor query_51_cast_fp16 = add(x = query_49_cast_fp16, y = var_3622_to_fp16)[name = tensor("query_51_cast_fp16")]; + tensor var_3625_to_fp16 = const()[name = tensor("op_3625_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239160832)))]; + tensor q_with_bias_v_25_cast_fp16 = add(x = query_49_cast_fp16, y = var_3625_to_fp16)[name = tensor("q_with_bias_v_25_cast_fp16")]; + tensor p_25_pad_type_0 = const()[name = tensor("p_25_pad_type_0"), val = tensor("valid")]; + tensor p_25_strides_0 = const()[name = tensor("p_25_strides_0"), val = tensor([1, 1])]; + tensor p_25_pad_0 = const()[name = tensor("p_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_25_dilations_0 = const()[name = tensor("p_25_dilations_0"), val = tensor([1, 1])]; + tensor p_25_groups_0 = const()[name = tensor("p_25_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239162944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239949440))), name = tensor("layers_12_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_25_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_25_dilations_0, groups = p_25_groups_0, pad = p_25_pad_0, pad_type = p_25_pad_type_0, strides = p_25_strides_0, weight = layers_12_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_25_cast_fp16")]; + tensor var_3636 = const()[name = tensor("op_3636"), val = tensor([1, 8, 128, 188])]; + tensor var_3637_cast_fp16 = reshape(shape = var_3636, x = q_with_bias_v_25_cast_fp16)[name = tensor("op_3637_cast_fp16")]; + tensor var_3638 = const()[name = tensor("op_3638"), val = tensor([1, 8, 128, -1])]; + tensor var_3639_cast_fp16 = reshape(shape = var_3638, x = p_25_cast_fp16)[name = tensor("op_3639_cast_fp16")]; + tensor matrix_bd_97_transpose_x_0 = const()[name = tensor("matrix_bd_97_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_97_transpose_y_0 = const()[name = tensor("matrix_bd_97_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_97_cast_fp16 = matmul(transpose_x = matrix_bd_97_transpose_x_0, transpose_y = matrix_bd_97_transpose_y_0, x = var_3637_cast_fp16, y = var_3639_cast_fp16)[name = tensor("matrix_bd_97_cast_fp16")]; + tensor matrix_bd_99_pad_0 = const()[name = tensor("matrix_bd_99_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_99_mode_0 = const()[name = tensor("matrix_bd_99_mode_0"), val = tensor("constant")]; + tensor const_142_to_fp16 = const()[name = tensor("const_142_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_99_cast_fp16 = pad(constant_val = const_142_to_fp16, mode = matrix_bd_99_mode_0, pad = matrix_bd_99_pad_0, x = matrix_bd_97_cast_fp16)[name = tensor("matrix_bd_99_cast_fp16")]; + tensor var_3648 = const()[name = tensor("op_3648"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_101_cast_fp16 = reshape(shape = var_3648, x = matrix_bd_99_cast_fp16)[name = tensor("matrix_bd_101_cast_fp16")]; + tensor var_3652_begin_0 = const()[name = tensor("op_3652_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3652_end_0 = const()[name = tensor("op_3652_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3652_end_mask_0 = const()[name = tensor("op_3652_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3652_cast_fp16 = slice_by_index(begin = var_3652_begin_0, end = var_3652_end_0, end_mask = var_3652_end_mask_0, x = matrix_bd_101_cast_fp16)[name = tensor("op_3652_cast_fp16")]; + tensor var_3653 = const()[name = tensor("op_3653"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_103_cast_fp16 = reshape(shape = var_3653, x = var_3652_cast_fp16)[name = tensor("matrix_bd_103_cast_fp16")]; + tensor var_3658_begin_0 = const()[name = tensor("op_3658_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3658_end_0 = const()[name = tensor("op_3658_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3658_end_mask_0 = const()[name = tensor("op_3658_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3658_cast_fp16 = slice_by_index(begin = var_3658_begin_0, end = var_3658_end_0, end_mask = var_3658_end_mask_0, x = matrix_bd_103_cast_fp16)[name = tensor("op_3658_cast_fp16")]; + tensor var_3659_to_fp16 = const()[name = tensor("op_3659_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_25_cast_fp16 = mul(x = var_3658_cast_fp16, y = var_3659_to_fp16)[name = tensor("qk_mask_25_cast_fp16")]; + tensor var_3663 = const()[name = tensor("op_3663"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_3663, x = query_51_cast_fp16)[name = tensor("mh_q_25_cast_fp16")]; + tensor var_3665_to_fp16 = const()[name = tensor("op_3665_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3666_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_3665_to_fp16)[name = tensor("op_3666_cast_fp16")]; + tensor var_3669 = const()[name = tensor("op_3669"), val = tensor([1, 8, 128, 188])]; + tensor var_3670_cast_fp16 = reshape(shape = var_3669, x = key_25_cast_fp16)[name = tensor("op_3670_cast_fp16")]; + tensor mh_w_49_transpose_x_0 = const()[name = tensor("mh_w_49_transpose_x_0"), val = tensor(true)]; + tensor mh_w_49_transpose_y_0 = const()[name = tensor("mh_w_49_transpose_y_0"), val = tensor(false)]; + tensor mh_w_49_cast_fp16 = matmul(transpose_x = mh_w_49_transpose_x_0, transpose_y = mh_w_49_transpose_y_0, x = var_3666_cast_fp16, y = var_3670_cast_fp16)[name = tensor("mh_w_49_cast_fp16")]; + tensor mh_w_51_cast_fp16 = add(x = mh_w_49_cast_fp16, y = qk_mask_25_cast_fp16)[name = tensor("mh_w_51_cast_fp16")]; + tensor var_3674_cast_fp16 = softmax(axis = var_3515, x = mh_w_51_cast_fp16)[name = tensor("op_3674_cast_fp16")]; + tensor var_3675 = const()[name = tensor("op_3675"), val = tensor([1, 8, 128, 188])]; + tensor var_3676_cast_fp16 = reshape(shape = var_3675, x = value_25_cast_fp16)[name = tensor("op_3676_cast_fp16")]; + tensor attn_25_transpose_x_0 = const()[name = tensor("attn_25_transpose_x_0"), val = tensor(false)]; + tensor attn_25_transpose_y_0 = const()[name = tensor("attn_25_transpose_y_0"), val = tensor(true)]; + tensor attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = var_3676_cast_fp16, y = var_3674_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3679 = const()[name = tensor("op_3679"), val = tensor([1, 1024, 1, 188])]; + tensor input_335_cast_fp16 = reshape(shape = var_3679, x = attn_25_cast_fp16)[name = tensor("input_335_cast_fp16")]; + tensor obj_53_pad_type_0 = const()[name = tensor("obj_53_pad_type_0"), val = tensor("valid")]; + tensor obj_53_strides_0 = const()[name = tensor("obj_53_strides_0"), val = tensor([1, 1])]; + tensor obj_53_pad_0 = const()[name = tensor("obj_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_53_dilations_0 = const()[name = tensor("obj_53_dilations_0"), val = tensor([1, 1])]; + tensor obj_53_groups_0 = const()[name = tensor("obj_53_groups_0"), val = tensor(1)]; + tensor layers_12_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239949632))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240736128))), name = tensor("layers_12_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_53_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_53_dilations_0, groups = obj_53_groups_0, pad = obj_53_pad_0, pad_type = obj_53_pad_type_0, strides = obj_53_strides_0, weight = layers_12_self_attn_o_proj_weight_to_fp16_palettized, x = input_335_cast_fp16)[name = tensor("obj_53_cast_fp16")]; + tensor inputs_125_cast_fp16 = add(x = inputs_123_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_125_cast_fp16")]; + tensor out_125_axes_0 = const()[name = tensor("out_125_axes_0"), val = tensor([1])]; + tensor var_3697_to_fp16 = const()[name = tensor("op_3697_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_3697_to_fp16, x = inputs_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_337_gamma_0_to_fp16 = const()[name = tensor("input_337_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240736320)))]; + tensor input_337_beta_0_to_fp16 = const()[name = tensor("input_337_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240738432)))]; + tensor input_337_epsilon_0_to_fp16 = const()[name = tensor("input_337_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_337_cast_fp16 = batch_norm(beta = input_337_beta_0_to_fp16, epsilon = input_337_epsilon_0_to_fp16, gamma = input_337_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("input_337_cast_fp16")]; + tensor input_339_pad_type_0 = const()[name = tensor("input_339_pad_type_0"), val = tensor("valid")]; + tensor input_339_strides_0 = const()[name = tensor("input_339_strides_0"), val = tensor([1, 1])]; + tensor input_339_pad_0 = const()[name = tensor("input_339_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_339_dilations_0 = const()[name = tensor("input_339_dilations_0"), val = tensor([1, 1])]; + tensor input_339_groups_0 = const()[name = tensor("input_339_groups_0"), val = tensor(1)]; + tensor layers_12_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(240740544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242313472))), name = tensor("layers_12_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_339_cast_fp16 = conv(dilations = input_339_dilations_0, groups = input_339_groups_0, pad = input_339_pad_0, pad_type = input_339_pad_type_0, strides = input_339_strides_0, weight = layers_12_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_337_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor input_341_split_num_splits_0 = const()[name = tensor("input_341_split_num_splits_0"), val = tensor(2)]; + tensor input_341_split_axis_0 = const()[name = tensor("input_341_split_axis_0"), val = tensor(1)]; + tensor input_341_split_cast_fp16_0, tensor input_341_split_cast_fp16_1 = split(axis = input_341_split_axis_0, num_splits = input_341_split_num_splits_0, x = input_339_cast_fp16)[name = tensor("input_341_split_cast_fp16")]; + tensor input_341_split_1_sigmoid_cast_fp16 = sigmoid(x = input_341_split_cast_fp16_1)[name = tensor("input_341_split_1_sigmoid_cast_fp16")]; + tensor input_341_cast_fp16 = mul(x = input_341_split_cast_fp16_0, y = input_341_split_1_sigmoid_cast_fp16)[name = tensor("input_341_cast_fp16")]; + tensor input_343_pad_type_0 = const()[name = tensor("input_343_pad_type_0"), val = tensor("custom")]; + tensor input_343_pad_0 = const()[name = tensor("input_343_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_343_groups_0 = const()[name = tensor("input_343_groups_0"), val = tensor(1024)]; + tensor input_343_strides_0 = const()[name = tensor("input_343_strides_0"), val = tensor([1, 1])]; + tensor input_343_dilations_0 = const()[name = tensor("input_343_dilations_0"), val = tensor([1, 1])]; + tensor const_292_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242313664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242320640))), name = tensor("const_292_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_293_to_fp16 = const()[name = tensor("const_293_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242320832)))]; + tensor input_345_cast_fp16 = conv(bias = const_293_to_fp16, dilations = input_343_dilations_0, groups = input_343_groups_0, pad = input_343_pad_0, pad_type = input_343_pad_type_0, strides = input_343_strides_0, weight = const_292_to_fp16_palettized, x = input_341_cast_fp16)[name = tensor("input_345_cast_fp16")]; + tensor input_347_cast_fp16 = silu(x = input_345_cast_fp16)[name = tensor("input_347_cast_fp16")]; + tensor x_77_pad_type_0 = const()[name = tensor("x_77_pad_type_0"), val = tensor("valid")]; + tensor x_77_strides_0 = const()[name = tensor("x_77_strides_0"), val = tensor([1, 1])]; + tensor x_77_pad_0 = const()[name = tensor("x_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_77_dilations_0 = const()[name = tensor("x_77_dilations_0"), val = tensor([1, 1])]; + tensor x_77_groups_0 = const()[name = tensor("x_77_groups_0"), val = tensor(1)]; + tensor layers_12_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242322944))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243109440))), name = tensor("layers_12_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_77_cast_fp16 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = layers_12_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_347_cast_fp16)[name = tensor("x_77_cast_fp16")]; + tensor inputs_127_cast_fp16 = add(x = inputs_125_cast_fp16, y = x_77_cast_fp16)[name = tensor("inputs_127_cast_fp16")]; + tensor out_127_axes_0 = const()[name = tensor("out_127_axes_0"), val = tensor([1])]; + tensor var_3745_to_fp16 = const()[name = tensor("op_3745_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_3745_to_fp16, x = inputs_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor input_349_gamma_0_to_fp16 = const()[name = tensor("input_349_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243109632)))]; + tensor input_349_beta_0_to_fp16 = const()[name = tensor("input_349_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243111744)))]; + tensor input_349_epsilon_0_to_fp16 = const()[name = tensor("input_349_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_349_cast_fp16 = batch_norm(beta = input_349_beta_0_to_fp16, epsilon = input_349_epsilon_0_to_fp16, gamma = input_349_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("input_349_cast_fp16")]; + tensor input_351_pad_type_0 = const()[name = tensor("input_351_pad_type_0"), val = tensor("valid")]; + tensor input_351_strides_0 = const()[name = tensor("input_351_strides_0"), val = tensor([1, 1])]; + tensor input_351_pad_0 = const()[name = tensor("input_351_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_351_dilations_0 = const()[name = tensor("input_351_dilations_0"), val = tensor([1, 1])]; + tensor input_351_groups_0 = const()[name = tensor("input_351_groups_0"), val = tensor(1)]; + tensor layers_12_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243113856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246259648))), name = tensor("layers_12_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_351_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_351_dilations_0, groups = input_351_groups_0, pad = input_351_pad_0, pad_type = input_351_pad_type_0, strides = input_351_strides_0, weight = layers_12_feed_forward2_fc1_weight_to_fp16_palettized, x = input_349_cast_fp16)[name = tensor("input_351_cast_fp16")]; + tensor input_353_cast_fp16 = silu(x = input_351_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor x_79_pad_type_0 = const()[name = tensor("x_79_pad_type_0"), val = tensor("valid")]; + tensor x_79_strides_0 = const()[name = tensor("x_79_strides_0"), val = tensor([1, 1])]; + tensor x_79_pad_0 = const()[name = tensor("x_79_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_79_dilations_0 = const()[name = tensor("x_79_dilations_0"), val = tensor([1, 1])]; + tensor x_79_groups_0 = const()[name = tensor("x_79_groups_0"), val = tensor(1)]; + tensor op_3773_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(246259840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249405632))), name = tensor("op_3773_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_3773_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_79_dilations_0, groups = x_79_groups_0, pad = x_79_pad_0, pad_type = x_79_pad_type_0, strides = x_79_strides_0, weight = op_3773_weight_0_to_fp16_palettized, x = input_353_cast_fp16)[name = tensor("op_3773_cast_fp16")]; + tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = var_3773_cast_fp16)[name = tensor("inputs_129_cast_fp16")]; + tensor out_129_axes_0 = const()[name = tensor("out_129_axes_0"), val = tensor([1])]; + tensor var_3783_to_fp16 = const()[name = tensor("op_3783_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_3783_to_fp16, x = inputs_129_cast_fp16)[name = tensor("out_129_cast_fp16")]; + tensor inputs_131_gamma_0_to_fp16 = const()[name = tensor("inputs_131_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249405824)))]; + tensor inputs_131_beta_0_to_fp16 = const()[name = tensor("inputs_131_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249407936)))]; + tensor inputs_131_epsilon_0_to_fp16 = const()[name = tensor("inputs_131_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_131_cast_fp16 = batch_norm(beta = inputs_131_beta_0_to_fp16, epsilon = inputs_131_epsilon_0_to_fp16, gamma = inputs_131_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; + tensor var_3797 = const()[name = tensor("op_3797"), val = tensor(3)]; + tensor out_131_axes_0 = const()[name = tensor("out_131_axes_0"), val = tensor([1])]; + tensor var_3828_to_fp16 = const()[name = tensor("op_3828_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_3828_to_fp16, x = inputs_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_355_gamma_0_to_fp16 = const()[name = tensor("input_355_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249410048)))]; + tensor input_355_beta_0_to_fp16 = const()[name = tensor("input_355_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249412160)))]; + tensor input_355_epsilon_0_to_fp16 = const()[name = tensor("input_355_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_355_cast_fp16 = batch_norm(beta = input_355_beta_0_to_fp16, epsilon = input_355_epsilon_0_to_fp16, gamma = input_355_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor input_357_pad_type_0 = const()[name = tensor("input_357_pad_type_0"), val = tensor("valid")]; + tensor input_357_strides_0 = const()[name = tensor("input_357_strides_0"), val = tensor([1, 1])]; + tensor input_357_pad_0 = const()[name = tensor("input_357_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_357_dilations_0 = const()[name = tensor("input_357_dilations_0"), val = tensor([1, 1])]; + tensor input_357_groups_0 = const()[name = tensor("input_357_groups_0"), val = tensor(1)]; + tensor layers_13_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(249414272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252560064))), name = tensor("layers_13_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_357_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_357_dilations_0, groups = input_357_groups_0, pad = input_357_pad_0, pad_type = input_357_pad_type_0, strides = input_357_strides_0, weight = layers_13_feed_forward1_fc1_weight_to_fp16_palettized, x = input_355_cast_fp16)[name = tensor("input_357_cast_fp16")]; + tensor input_359_cast_fp16 = silu(x = input_357_cast_fp16)[name = tensor("input_359_cast_fp16")]; + tensor x_81_pad_type_0 = const()[name = tensor("x_81_pad_type_0"), val = tensor("valid")]; + tensor x_81_strides_0 = const()[name = tensor("x_81_strides_0"), val = tensor([1, 1])]; + tensor x_81_pad_0 = const()[name = tensor("x_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_81_dilations_0 = const()[name = tensor("x_81_dilations_0"), val = tensor([1, 1])]; + tensor x_81_groups_0 = const()[name = tensor("x_81_groups_0"), val = tensor(1)]; + tensor op_3856_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252560256))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255706048))), name = tensor("op_3856_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_3856_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_81_dilations_0, groups = x_81_groups_0, pad = x_81_pad_0, pad_type = x_81_pad_type_0, strides = x_81_strides_0, weight = op_3856_weight_0_to_fp16_palettized, x = input_359_cast_fp16)[name = tensor("op_3856_cast_fp16")]; + tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = var_3856_cast_fp16)[name = tensor("inputs_133_cast_fp16")]; + tensor out_133_axes_0 = const()[name = tensor("out_133_axes_0"), val = tensor([1])]; + tensor var_3866_to_fp16 = const()[name = tensor("op_3866_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_3866_to_fp16, x = inputs_133_cast_fp16)[name = tensor("out_133_cast_fp16")]; + tensor obj_55_gamma_0_to_fp16 = const()[name = tensor("obj_55_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255706240)))]; + tensor obj_55_beta_0_to_fp16 = const()[name = tensor("obj_55_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255708352)))]; + tensor obj_55_epsilon_0_to_fp16 = const()[name = tensor("obj_55_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_55_cast_fp16 = batch_norm(beta = obj_55_beta_0_to_fp16, epsilon = obj_55_epsilon_0_to_fp16, gamma = obj_55_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_133_cast_fp16)[name = tensor("obj_55_cast_fp16")]; + tensor query_53_pad_type_0 = const()[name = tensor("query_53_pad_type_0"), val = tensor("valid")]; + tensor query_53_strides_0 = const()[name = tensor("query_53_strides_0"), val = tensor([1, 1])]; + tensor query_53_pad_0 = const()[name = tensor("query_53_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_53_dilations_0 = const()[name = tensor("query_53_dilations_0"), val = tensor([1, 1])]; + tensor query_53_groups_0 = const()[name = tensor("query_53_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255710464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256496960))), name = tensor("layers_13_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_53_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_53_dilations_0, groups = query_53_groups_0, pad = query_53_pad_0, pad_type = query_53_pad_type_0, strides = query_53_strides_0, weight = layers_13_self_attn_q_proj_weight_to_fp16_palettized, x = obj_55_cast_fp16)[name = tensor("query_53_cast_fp16")]; + tensor key_27_pad_type_0 = const()[name = tensor("key_27_pad_type_0"), val = tensor("valid")]; + tensor key_27_strides_0 = const()[name = tensor("key_27_strides_0"), val = tensor([1, 1])]; + tensor key_27_pad_0 = const()[name = tensor("key_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_27_dilations_0 = const()[name = tensor("key_27_dilations_0"), val = tensor([1, 1])]; + tensor key_27_groups_0 = const()[name = tensor("key_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256497152))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257283648))), name = tensor("layers_13_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_27_cast_fp16 = conv(dilations = key_27_dilations_0, groups = key_27_groups_0, pad = key_27_pad_0, pad_type = key_27_pad_type_0, strides = key_27_strides_0, weight = layers_13_self_attn_k_proj_weight_to_fp16_palettized, x = obj_55_cast_fp16)[name = tensor("key_27_cast_fp16")]; + tensor value_27_pad_type_0 = const()[name = tensor("value_27_pad_type_0"), val = tensor("valid")]; + tensor value_27_strides_0 = const()[name = tensor("value_27_strides_0"), val = tensor([1, 1])]; + tensor value_27_pad_0 = const()[name = tensor("value_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_27_dilations_0 = const()[name = tensor("value_27_dilations_0"), val = tensor([1, 1])]; + tensor value_27_groups_0 = const()[name = tensor("value_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(257283840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258070336))), name = tensor("layers_13_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_27_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_27_dilations_0, groups = value_27_groups_0, pad = value_27_pad_0, pad_type = value_27_pad_type_0, strides = value_27_strides_0, weight = layers_13_self_attn_v_proj_weight_to_fp16_palettized, x = obj_55_cast_fp16)[name = tensor("value_27_cast_fp16")]; + tensor var_3904_to_fp16 = const()[name = tensor("op_3904_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258070528)))]; + tensor query_55_cast_fp16 = add(x = query_53_cast_fp16, y = var_3904_to_fp16)[name = tensor("query_55_cast_fp16")]; + tensor var_3907_to_fp16 = const()[name = tensor("op_3907_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258072640)))]; + tensor q_with_bias_v_27_cast_fp16 = add(x = query_53_cast_fp16, y = var_3907_to_fp16)[name = tensor("q_with_bias_v_27_cast_fp16")]; + tensor p_27_pad_type_0 = const()[name = tensor("p_27_pad_type_0"), val = tensor("valid")]; + tensor p_27_strides_0 = const()[name = tensor("p_27_strides_0"), val = tensor([1, 1])]; + tensor p_27_pad_0 = const()[name = tensor("p_27_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_27_dilations_0 = const()[name = tensor("p_27_dilations_0"), val = tensor([1, 1])]; + tensor p_27_groups_0 = const()[name = tensor("p_27_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258074752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258861248))), name = tensor("layers_13_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_27_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_27_dilations_0, groups = p_27_groups_0, pad = p_27_pad_0, pad_type = p_27_pad_type_0, strides = p_27_strides_0, weight = layers_13_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_27_cast_fp16")]; + tensor var_3918 = const()[name = tensor("op_3918"), val = tensor([1, 8, 128, 188])]; + tensor var_3919_cast_fp16 = reshape(shape = var_3918, x = q_with_bias_v_27_cast_fp16)[name = tensor("op_3919_cast_fp16")]; + tensor var_3920 = const()[name = tensor("op_3920"), val = tensor([1, 8, 128, -1])]; + tensor var_3921_cast_fp16 = reshape(shape = var_3920, x = p_27_cast_fp16)[name = tensor("op_3921_cast_fp16")]; + tensor matrix_bd_105_transpose_x_0 = const()[name = tensor("matrix_bd_105_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_105_transpose_y_0 = const()[name = tensor("matrix_bd_105_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_105_cast_fp16 = matmul(transpose_x = matrix_bd_105_transpose_x_0, transpose_y = matrix_bd_105_transpose_y_0, x = var_3919_cast_fp16, y = var_3921_cast_fp16)[name = tensor("matrix_bd_105_cast_fp16")]; + tensor matrix_bd_107_pad_0 = const()[name = tensor("matrix_bd_107_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_107_mode_0 = const()[name = tensor("matrix_bd_107_mode_0"), val = tensor("constant")]; + tensor const_153_to_fp16 = const()[name = tensor("const_153_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_107_cast_fp16 = pad(constant_val = const_153_to_fp16, mode = matrix_bd_107_mode_0, pad = matrix_bd_107_pad_0, x = matrix_bd_105_cast_fp16)[name = tensor("matrix_bd_107_cast_fp16")]; + tensor var_3930 = const()[name = tensor("op_3930"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_109_cast_fp16 = reshape(shape = var_3930, x = matrix_bd_107_cast_fp16)[name = tensor("matrix_bd_109_cast_fp16")]; + tensor var_3934_begin_0 = const()[name = tensor("op_3934_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3934_end_0 = const()[name = tensor("op_3934_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3934_end_mask_0 = const()[name = tensor("op_3934_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3934_cast_fp16 = slice_by_index(begin = var_3934_begin_0, end = var_3934_end_0, end_mask = var_3934_end_mask_0, x = matrix_bd_109_cast_fp16)[name = tensor("op_3934_cast_fp16")]; + tensor var_3935 = const()[name = tensor("op_3935"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_111_cast_fp16 = reshape(shape = var_3935, x = var_3934_cast_fp16)[name = tensor("matrix_bd_111_cast_fp16")]; + tensor var_3940_begin_0 = const()[name = tensor("op_3940_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3940_end_0 = const()[name = tensor("op_3940_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3940_end_mask_0 = const()[name = tensor("op_3940_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3940_cast_fp16 = slice_by_index(begin = var_3940_begin_0, end = var_3940_end_0, end_mask = var_3940_end_mask_0, x = matrix_bd_111_cast_fp16)[name = tensor("op_3940_cast_fp16")]; + tensor var_3941_to_fp16 = const()[name = tensor("op_3941_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_27_cast_fp16 = mul(x = var_3940_cast_fp16, y = var_3941_to_fp16)[name = tensor("qk_mask_27_cast_fp16")]; + tensor var_3945 = const()[name = tensor("op_3945"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_3945, x = query_55_cast_fp16)[name = tensor("mh_q_27_cast_fp16")]; + tensor var_3947_to_fp16 = const()[name = tensor("op_3947_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3948_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_3947_to_fp16)[name = tensor("op_3948_cast_fp16")]; + tensor var_3951 = const()[name = tensor("op_3951"), val = tensor([1, 8, 128, 188])]; + tensor var_3952_cast_fp16 = reshape(shape = var_3951, x = key_27_cast_fp16)[name = tensor("op_3952_cast_fp16")]; + tensor mh_w_53_transpose_x_0 = const()[name = tensor("mh_w_53_transpose_x_0"), val = tensor(true)]; + tensor mh_w_53_transpose_y_0 = const()[name = tensor("mh_w_53_transpose_y_0"), val = tensor(false)]; + tensor mh_w_53_cast_fp16 = matmul(transpose_x = mh_w_53_transpose_x_0, transpose_y = mh_w_53_transpose_y_0, x = var_3948_cast_fp16, y = var_3952_cast_fp16)[name = tensor("mh_w_53_cast_fp16")]; + tensor mh_w_55_cast_fp16 = add(x = mh_w_53_cast_fp16, y = qk_mask_27_cast_fp16)[name = tensor("mh_w_55_cast_fp16")]; + tensor var_3956_cast_fp16 = softmax(axis = var_3797, x = mh_w_55_cast_fp16)[name = tensor("op_3956_cast_fp16")]; + tensor var_3957 = const()[name = tensor("op_3957"), val = tensor([1, 8, 128, 188])]; + tensor var_3958_cast_fp16 = reshape(shape = var_3957, x = value_27_cast_fp16)[name = tensor("op_3958_cast_fp16")]; + tensor attn_27_transpose_x_0 = const()[name = tensor("attn_27_transpose_x_0"), val = tensor(false)]; + tensor attn_27_transpose_y_0 = const()[name = tensor("attn_27_transpose_y_0"), val = tensor(true)]; + tensor attn_27_cast_fp16 = matmul(transpose_x = attn_27_transpose_x_0, transpose_y = attn_27_transpose_y_0, x = var_3958_cast_fp16, y = var_3956_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3961 = const()[name = tensor("op_3961"), val = tensor([1, 1024, 1, 188])]; + tensor input_361_cast_fp16 = reshape(shape = var_3961, x = attn_27_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor obj_57_pad_type_0 = const()[name = tensor("obj_57_pad_type_0"), val = tensor("valid")]; + tensor obj_57_strides_0 = const()[name = tensor("obj_57_strides_0"), val = tensor([1, 1])]; + tensor obj_57_pad_0 = const()[name = tensor("obj_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_57_dilations_0 = const()[name = tensor("obj_57_dilations_0"), val = tensor([1, 1])]; + tensor obj_57_groups_0 = const()[name = tensor("obj_57_groups_0"), val = tensor(1)]; + tensor layers_13_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(258861440))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259647936))), name = tensor("layers_13_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_57_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_57_dilations_0, groups = obj_57_groups_0, pad = obj_57_pad_0, pad_type = obj_57_pad_type_0, strides = obj_57_strides_0, weight = layers_13_self_attn_o_proj_weight_to_fp16_palettized, x = input_361_cast_fp16)[name = tensor("obj_57_cast_fp16")]; + tensor inputs_135_cast_fp16 = add(x = inputs_133_cast_fp16, y = obj_57_cast_fp16)[name = tensor("inputs_135_cast_fp16")]; + tensor out_135_axes_0 = const()[name = tensor("out_135_axes_0"), val = tensor([1])]; + tensor var_3979_to_fp16 = const()[name = tensor("op_3979_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_3979_to_fp16, x = inputs_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor input_363_gamma_0_to_fp16 = const()[name = tensor("input_363_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259648128)))]; + tensor input_363_beta_0_to_fp16 = const()[name = tensor("input_363_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259650240)))]; + tensor input_363_epsilon_0_to_fp16 = const()[name = tensor("input_363_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_363_cast_fp16 = batch_norm(beta = input_363_beta_0_to_fp16, epsilon = input_363_epsilon_0_to_fp16, gamma = input_363_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("input_363_cast_fp16")]; + tensor input_365_pad_type_0 = const()[name = tensor("input_365_pad_type_0"), val = tensor("valid")]; + tensor input_365_strides_0 = const()[name = tensor("input_365_strides_0"), val = tensor([1, 1])]; + tensor input_365_pad_0 = const()[name = tensor("input_365_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_365_dilations_0 = const()[name = tensor("input_365_dilations_0"), val = tensor([1, 1])]; + tensor input_365_groups_0 = const()[name = tensor("input_365_groups_0"), val = tensor(1)]; + tensor layers_13_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259652352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261225280))), name = tensor("layers_13_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_365_cast_fp16 = conv(dilations = input_365_dilations_0, groups = input_365_groups_0, pad = input_365_pad_0, pad_type = input_365_pad_type_0, strides = input_365_strides_0, weight = layers_13_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_363_cast_fp16)[name = tensor("input_365_cast_fp16")]; + tensor input_367_split_num_splits_0 = const()[name = tensor("input_367_split_num_splits_0"), val = tensor(2)]; + tensor input_367_split_axis_0 = const()[name = tensor("input_367_split_axis_0"), val = tensor(1)]; + tensor input_367_split_cast_fp16_0, tensor input_367_split_cast_fp16_1 = split(axis = input_367_split_axis_0, num_splits = input_367_split_num_splits_0, x = input_365_cast_fp16)[name = tensor("input_367_split_cast_fp16")]; + tensor input_367_split_1_sigmoid_cast_fp16 = sigmoid(x = input_367_split_cast_fp16_1)[name = tensor("input_367_split_1_sigmoid_cast_fp16")]; + tensor input_367_cast_fp16 = mul(x = input_367_split_cast_fp16_0, y = input_367_split_1_sigmoid_cast_fp16)[name = tensor("input_367_cast_fp16")]; + tensor input_369_pad_type_0 = const()[name = tensor("input_369_pad_type_0"), val = tensor("custom")]; + tensor input_369_pad_0 = const()[name = tensor("input_369_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_369_groups_0 = const()[name = tensor("input_369_groups_0"), val = tensor(1024)]; + tensor input_369_strides_0 = const()[name = tensor("input_369_strides_0"), val = tensor([1, 1])]; + tensor input_369_dilations_0 = const()[name = tensor("input_369_dilations_0"), val = tensor([1, 1])]; + tensor const_294_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261225472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261232448))), name = tensor("const_294_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_295_to_fp16 = const()[name = tensor("const_295_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261232640)))]; + tensor input_371_cast_fp16 = conv(bias = const_295_to_fp16, dilations = input_369_dilations_0, groups = input_369_groups_0, pad = input_369_pad_0, pad_type = input_369_pad_type_0, strides = input_369_strides_0, weight = const_294_to_fp16_palettized, x = input_367_cast_fp16)[name = tensor("input_371_cast_fp16")]; + tensor input_373_cast_fp16 = silu(x = input_371_cast_fp16)[name = tensor("input_373_cast_fp16")]; + tensor x_83_pad_type_0 = const()[name = tensor("x_83_pad_type_0"), val = tensor("valid")]; + tensor x_83_strides_0 = const()[name = tensor("x_83_strides_0"), val = tensor([1, 1])]; + tensor x_83_pad_0 = const()[name = tensor("x_83_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_83_dilations_0 = const()[name = tensor("x_83_dilations_0"), val = tensor([1, 1])]; + tensor x_83_groups_0 = const()[name = tensor("x_83_groups_0"), val = tensor(1)]; + tensor layers_13_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261234752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262021248))), name = tensor("layers_13_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_83_cast_fp16 = conv(dilations = x_83_dilations_0, groups = x_83_groups_0, pad = x_83_pad_0, pad_type = x_83_pad_type_0, strides = x_83_strides_0, weight = layers_13_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_373_cast_fp16)[name = tensor("x_83_cast_fp16")]; + tensor inputs_137_cast_fp16 = add(x = inputs_135_cast_fp16, y = x_83_cast_fp16)[name = tensor("inputs_137_cast_fp16")]; + tensor out_137_axes_0 = const()[name = tensor("out_137_axes_0"), val = tensor([1])]; + tensor var_4027_to_fp16 = const()[name = tensor("op_4027_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_4027_to_fp16, x = inputs_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_375_gamma_0_to_fp16 = const()[name = tensor("input_375_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262021440)))]; + tensor input_375_beta_0_to_fp16 = const()[name = tensor("input_375_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262023552)))]; + tensor input_375_epsilon_0_to_fp16 = const()[name = tensor("input_375_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_375_cast_fp16 = batch_norm(beta = input_375_beta_0_to_fp16, epsilon = input_375_epsilon_0_to_fp16, gamma = input_375_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("input_375_cast_fp16")]; + tensor input_377_pad_type_0 = const()[name = tensor("input_377_pad_type_0"), val = tensor("valid")]; + tensor input_377_strides_0 = const()[name = tensor("input_377_strides_0"), val = tensor([1, 1])]; + tensor input_377_pad_0 = const()[name = tensor("input_377_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_377_dilations_0 = const()[name = tensor("input_377_dilations_0"), val = tensor([1, 1])]; + tensor input_377_groups_0 = const()[name = tensor("input_377_groups_0"), val = tensor(1)]; + tensor layers_13_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262025664))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265171456))), name = tensor("layers_13_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_377_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_377_dilations_0, groups = input_377_groups_0, pad = input_377_pad_0, pad_type = input_377_pad_type_0, strides = input_377_strides_0, weight = layers_13_feed_forward2_fc1_weight_to_fp16_palettized, x = input_375_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor input_379_cast_fp16 = silu(x = input_377_cast_fp16)[name = tensor("input_379_cast_fp16")]; + tensor x_85_pad_type_0 = const()[name = tensor("x_85_pad_type_0"), val = tensor("valid")]; + tensor x_85_strides_0 = const()[name = tensor("x_85_strides_0"), val = tensor([1, 1])]; + tensor x_85_pad_0 = const()[name = tensor("x_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_85_dilations_0 = const()[name = tensor("x_85_dilations_0"), val = tensor([1, 1])]; + tensor x_85_groups_0 = const()[name = tensor("x_85_groups_0"), val = tensor(1)]; + tensor op_4055_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(265171648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268317440))), name = tensor("op_4055_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4055_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_85_dilations_0, groups = x_85_groups_0, pad = x_85_pad_0, pad_type = x_85_pad_type_0, strides = x_85_strides_0, weight = op_4055_weight_0_to_fp16_palettized, x = input_379_cast_fp16)[name = tensor("op_4055_cast_fp16")]; + tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = var_4055_cast_fp16)[name = tensor("inputs_139_cast_fp16")]; + tensor out_139_axes_0 = const()[name = tensor("out_139_axes_0"), val = tensor([1])]; + tensor var_4065_to_fp16 = const()[name = tensor("op_4065_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_4065_to_fp16, x = inputs_139_cast_fp16)[name = tensor("out_139_cast_fp16")]; + tensor inputs_141_gamma_0_to_fp16 = const()[name = tensor("inputs_141_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268317632)))]; + tensor inputs_141_beta_0_to_fp16 = const()[name = tensor("inputs_141_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268319744)))]; + tensor inputs_141_epsilon_0_to_fp16 = const()[name = tensor("inputs_141_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_141_cast_fp16 = batch_norm(beta = inputs_141_beta_0_to_fp16, epsilon = inputs_141_epsilon_0_to_fp16, gamma = inputs_141_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; + tensor var_4079 = const()[name = tensor("op_4079"), val = tensor(3)]; + tensor out_141_axes_0 = const()[name = tensor("out_141_axes_0"), val = tensor([1])]; + tensor var_4110_to_fp16 = const()[name = tensor("op_4110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_4110_to_fp16, x = inputs_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor input_381_gamma_0_to_fp16 = const()[name = tensor("input_381_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268321856)))]; + tensor input_381_beta_0_to_fp16 = const()[name = tensor("input_381_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268323968)))]; + tensor input_381_epsilon_0_to_fp16 = const()[name = tensor("input_381_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_381_cast_fp16 = batch_norm(beta = input_381_beta_0_to_fp16, epsilon = input_381_epsilon_0_to_fp16, gamma = input_381_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("input_381_cast_fp16")]; + tensor input_383_pad_type_0 = const()[name = tensor("input_383_pad_type_0"), val = tensor("valid")]; + tensor input_383_strides_0 = const()[name = tensor("input_383_strides_0"), val = tensor([1, 1])]; + tensor input_383_pad_0 = const()[name = tensor("input_383_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_383_dilations_0 = const()[name = tensor("input_383_dilations_0"), val = tensor([1, 1])]; + tensor input_383_groups_0 = const()[name = tensor("input_383_groups_0"), val = tensor(1)]; + tensor layers_14_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268326080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271471872))), name = tensor("layers_14_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_383_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_383_dilations_0, groups = input_383_groups_0, pad = input_383_pad_0, pad_type = input_383_pad_type_0, strides = input_383_strides_0, weight = layers_14_feed_forward1_fc1_weight_to_fp16_palettized, x = input_381_cast_fp16)[name = tensor("input_383_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = input_383_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor x_87_pad_type_0 = const()[name = tensor("x_87_pad_type_0"), val = tensor("valid")]; + tensor x_87_strides_0 = const()[name = tensor("x_87_strides_0"), val = tensor([1, 1])]; + tensor x_87_pad_0 = const()[name = tensor("x_87_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_87_dilations_0 = const()[name = tensor("x_87_dilations_0"), val = tensor([1, 1])]; + tensor x_87_groups_0 = const()[name = tensor("x_87_groups_0"), val = tensor(1)]; + tensor op_4138_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(271472064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274617856))), name = tensor("op_4138_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4138_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, 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 = op_4138_weight_0_to_fp16_palettized, x = input_385_cast_fp16)[name = tensor("op_4138_cast_fp16")]; + tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = var_4138_cast_fp16)[name = tensor("inputs_143_cast_fp16")]; + tensor out_143_axes_0 = const()[name = tensor("out_143_axes_0"), val = tensor([1])]; + tensor var_4148_to_fp16 = const()[name = tensor("op_4148_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_4148_to_fp16, x = inputs_143_cast_fp16)[name = tensor("out_143_cast_fp16")]; + tensor obj_59_gamma_0_to_fp16 = const()[name = tensor("obj_59_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274618048)))]; + tensor obj_59_beta_0_to_fp16 = const()[name = tensor("obj_59_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274620160)))]; + tensor obj_59_epsilon_0_to_fp16 = const()[name = tensor("obj_59_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_59_cast_fp16 = batch_norm(beta = obj_59_beta_0_to_fp16, epsilon = obj_59_epsilon_0_to_fp16, gamma = obj_59_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_143_cast_fp16)[name = tensor("obj_59_cast_fp16")]; + tensor query_57_pad_type_0 = const()[name = tensor("query_57_pad_type_0"), val = tensor("valid")]; + tensor query_57_strides_0 = const()[name = tensor("query_57_strides_0"), val = tensor([1, 1])]; + tensor query_57_pad_0 = const()[name = tensor("query_57_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_57_dilations_0 = const()[name = tensor("query_57_dilations_0"), val = tensor([1, 1])]; + tensor query_57_groups_0 = const()[name = tensor("query_57_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(274622272))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275408768))), name = tensor("layers_14_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_57_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_57_dilations_0, groups = query_57_groups_0, pad = query_57_pad_0, pad_type = query_57_pad_type_0, strides = query_57_strides_0, weight = layers_14_self_attn_q_proj_weight_to_fp16_palettized, x = obj_59_cast_fp16)[name = tensor("query_57_cast_fp16")]; + tensor key_29_pad_type_0 = const()[name = tensor("key_29_pad_type_0"), val = tensor("valid")]; + tensor key_29_strides_0 = const()[name = tensor("key_29_strides_0"), val = tensor([1, 1])]; + tensor key_29_pad_0 = const()[name = tensor("key_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_29_dilations_0 = const()[name = tensor("key_29_dilations_0"), val = tensor([1, 1])]; + tensor key_29_groups_0 = const()[name = tensor("key_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(275408960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276195456))), name = tensor("layers_14_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_29_cast_fp16 = conv(dilations = key_29_dilations_0, groups = key_29_groups_0, pad = key_29_pad_0, pad_type = key_29_pad_type_0, strides = key_29_strides_0, weight = layers_14_self_attn_k_proj_weight_to_fp16_palettized, x = obj_59_cast_fp16)[name = tensor("key_29_cast_fp16")]; + tensor value_29_pad_type_0 = const()[name = tensor("value_29_pad_type_0"), val = tensor("valid")]; + tensor value_29_strides_0 = const()[name = tensor("value_29_strides_0"), val = tensor([1, 1])]; + tensor value_29_pad_0 = const()[name = tensor("value_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_29_dilations_0 = const()[name = tensor("value_29_dilations_0"), val = tensor([1, 1])]; + tensor value_29_groups_0 = const()[name = tensor("value_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276195648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276982144))), name = tensor("layers_14_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_29_dilations_0, groups = value_29_groups_0, pad = value_29_pad_0, pad_type = value_29_pad_type_0, strides = value_29_strides_0, weight = layers_14_self_attn_v_proj_weight_to_fp16_palettized, x = obj_59_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_4186_to_fp16 = const()[name = tensor("op_4186_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276982336)))]; + tensor query_59_cast_fp16 = add(x = query_57_cast_fp16, y = var_4186_to_fp16)[name = tensor("query_59_cast_fp16")]; + tensor var_4189_to_fp16 = const()[name = tensor("op_4189_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276984448)))]; + tensor q_with_bias_v_29_cast_fp16 = add(x = query_57_cast_fp16, y = var_4189_to_fp16)[name = tensor("q_with_bias_v_29_cast_fp16")]; + tensor p_29_pad_type_0 = const()[name = tensor("p_29_pad_type_0"), val = tensor("valid")]; + tensor p_29_strides_0 = const()[name = tensor("p_29_strides_0"), val = tensor([1, 1])]; + tensor p_29_pad_0 = const()[name = tensor("p_29_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_29_dilations_0 = const()[name = tensor("p_29_dilations_0"), val = tensor([1, 1])]; + tensor p_29_groups_0 = const()[name = tensor("p_29_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(276986560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277773056))), name = tensor("layers_14_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_29_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_29_dilations_0, groups = p_29_groups_0, pad = p_29_pad_0, pad_type = p_29_pad_type_0, strides = p_29_strides_0, weight = layers_14_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_29_cast_fp16")]; + tensor var_4200 = const()[name = tensor("op_4200"), val = tensor([1, 8, 128, 188])]; + tensor var_4201_cast_fp16 = reshape(shape = var_4200, x = q_with_bias_v_29_cast_fp16)[name = tensor("op_4201_cast_fp16")]; + tensor var_4202 = const()[name = tensor("op_4202"), val = tensor([1, 8, 128, -1])]; + tensor var_4203_cast_fp16 = reshape(shape = var_4202, x = p_29_cast_fp16)[name = tensor("op_4203_cast_fp16")]; + tensor matrix_bd_113_transpose_x_0 = const()[name = tensor("matrix_bd_113_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_113_transpose_y_0 = const()[name = tensor("matrix_bd_113_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_113_cast_fp16 = matmul(transpose_x = matrix_bd_113_transpose_x_0, transpose_y = matrix_bd_113_transpose_y_0, x = var_4201_cast_fp16, y = var_4203_cast_fp16)[name = tensor("matrix_bd_113_cast_fp16")]; + tensor matrix_bd_115_pad_0 = const()[name = tensor("matrix_bd_115_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_115_mode_0 = const()[name = tensor("matrix_bd_115_mode_0"), val = tensor("constant")]; + tensor const_164_to_fp16 = const()[name = tensor("const_164_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_115_cast_fp16 = pad(constant_val = const_164_to_fp16, mode = matrix_bd_115_mode_0, pad = matrix_bd_115_pad_0, x = matrix_bd_113_cast_fp16)[name = tensor("matrix_bd_115_cast_fp16")]; + tensor var_4212 = const()[name = tensor("op_4212"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_117_cast_fp16 = reshape(shape = var_4212, x = matrix_bd_115_cast_fp16)[name = tensor("matrix_bd_117_cast_fp16")]; + tensor var_4216_begin_0 = const()[name = tensor("op_4216_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4216_end_0 = const()[name = tensor("op_4216_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_4216_end_mask_0 = const()[name = tensor("op_4216_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4216_cast_fp16 = slice_by_index(begin = var_4216_begin_0, end = var_4216_end_0, end_mask = var_4216_end_mask_0, x = matrix_bd_117_cast_fp16)[name = tensor("op_4216_cast_fp16")]; + tensor var_4217 = const()[name = tensor("op_4217"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_119_cast_fp16 = reshape(shape = var_4217, x = var_4216_cast_fp16)[name = tensor("matrix_bd_119_cast_fp16")]; + tensor var_4222_begin_0 = const()[name = tensor("op_4222_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4222_end_0 = const()[name = tensor("op_4222_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_4222_end_mask_0 = const()[name = tensor("op_4222_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_4222_cast_fp16 = slice_by_index(begin = var_4222_begin_0, end = var_4222_end_0, end_mask = var_4222_end_mask_0, x = matrix_bd_119_cast_fp16)[name = tensor("op_4222_cast_fp16")]; + tensor var_4223_to_fp16 = const()[name = tensor("op_4223_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_29_cast_fp16 = mul(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = tensor("qk_mask_29_cast_fp16")]; + tensor var_4227 = const()[name = tensor("op_4227"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_4227, x = query_59_cast_fp16)[name = tensor("mh_q_29_cast_fp16")]; + tensor var_4229_to_fp16 = const()[name = tensor("op_4229_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_4230_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_4229_to_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor var_4233 = const()[name = tensor("op_4233"), val = tensor([1, 8, 128, 188])]; + tensor var_4234_cast_fp16 = reshape(shape = var_4233, x = key_29_cast_fp16)[name = tensor("op_4234_cast_fp16")]; + tensor mh_w_57_transpose_x_0 = const()[name = tensor("mh_w_57_transpose_x_0"), val = tensor(true)]; + tensor mh_w_57_transpose_y_0 = const()[name = tensor("mh_w_57_transpose_y_0"), val = tensor(false)]; + tensor mh_w_57_cast_fp16 = matmul(transpose_x = mh_w_57_transpose_x_0, transpose_y = mh_w_57_transpose_y_0, x = var_4230_cast_fp16, y = var_4234_cast_fp16)[name = tensor("mh_w_57_cast_fp16")]; + tensor mh_w_59_cast_fp16 = add(x = mh_w_57_cast_fp16, y = qk_mask_29_cast_fp16)[name = tensor("mh_w_59_cast_fp16")]; + tensor var_4238_cast_fp16 = softmax(axis = var_4079, x = mh_w_59_cast_fp16)[name = tensor("op_4238_cast_fp16")]; + tensor var_4239 = const()[name = tensor("op_4239"), val = tensor([1, 8, 128, 188])]; + tensor var_4240_cast_fp16 = reshape(shape = var_4239, x = value_29_cast_fp16)[name = tensor("op_4240_cast_fp16")]; + tensor attn_29_transpose_x_0 = const()[name = tensor("attn_29_transpose_x_0"), val = tensor(false)]; + tensor attn_29_transpose_y_0 = const()[name = tensor("attn_29_transpose_y_0"), val = tensor(true)]; + tensor attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = var_4240_cast_fp16, y = var_4238_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_4243 = const()[name = tensor("op_4243"), val = tensor([1, 1024, 1, 188])]; + tensor input_387_cast_fp16 = reshape(shape = var_4243, x = attn_29_cast_fp16)[name = tensor("input_387_cast_fp16")]; + tensor obj_61_pad_type_0 = const()[name = tensor("obj_61_pad_type_0"), val = tensor("valid")]; + tensor obj_61_strides_0 = const()[name = tensor("obj_61_strides_0"), val = tensor([1, 1])]; + tensor obj_61_pad_0 = const()[name = tensor("obj_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_61_dilations_0 = const()[name = tensor("obj_61_dilations_0"), val = tensor([1, 1])]; + tensor obj_61_groups_0 = const()[name = tensor("obj_61_groups_0"), val = tensor(1)]; + tensor layers_14_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277773248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278559744))), name = tensor("layers_14_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_61_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_61_dilations_0, groups = obj_61_groups_0, pad = obj_61_pad_0, pad_type = obj_61_pad_type_0, strides = obj_61_strides_0, weight = layers_14_self_attn_o_proj_weight_to_fp16_palettized, x = input_387_cast_fp16)[name = tensor("obj_61_cast_fp16")]; + tensor inputs_145_cast_fp16 = add(x = inputs_143_cast_fp16, y = obj_61_cast_fp16)[name = tensor("inputs_145_cast_fp16")]; + tensor out_145_axes_0 = const()[name = tensor("out_145_axes_0"), val = tensor([1])]; + tensor var_4261_to_fp16 = const()[name = tensor("op_4261_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_145_cast_fp16 = layer_norm(axes = out_145_axes_0, epsilon = var_4261_to_fp16, x = inputs_145_cast_fp16)[name = tensor("out_145_cast_fp16")]; + tensor input_389_gamma_0_to_fp16 = const()[name = tensor("input_389_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278559936)))]; + tensor input_389_beta_0_to_fp16 = const()[name = tensor("input_389_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278562048)))]; + tensor input_389_epsilon_0_to_fp16 = const()[name = tensor("input_389_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_389_cast_fp16 = batch_norm(beta = input_389_beta_0_to_fp16, epsilon = input_389_epsilon_0_to_fp16, gamma = input_389_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor("input_389_cast_fp16")]; + tensor input_391_pad_type_0 = const()[name = tensor("input_391_pad_type_0"), val = tensor("valid")]; + tensor input_391_strides_0 = const()[name = tensor("input_391_strides_0"), val = tensor([1, 1])]; + tensor input_391_pad_0 = const()[name = tensor("input_391_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_391_dilations_0 = const()[name = tensor("input_391_dilations_0"), val = tensor([1, 1])]; + tensor input_391_groups_0 = const()[name = tensor("input_391_groups_0"), val = tensor(1)]; + tensor layers_14_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278564160))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280137088))), name = tensor("layers_14_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_391_cast_fp16 = conv(dilations = input_391_dilations_0, groups = input_391_groups_0, pad = input_391_pad_0, pad_type = input_391_pad_type_0, strides = input_391_strides_0, weight = layers_14_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_389_cast_fp16)[name = tensor("input_391_cast_fp16")]; + tensor input_393_split_num_splits_0 = const()[name = tensor("input_393_split_num_splits_0"), val = tensor(2)]; + tensor input_393_split_axis_0 = const()[name = tensor("input_393_split_axis_0"), val = tensor(1)]; + tensor input_393_split_cast_fp16_0, tensor input_393_split_cast_fp16_1 = split(axis = input_393_split_axis_0, num_splits = input_393_split_num_splits_0, x = input_391_cast_fp16)[name = tensor("input_393_split_cast_fp16")]; + tensor input_393_split_1_sigmoid_cast_fp16 = sigmoid(x = input_393_split_cast_fp16_1)[name = tensor("input_393_split_1_sigmoid_cast_fp16")]; + tensor input_393_cast_fp16 = mul(x = input_393_split_cast_fp16_0, y = input_393_split_1_sigmoid_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor input_395_pad_type_0 = const()[name = tensor("input_395_pad_type_0"), val = tensor("custom")]; + tensor input_395_pad_0 = const()[name = tensor("input_395_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_395_groups_0 = const()[name = tensor("input_395_groups_0"), val = tensor(1024)]; + tensor input_395_strides_0 = const()[name = tensor("input_395_strides_0"), val = tensor([1, 1])]; + tensor input_395_dilations_0 = const()[name = tensor("input_395_dilations_0"), val = tensor([1, 1])]; + tensor const_296_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280137280))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280144256))), name = tensor("const_296_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_297_to_fp16 = const()[name = tensor("const_297_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280144448)))]; + tensor input_397_cast_fp16 = conv(bias = const_297_to_fp16, dilations = input_395_dilations_0, groups = input_395_groups_0, pad = input_395_pad_0, pad_type = input_395_pad_type_0, strides = input_395_strides_0, weight = const_296_to_fp16_palettized, x = input_393_cast_fp16)[name = tensor("input_397_cast_fp16")]; + tensor input_399_cast_fp16 = silu(x = input_397_cast_fp16)[name = tensor("input_399_cast_fp16")]; + tensor x_89_pad_type_0 = const()[name = tensor("x_89_pad_type_0"), val = tensor("valid")]; + tensor x_89_strides_0 = const()[name = tensor("x_89_strides_0"), val = tensor([1, 1])]; + tensor x_89_pad_0 = const()[name = tensor("x_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_89_dilations_0 = const()[name = tensor("x_89_dilations_0"), val = tensor([1, 1])]; + tensor x_89_groups_0 = const()[name = tensor("x_89_groups_0"), val = tensor(1)]; + tensor layers_14_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280146560))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280933056))), name = tensor("layers_14_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_89_cast_fp16 = conv(dilations = x_89_dilations_0, groups = x_89_groups_0, pad = x_89_pad_0, pad_type = x_89_pad_type_0, strides = x_89_strides_0, weight = layers_14_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_399_cast_fp16)[name = tensor("x_89_cast_fp16")]; + tensor inputs_147_cast_fp16 = add(x = inputs_145_cast_fp16, y = x_89_cast_fp16)[name = tensor("inputs_147_cast_fp16")]; + tensor out_147_axes_0 = const()[name = tensor("out_147_axes_0"), val = tensor([1])]; + tensor var_4309_to_fp16 = const()[name = tensor("op_4309_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_147_cast_fp16 = layer_norm(axes = out_147_axes_0, epsilon = var_4309_to_fp16, x = inputs_147_cast_fp16)[name = tensor("out_147_cast_fp16")]; + tensor input_401_gamma_0_to_fp16 = const()[name = tensor("input_401_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280933248)))]; + tensor input_401_beta_0_to_fp16 = const()[name = tensor("input_401_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280935360)))]; + tensor input_401_epsilon_0_to_fp16 = const()[name = tensor("input_401_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_401_cast_fp16 = batch_norm(beta = input_401_beta_0_to_fp16, epsilon = input_401_epsilon_0_to_fp16, gamma = input_401_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor input_403_pad_type_0 = const()[name = tensor("input_403_pad_type_0"), val = tensor("valid")]; + tensor input_403_strides_0 = const()[name = tensor("input_403_strides_0"), val = tensor([1, 1])]; + tensor input_403_pad_0 = const()[name = tensor("input_403_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_403_dilations_0 = const()[name = tensor("input_403_dilations_0"), val = tensor([1, 1])]; + tensor input_403_groups_0 = const()[name = tensor("input_403_groups_0"), val = tensor(1)]; + tensor layers_14_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280937472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284083264))), name = tensor("layers_14_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_403_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_403_dilations_0, groups = input_403_groups_0, pad = input_403_pad_0, pad_type = input_403_pad_type_0, strides = input_403_strides_0, weight = layers_14_feed_forward2_fc1_weight_to_fp16_palettized, x = input_401_cast_fp16)[name = tensor("input_403_cast_fp16")]; + tensor input_405_cast_fp16 = silu(x = input_403_cast_fp16)[name = tensor("input_405_cast_fp16")]; + tensor x_91_pad_type_0 = const()[name = tensor("x_91_pad_type_0"), val = tensor("valid")]; + tensor x_91_strides_0 = const()[name = tensor("x_91_strides_0"), val = tensor([1, 1])]; + tensor x_91_pad_0 = const()[name = tensor("x_91_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_91_dilations_0 = const()[name = tensor("x_91_dilations_0"), val = tensor([1, 1])]; + tensor x_91_groups_0 = const()[name = tensor("x_91_groups_0"), val = tensor(1)]; + tensor op_4337_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284083456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287229248))), name = tensor("op_4337_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4337_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_91_dilations_0, groups = x_91_groups_0, pad = x_91_pad_0, pad_type = x_91_pad_type_0, strides = x_91_strides_0, weight = op_4337_weight_0_to_fp16_palettized, x = input_405_cast_fp16)[name = tensor("op_4337_cast_fp16")]; + tensor inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = var_4337_cast_fp16)[name = tensor("inputs_149_cast_fp16")]; + tensor out_149_axes_0 = const()[name = tensor("out_149_axes_0"), val = tensor([1])]; + tensor var_4347_to_fp16 = const()[name = tensor("op_4347_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_149_cast_fp16 = layer_norm(axes = out_149_axes_0, epsilon = var_4347_to_fp16, x = inputs_149_cast_fp16)[name = tensor("out_149_cast_fp16")]; + tensor inputs_151_gamma_0_to_fp16 = const()[name = tensor("inputs_151_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287229440)))]; + tensor inputs_151_beta_0_to_fp16 = const()[name = tensor("inputs_151_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287231552)))]; + tensor inputs_151_epsilon_0_to_fp16 = const()[name = tensor("inputs_151_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_151_cast_fp16 = batch_norm(beta = inputs_151_beta_0_to_fp16, epsilon = inputs_151_epsilon_0_to_fp16, gamma = inputs_151_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor("inputs_151_cast_fp16")]; + tensor var_4361 = const()[name = tensor("op_4361"), val = tensor(3)]; + tensor out_151_axes_0 = const()[name = tensor("out_151_axes_0"), val = tensor([1])]; + tensor var_4392_to_fp16 = const()[name = tensor("op_4392_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_151_cast_fp16 = layer_norm(axes = out_151_axes_0, epsilon = var_4392_to_fp16, x = inputs_151_cast_fp16)[name = tensor("out_151_cast_fp16")]; + tensor input_407_gamma_0_to_fp16 = const()[name = tensor("input_407_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287233664)))]; + tensor input_407_beta_0_to_fp16 = const()[name = tensor("input_407_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287235776)))]; + tensor input_407_epsilon_0_to_fp16 = const()[name = tensor("input_407_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_407_cast_fp16 = batch_norm(beta = input_407_beta_0_to_fp16, epsilon = input_407_epsilon_0_to_fp16, gamma = input_407_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor input_409_pad_type_0 = const()[name = tensor("input_409_pad_type_0"), val = tensor("valid")]; + tensor input_409_strides_0 = const()[name = tensor("input_409_strides_0"), val = tensor([1, 1])]; + tensor input_409_pad_0 = const()[name = tensor("input_409_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_409_dilations_0 = const()[name = tensor("input_409_dilations_0"), val = tensor([1, 1])]; + tensor input_409_groups_0 = const()[name = tensor("input_409_groups_0"), val = tensor(1)]; + tensor layers_15_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(287237888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290383680))), name = tensor("layers_15_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_409_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_409_dilations_0, groups = input_409_groups_0, pad = input_409_pad_0, pad_type = input_409_pad_type_0, strides = input_409_strides_0, weight = layers_15_feed_forward1_fc1_weight_to_fp16_palettized, x = input_407_cast_fp16)[name = tensor("input_409_cast_fp16")]; + tensor input_411_cast_fp16 = silu(x = input_409_cast_fp16)[name = tensor("input_411_cast_fp16")]; + tensor x_93_pad_type_0 = const()[name = tensor("x_93_pad_type_0"), val = tensor("valid")]; + tensor x_93_strides_0 = const()[name = tensor("x_93_strides_0"), val = tensor([1, 1])]; + tensor x_93_pad_0 = const()[name = tensor("x_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_93_dilations_0 = const()[name = tensor("x_93_dilations_0"), val = tensor([1, 1])]; + tensor x_93_groups_0 = const()[name = tensor("x_93_groups_0"), val = tensor(1)]; + tensor op_4420_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(290383872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293529664))), name = tensor("op_4420_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4420_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_93_dilations_0, groups = x_93_groups_0, pad = x_93_pad_0, pad_type = x_93_pad_type_0, strides = x_93_strides_0, weight = op_4420_weight_0_to_fp16_palettized, x = input_411_cast_fp16)[name = tensor("op_4420_cast_fp16")]; + tensor inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = var_4420_cast_fp16)[name = tensor("inputs_153_cast_fp16")]; + tensor out_153_axes_0 = const()[name = tensor("out_153_axes_0"), val = tensor([1])]; + tensor var_4430_to_fp16 = const()[name = tensor("op_4430_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_153_cast_fp16 = layer_norm(axes = out_153_axes_0, epsilon = var_4430_to_fp16, x = inputs_153_cast_fp16)[name = tensor("out_153_cast_fp16")]; + tensor obj_63_gamma_0_to_fp16 = const()[name = tensor("obj_63_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293529856)))]; + tensor obj_63_beta_0_to_fp16 = const()[name = tensor("obj_63_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293531968)))]; + tensor obj_63_epsilon_0_to_fp16 = const()[name = tensor("obj_63_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_63_cast_fp16 = batch_norm(beta = obj_63_beta_0_to_fp16, epsilon = obj_63_epsilon_0_to_fp16, gamma = obj_63_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_153_cast_fp16)[name = tensor("obj_63_cast_fp16")]; + tensor query_61_pad_type_0 = const()[name = tensor("query_61_pad_type_0"), val = tensor("valid")]; + tensor query_61_strides_0 = const()[name = tensor("query_61_strides_0"), val = tensor([1, 1])]; + tensor query_61_pad_0 = const()[name = tensor("query_61_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_61_dilations_0 = const()[name = tensor("query_61_dilations_0"), val = tensor([1, 1])]; + tensor query_61_groups_0 = const()[name = tensor("query_61_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293534080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294320576))), name = tensor("layers_15_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_61_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_61_dilations_0, groups = query_61_groups_0, pad = query_61_pad_0, pad_type = query_61_pad_type_0, strides = query_61_strides_0, weight = layers_15_self_attn_q_proj_weight_to_fp16_palettized, x = obj_63_cast_fp16)[name = tensor("query_61_cast_fp16")]; + tensor key_31_pad_type_0 = const()[name = tensor("key_31_pad_type_0"), val = tensor("valid")]; + tensor key_31_strides_0 = const()[name = tensor("key_31_strides_0"), val = tensor([1, 1])]; + tensor key_31_pad_0 = const()[name = tensor("key_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_31_dilations_0 = const()[name = tensor("key_31_dilations_0"), val = tensor([1, 1])]; + tensor key_31_groups_0 = const()[name = tensor("key_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294320768))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295107264))), name = tensor("layers_15_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_31_cast_fp16 = conv(dilations = key_31_dilations_0, groups = key_31_groups_0, pad = key_31_pad_0, pad_type = key_31_pad_type_0, strides = key_31_strides_0, weight = layers_15_self_attn_k_proj_weight_to_fp16_palettized, x = obj_63_cast_fp16)[name = tensor("key_31_cast_fp16")]; + tensor value_31_pad_type_0 = const()[name = tensor("value_31_pad_type_0"), val = tensor("valid")]; + tensor value_31_strides_0 = const()[name = tensor("value_31_strides_0"), val = tensor([1, 1])]; + tensor value_31_pad_0 = const()[name = tensor("value_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_31_dilations_0 = const()[name = tensor("value_31_dilations_0"), val = tensor([1, 1])]; + tensor value_31_groups_0 = const()[name = tensor("value_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295107456))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295893952))), name = tensor("layers_15_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_31_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_31_dilations_0, groups = value_31_groups_0, pad = value_31_pad_0, pad_type = value_31_pad_type_0, strides = value_31_strides_0, weight = layers_15_self_attn_v_proj_weight_to_fp16_palettized, x = obj_63_cast_fp16)[name = tensor("value_31_cast_fp16")]; + tensor var_4468_to_fp16 = const()[name = tensor("op_4468_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295894144)))]; + tensor query_63_cast_fp16 = add(x = query_61_cast_fp16, y = var_4468_to_fp16)[name = tensor("query_63_cast_fp16")]; + tensor var_4471_to_fp16 = const()[name = tensor("op_4471_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295896256)))]; + tensor q_with_bias_v_31_cast_fp16 = add(x = query_61_cast_fp16, y = var_4471_to_fp16)[name = tensor("q_with_bias_v_31_cast_fp16")]; + tensor p_31_pad_type_0 = const()[name = tensor("p_31_pad_type_0"), val = tensor("valid")]; + tensor p_31_strides_0 = const()[name = tensor("p_31_strides_0"), val = tensor([1, 1])]; + tensor p_31_pad_0 = const()[name = tensor("p_31_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_31_dilations_0 = const()[name = tensor("p_31_dilations_0"), val = tensor([1, 1])]; + tensor p_31_groups_0 = const()[name = tensor("p_31_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295898368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296684864))), name = tensor("layers_15_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_31_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_31_dilations_0, groups = p_31_groups_0, pad = p_31_pad_0, pad_type = p_31_pad_type_0, strides = p_31_strides_0, weight = layers_15_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_31_cast_fp16")]; + tensor var_4482 = const()[name = tensor("op_4482"), val = tensor([1, 8, 128, 188])]; + tensor var_4483_cast_fp16 = reshape(shape = var_4482, x = q_with_bias_v_31_cast_fp16)[name = tensor("op_4483_cast_fp16")]; + tensor var_4484 = const()[name = tensor("op_4484"), val = tensor([1, 8, 128, -1])]; + tensor var_4485_cast_fp16 = reshape(shape = var_4484, x = p_31_cast_fp16)[name = tensor("op_4485_cast_fp16")]; + tensor matrix_bd_121_transpose_x_0 = const()[name = tensor("matrix_bd_121_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_121_transpose_y_0 = const()[name = tensor("matrix_bd_121_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_121_cast_fp16 = matmul(transpose_x = matrix_bd_121_transpose_x_0, transpose_y = matrix_bd_121_transpose_y_0, x = var_4483_cast_fp16, y = var_4485_cast_fp16)[name = tensor("matrix_bd_121_cast_fp16")]; + tensor matrix_bd_123_pad_0 = const()[name = tensor("matrix_bd_123_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_123_mode_0 = const()[name = tensor("matrix_bd_123_mode_0"), val = tensor("constant")]; + tensor const_175_to_fp16 = const()[name = tensor("const_175_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_123_cast_fp16 = pad(constant_val = const_175_to_fp16, mode = matrix_bd_123_mode_0, pad = matrix_bd_123_pad_0, x = matrix_bd_121_cast_fp16)[name = tensor("matrix_bd_123_cast_fp16")]; + tensor var_4494 = const()[name = tensor("op_4494"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_125_cast_fp16 = reshape(shape = var_4494, x = matrix_bd_123_cast_fp16)[name = tensor("matrix_bd_125_cast_fp16")]; + tensor var_4498_begin_0 = const()[name = tensor("op_4498_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4498_end_0 = const()[name = tensor("op_4498_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_4498_end_mask_0 = const()[name = tensor("op_4498_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4498_cast_fp16 = slice_by_index(begin = var_4498_begin_0, end = var_4498_end_0, end_mask = var_4498_end_mask_0, x = matrix_bd_125_cast_fp16)[name = tensor("op_4498_cast_fp16")]; + tensor var_4499 = const()[name = tensor("op_4499"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_127_cast_fp16 = reshape(shape = var_4499, x = var_4498_cast_fp16)[name = tensor("matrix_bd_127_cast_fp16")]; + tensor var_4504_begin_0 = const()[name = tensor("op_4504_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4504_end_0 = const()[name = tensor("op_4504_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_4504_end_mask_0 = const()[name = tensor("op_4504_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_4504_cast_fp16 = slice_by_index(begin = var_4504_begin_0, end = var_4504_end_0, end_mask = var_4504_end_mask_0, x = matrix_bd_127_cast_fp16)[name = tensor("op_4504_cast_fp16")]; + tensor var_4505_to_fp16 = const()[name = tensor("op_4505_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_31_cast_fp16 = mul(x = var_4504_cast_fp16, y = var_4505_to_fp16)[name = tensor("qk_mask_31_cast_fp16")]; + tensor var_4509 = const()[name = tensor("op_4509"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_4509, x = query_63_cast_fp16)[name = tensor("mh_q_31_cast_fp16")]; + tensor var_4511_to_fp16 = const()[name = tensor("op_4511_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_4512_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_4511_to_fp16)[name = tensor("op_4512_cast_fp16")]; + tensor var_4515 = const()[name = tensor("op_4515"), val = tensor([1, 8, 128, 188])]; + tensor var_4516_cast_fp16 = reshape(shape = var_4515, x = key_31_cast_fp16)[name = tensor("op_4516_cast_fp16")]; + tensor mh_w_61_transpose_x_0 = const()[name = tensor("mh_w_61_transpose_x_0"), val = tensor(true)]; + tensor mh_w_61_transpose_y_0 = const()[name = tensor("mh_w_61_transpose_y_0"), val = tensor(false)]; + tensor mh_w_61_cast_fp16 = matmul(transpose_x = mh_w_61_transpose_x_0, transpose_y = mh_w_61_transpose_y_0, x = var_4512_cast_fp16, y = var_4516_cast_fp16)[name = tensor("mh_w_61_cast_fp16")]; + tensor mh_w_63_cast_fp16 = add(x = mh_w_61_cast_fp16, y = qk_mask_31_cast_fp16)[name = tensor("mh_w_63_cast_fp16")]; + tensor var_4520_cast_fp16 = softmax(axis = var_4361, x = mh_w_63_cast_fp16)[name = tensor("op_4520_cast_fp16")]; + tensor var_4521 = const()[name = tensor("op_4521"), val = tensor([1, 8, 128, 188])]; + tensor var_4522_cast_fp16 = reshape(shape = var_4521, x = value_31_cast_fp16)[name = tensor("op_4522_cast_fp16")]; + tensor attn_31_transpose_x_0 = const()[name = tensor("attn_31_transpose_x_0"), val = tensor(false)]; + tensor attn_31_transpose_y_0 = const()[name = tensor("attn_31_transpose_y_0"), val = tensor(true)]; + tensor attn_31_cast_fp16 = matmul(transpose_x = attn_31_transpose_x_0, transpose_y = attn_31_transpose_y_0, x = var_4522_cast_fp16, y = var_4520_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_4525 = const()[name = tensor("op_4525"), val = tensor([1, 1024, 1, 188])]; + tensor input_413_cast_fp16 = reshape(shape = var_4525, x = attn_31_cast_fp16)[name = tensor("input_413_cast_fp16")]; + tensor obj_65_pad_type_0 = const()[name = tensor("obj_65_pad_type_0"), val = tensor("valid")]; + tensor obj_65_strides_0 = const()[name = tensor("obj_65_strides_0"), val = tensor([1, 1])]; + tensor obj_65_pad_0 = const()[name = tensor("obj_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_65_dilations_0 = const()[name = tensor("obj_65_dilations_0"), val = tensor([1, 1])]; + tensor obj_65_groups_0 = const()[name = tensor("obj_65_groups_0"), val = tensor(1)]; + tensor layers_15_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(296685056))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297471552))), name = tensor("layers_15_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_65_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_65_dilations_0, groups = obj_65_groups_0, pad = obj_65_pad_0, pad_type = obj_65_pad_type_0, strides = obj_65_strides_0, weight = layers_15_self_attn_o_proj_weight_to_fp16_palettized, x = input_413_cast_fp16)[name = tensor("obj_65_cast_fp16")]; + tensor inputs_155_cast_fp16 = add(x = inputs_153_cast_fp16, y = obj_65_cast_fp16)[name = tensor("inputs_155_cast_fp16")]; + tensor out_155_axes_0 = const()[name = tensor("out_155_axes_0"), val = tensor([1])]; + tensor var_4543_to_fp16 = const()[name = tensor("op_4543_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_155_cast_fp16 = layer_norm(axes = out_155_axes_0, epsilon = var_4543_to_fp16, x = inputs_155_cast_fp16)[name = tensor("out_155_cast_fp16")]; + tensor input_415_gamma_0_to_fp16 = const()[name = tensor("input_415_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297471744)))]; + tensor input_415_beta_0_to_fp16 = const()[name = tensor("input_415_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297473856)))]; + tensor input_415_epsilon_0_to_fp16 = const()[name = tensor("input_415_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_415_cast_fp16 = batch_norm(beta = input_415_beta_0_to_fp16, epsilon = input_415_epsilon_0_to_fp16, gamma = input_415_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor("input_415_cast_fp16")]; + tensor input_417_pad_type_0 = const()[name = tensor("input_417_pad_type_0"), val = tensor("valid")]; + tensor input_417_strides_0 = const()[name = tensor("input_417_strides_0"), val = tensor([1, 1])]; + tensor input_417_pad_0 = const()[name = tensor("input_417_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_417_dilations_0 = const()[name = tensor("input_417_dilations_0"), val = tensor([1, 1])]; + tensor input_417_groups_0 = const()[name = tensor("input_417_groups_0"), val = tensor(1)]; + tensor layers_15_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(297475968))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299048896))), name = tensor("layers_15_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_417_cast_fp16 = conv(dilations = input_417_dilations_0, groups = input_417_groups_0, pad = input_417_pad_0, pad_type = input_417_pad_type_0, strides = input_417_strides_0, weight = layers_15_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_415_cast_fp16)[name = tensor("input_417_cast_fp16")]; + tensor input_419_split_num_splits_0 = const()[name = tensor("input_419_split_num_splits_0"), val = tensor(2)]; + tensor input_419_split_axis_0 = const()[name = tensor("input_419_split_axis_0"), val = tensor(1)]; + tensor input_419_split_cast_fp16_0, tensor input_419_split_cast_fp16_1 = split(axis = input_419_split_axis_0, num_splits = input_419_split_num_splits_0, x = input_417_cast_fp16)[name = tensor("input_419_split_cast_fp16")]; + tensor input_419_split_1_sigmoid_cast_fp16 = sigmoid(x = input_419_split_cast_fp16_1)[name = tensor("input_419_split_1_sigmoid_cast_fp16")]; + tensor input_419_cast_fp16 = mul(x = input_419_split_cast_fp16_0, y = input_419_split_1_sigmoid_cast_fp16)[name = tensor("input_419_cast_fp16")]; + tensor input_421_pad_type_0 = const()[name = tensor("input_421_pad_type_0"), val = tensor("custom")]; + tensor input_421_pad_0 = const()[name = tensor("input_421_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_421_groups_0 = const()[name = tensor("input_421_groups_0"), val = tensor(1024)]; + tensor input_421_strides_0 = const()[name = tensor("input_421_strides_0"), val = tensor([1, 1])]; + tensor input_421_dilations_0 = const()[name = tensor("input_421_dilations_0"), val = tensor([1, 1])]; + tensor const_298_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299049088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299056064))), name = tensor("const_298_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_299_to_fp16 = const()[name = tensor("const_299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299056256)))]; + tensor input_423_cast_fp16 = conv(bias = const_299_to_fp16, dilations = input_421_dilations_0, groups = input_421_groups_0, pad = input_421_pad_0, pad_type = input_421_pad_type_0, strides = input_421_strides_0, weight = const_298_to_fp16_palettized, x = input_419_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor input_425_cast_fp16 = silu(x = input_423_cast_fp16)[name = tensor("input_425_cast_fp16")]; + tensor x_95_pad_type_0 = const()[name = tensor("x_95_pad_type_0"), val = tensor("valid")]; + tensor x_95_strides_0 = const()[name = tensor("x_95_strides_0"), val = tensor([1, 1])]; + tensor x_95_pad_0 = const()[name = tensor("x_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_95_dilations_0 = const()[name = tensor("x_95_dilations_0"), val = tensor([1, 1])]; + tensor x_95_groups_0 = const()[name = tensor("x_95_groups_0"), val = tensor(1)]; + tensor layers_15_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299058368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299844864))), name = tensor("layers_15_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_95_cast_fp16 = conv(dilations = x_95_dilations_0, groups = x_95_groups_0, pad = x_95_pad_0, pad_type = x_95_pad_type_0, strides = x_95_strides_0, weight = layers_15_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_425_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor inputs_157_cast_fp16 = add(x = inputs_155_cast_fp16, y = x_95_cast_fp16)[name = tensor("inputs_157_cast_fp16")]; + tensor out_157_axes_0 = const()[name = tensor("out_157_axes_0"), val = tensor([1])]; + tensor var_4591_to_fp16 = const()[name = tensor("op_4591_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_157_cast_fp16 = layer_norm(axes = out_157_axes_0, epsilon = var_4591_to_fp16, x = inputs_157_cast_fp16)[name = tensor("out_157_cast_fp16")]; + tensor input_427_gamma_0_to_fp16 = const()[name = tensor("input_427_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299845056)))]; + tensor input_427_beta_0_to_fp16 = const()[name = tensor("input_427_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299847168)))]; + tensor input_427_epsilon_0_to_fp16 = const()[name = tensor("input_427_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_427_cast_fp16 = batch_norm(beta = input_427_beta_0_to_fp16, epsilon = input_427_epsilon_0_to_fp16, gamma = input_427_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor("input_427_cast_fp16")]; + tensor input_429_pad_type_0 = const()[name = tensor("input_429_pad_type_0"), val = tensor("valid")]; + tensor input_429_strides_0 = const()[name = tensor("input_429_strides_0"), val = tensor([1, 1])]; + tensor input_429_pad_0 = const()[name = tensor("input_429_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_429_dilations_0 = const()[name = tensor("input_429_dilations_0"), val = tensor([1, 1])]; + tensor input_429_groups_0 = const()[name = tensor("input_429_groups_0"), val = tensor(1)]; + tensor layers_15_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(299849280))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302995072))), name = tensor("layers_15_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_429_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_429_dilations_0, groups = input_429_groups_0, pad = input_429_pad_0, pad_type = input_429_pad_type_0, strides = input_429_strides_0, weight = layers_15_feed_forward2_fc1_weight_to_fp16_palettized, x = input_427_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor input_431_cast_fp16 = silu(x = input_429_cast_fp16)[name = tensor("input_431_cast_fp16")]; + tensor x_97_pad_type_0 = const()[name = tensor("x_97_pad_type_0"), val = tensor("valid")]; + tensor x_97_strides_0 = const()[name = tensor("x_97_strides_0"), val = tensor([1, 1])]; + tensor x_97_pad_0 = const()[name = tensor("x_97_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_97_dilations_0 = const()[name = tensor("x_97_dilations_0"), val = tensor([1, 1])]; + tensor x_97_groups_0 = const()[name = tensor("x_97_groups_0"), val = tensor(1)]; + tensor op_4619_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302995264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306141056))), name = tensor("op_4619_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4619_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_97_dilations_0, groups = x_97_groups_0, pad = x_97_pad_0, pad_type = x_97_pad_type_0, strides = x_97_strides_0, weight = op_4619_weight_0_to_fp16_palettized, x = input_431_cast_fp16)[name = tensor("op_4619_cast_fp16")]; + tensor inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = var_4619_cast_fp16)[name = tensor("inputs_159_cast_fp16")]; + tensor out_159_axes_0 = const()[name = tensor("out_159_axes_0"), val = tensor([1])]; + tensor var_4629_to_fp16 = const()[name = tensor("op_4629_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_159_cast_fp16 = layer_norm(axes = out_159_axes_0, epsilon = var_4629_to_fp16, x = inputs_159_cast_fp16)[name = tensor("out_159_cast_fp16")]; + tensor inputs_161_gamma_0_to_fp16 = const()[name = tensor("inputs_161_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306141248)))]; + tensor inputs_161_beta_0_to_fp16 = const()[name = tensor("inputs_161_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306143360)))]; + tensor inputs_161_epsilon_0_to_fp16 = const()[name = tensor("inputs_161_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_161_cast_fp16 = batch_norm(beta = inputs_161_beta_0_to_fp16, epsilon = inputs_161_epsilon_0_to_fp16, gamma = inputs_161_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor("inputs_161_cast_fp16")]; + tensor var_4643 = const()[name = tensor("op_4643"), val = tensor(3)]; + tensor out_161_axes_0 = const()[name = tensor("out_161_axes_0"), val = tensor([1])]; + tensor var_4674_to_fp16 = const()[name = tensor("op_4674_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_161_cast_fp16 = layer_norm(axes = out_161_axes_0, epsilon = var_4674_to_fp16, x = inputs_161_cast_fp16)[name = tensor("out_161_cast_fp16")]; + tensor input_433_gamma_0_to_fp16 = const()[name = tensor("input_433_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306145472)))]; + tensor input_433_beta_0_to_fp16 = const()[name = tensor("input_433_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306147584)))]; + tensor input_433_epsilon_0_to_fp16 = const()[name = tensor("input_433_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_433_cast_fp16 = batch_norm(beta = input_433_beta_0_to_fp16, epsilon = input_433_epsilon_0_to_fp16, gamma = input_433_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor("input_433_cast_fp16")]; + tensor input_435_pad_type_0 = const()[name = tensor("input_435_pad_type_0"), val = tensor("valid")]; + tensor input_435_strides_0 = const()[name = tensor("input_435_strides_0"), val = tensor([1, 1])]; + tensor input_435_pad_0 = const()[name = tensor("input_435_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_435_dilations_0 = const()[name = tensor("input_435_dilations_0"), val = tensor([1, 1])]; + tensor input_435_groups_0 = const()[name = tensor("input_435_groups_0"), val = tensor(1)]; + tensor layers_16_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(306149696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309295488))), name = tensor("layers_16_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_435_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_435_dilations_0, groups = input_435_groups_0, pad = input_435_pad_0, pad_type = input_435_pad_type_0, strides = input_435_strides_0, weight = layers_16_feed_forward1_fc1_weight_to_fp16_palettized, x = input_433_cast_fp16)[name = tensor("input_435_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = input_435_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor x_99_pad_type_0 = const()[name = tensor("x_99_pad_type_0"), val = tensor("valid")]; + tensor x_99_strides_0 = const()[name = tensor("x_99_strides_0"), val = tensor([1, 1])]; + tensor x_99_pad_0 = const()[name = tensor("x_99_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_99_dilations_0 = const()[name = tensor("x_99_dilations_0"), val = tensor([1, 1])]; + tensor x_99_groups_0 = const()[name = tensor("x_99_groups_0"), val = tensor(1)]; + tensor op_4702_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309295680))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312441472))), name = tensor("op_4702_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4702_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = op_4702_weight_0_to_fp16_palettized, x = input_437_cast_fp16)[name = tensor("op_4702_cast_fp16")]; + tensor inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = var_4702_cast_fp16)[name = tensor("inputs_163_cast_fp16")]; + tensor out_163_axes_0 = const()[name = tensor("out_163_axes_0"), val = tensor([1])]; + tensor var_4712_to_fp16 = const()[name = tensor("op_4712_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_163_cast_fp16 = layer_norm(axes = out_163_axes_0, epsilon = var_4712_to_fp16, x = inputs_163_cast_fp16)[name = tensor("out_163_cast_fp16")]; + tensor obj_67_gamma_0_to_fp16 = const()[name = tensor("obj_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312441664)))]; + tensor obj_67_beta_0_to_fp16 = const()[name = tensor("obj_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312443776)))]; + tensor obj_67_epsilon_0_to_fp16 = const()[name = tensor("obj_67_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_67_cast_fp16 = batch_norm(beta = obj_67_beta_0_to_fp16, epsilon = obj_67_epsilon_0_to_fp16, gamma = obj_67_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_163_cast_fp16)[name = tensor("obj_67_cast_fp16")]; + tensor query_65_pad_type_0 = const()[name = tensor("query_65_pad_type_0"), val = tensor("valid")]; + tensor query_65_strides_0 = const()[name = tensor("query_65_strides_0"), val = tensor([1, 1])]; + tensor query_65_pad_0 = const()[name = tensor("query_65_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_65_dilations_0 = const()[name = tensor("query_65_dilations_0"), val = tensor([1, 1])]; + tensor query_65_groups_0 = const()[name = tensor("query_65_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(312445888))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313232384))), name = tensor("layers_16_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_65_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_65_dilations_0, groups = query_65_groups_0, pad = query_65_pad_0, pad_type = query_65_pad_type_0, strides = query_65_strides_0, weight = layers_16_self_attn_q_proj_weight_to_fp16_palettized, x = obj_67_cast_fp16)[name = tensor("query_65_cast_fp16")]; + tensor key_33_pad_type_0 = const()[name = tensor("key_33_pad_type_0"), val = tensor("valid")]; + tensor key_33_strides_0 = const()[name = tensor("key_33_strides_0"), val = tensor([1, 1])]; + tensor key_33_pad_0 = const()[name = tensor("key_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_33_dilations_0 = const()[name = tensor("key_33_dilations_0"), val = tensor([1, 1])]; + tensor key_33_groups_0 = const()[name = tensor("key_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(313232576))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314019072))), name = tensor("layers_16_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_33_cast_fp16 = conv(dilations = key_33_dilations_0, groups = key_33_groups_0, pad = key_33_pad_0, pad_type = key_33_pad_type_0, strides = key_33_strides_0, weight = layers_16_self_attn_k_proj_weight_to_fp16_palettized, x = obj_67_cast_fp16)[name = tensor("key_33_cast_fp16")]; + tensor value_33_pad_type_0 = const()[name = tensor("value_33_pad_type_0"), val = tensor("valid")]; + tensor value_33_strides_0 = const()[name = tensor("value_33_strides_0"), val = tensor([1, 1])]; + tensor value_33_pad_0 = const()[name = tensor("value_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_33_dilations_0 = const()[name = tensor("value_33_dilations_0"), val = tensor([1, 1])]; + tensor value_33_groups_0 = const()[name = tensor("value_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314019264))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314805760))), name = tensor("layers_16_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_33_dilations_0, groups = value_33_groups_0, pad = value_33_pad_0, pad_type = value_33_pad_type_0, strides = value_33_strides_0, weight = layers_16_self_attn_v_proj_weight_to_fp16_palettized, x = obj_67_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_4750_to_fp16 = const()[name = tensor("op_4750_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314805952)))]; + tensor query_67_cast_fp16 = add(x = query_65_cast_fp16, y = var_4750_to_fp16)[name = tensor("query_67_cast_fp16")]; + tensor var_4753_to_fp16 = const()[name = tensor("op_4753_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314808064)))]; + tensor q_with_bias_v_33_cast_fp16 = add(x = query_65_cast_fp16, y = var_4753_to_fp16)[name = tensor("q_with_bias_v_33_cast_fp16")]; + tensor p_33_pad_type_0 = const()[name = tensor("p_33_pad_type_0"), val = tensor("valid")]; + tensor p_33_strides_0 = const()[name = tensor("p_33_strides_0"), val = tensor([1, 1])]; + tensor p_33_pad_0 = const()[name = tensor("p_33_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_33_dilations_0 = const()[name = tensor("p_33_dilations_0"), val = tensor([1, 1])]; + tensor p_33_groups_0 = const()[name = tensor("p_33_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(314810176))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315596672))), name = tensor("layers_16_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_33_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_33_dilations_0, groups = p_33_groups_0, pad = p_33_pad_0, pad_type = p_33_pad_type_0, strides = p_33_strides_0, weight = layers_16_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_33_cast_fp16")]; + tensor var_4764 = const()[name = tensor("op_4764"), val = tensor([1, 8, 128, 188])]; + tensor var_4765_cast_fp16 = reshape(shape = var_4764, x = q_with_bias_v_33_cast_fp16)[name = tensor("op_4765_cast_fp16")]; + tensor var_4766 = const()[name = tensor("op_4766"), val = tensor([1, 8, 128, -1])]; + tensor var_4767_cast_fp16 = reshape(shape = var_4766, x = p_33_cast_fp16)[name = tensor("op_4767_cast_fp16")]; + tensor matrix_bd_129_transpose_x_0 = const()[name = tensor("matrix_bd_129_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_129_transpose_y_0 = const()[name = tensor("matrix_bd_129_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_129_cast_fp16 = matmul(transpose_x = matrix_bd_129_transpose_x_0, transpose_y = matrix_bd_129_transpose_y_0, x = var_4765_cast_fp16, y = var_4767_cast_fp16)[name = tensor("matrix_bd_129_cast_fp16")]; + tensor matrix_bd_131_pad_0 = const()[name = tensor("matrix_bd_131_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_131_mode_0 = const()[name = tensor("matrix_bd_131_mode_0"), val = tensor("constant")]; + tensor const_186_to_fp16 = const()[name = tensor("const_186_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_131_cast_fp16 = pad(constant_val = const_186_to_fp16, mode = matrix_bd_131_mode_0, pad = matrix_bd_131_pad_0, x = matrix_bd_129_cast_fp16)[name = tensor("matrix_bd_131_cast_fp16")]; + tensor var_4776 = const()[name = tensor("op_4776"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_133_cast_fp16 = reshape(shape = var_4776, x = matrix_bd_131_cast_fp16)[name = tensor("matrix_bd_133_cast_fp16")]; + tensor var_4780_begin_0 = const()[name = tensor("op_4780_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4780_end_0 = const()[name = tensor("op_4780_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_4780_end_mask_0 = const()[name = tensor("op_4780_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4780_cast_fp16 = slice_by_index(begin = var_4780_begin_0, end = var_4780_end_0, end_mask = var_4780_end_mask_0, x = matrix_bd_133_cast_fp16)[name = tensor("op_4780_cast_fp16")]; + tensor var_4781 = const()[name = tensor("op_4781"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_135_cast_fp16 = reshape(shape = var_4781, x = var_4780_cast_fp16)[name = tensor("matrix_bd_135_cast_fp16")]; + tensor var_4786_begin_0 = const()[name = tensor("op_4786_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4786_end_0 = const()[name = tensor("op_4786_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_4786_end_mask_0 = const()[name = tensor("op_4786_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_4786_cast_fp16 = slice_by_index(begin = var_4786_begin_0, end = var_4786_end_0, end_mask = var_4786_end_mask_0, x = matrix_bd_135_cast_fp16)[name = tensor("op_4786_cast_fp16")]; + tensor var_4787_to_fp16 = const()[name = tensor("op_4787_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_33_cast_fp16 = mul(x = var_4786_cast_fp16, y = var_4787_to_fp16)[name = tensor("qk_mask_33_cast_fp16")]; + tensor var_4791 = const()[name = tensor("op_4791"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_4791, x = query_67_cast_fp16)[name = tensor("mh_q_33_cast_fp16")]; + tensor var_4793_to_fp16 = const()[name = tensor("op_4793_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_4794_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_4793_to_fp16)[name = tensor("op_4794_cast_fp16")]; + tensor var_4797 = const()[name = tensor("op_4797"), val = tensor([1, 8, 128, 188])]; + tensor var_4798_cast_fp16 = reshape(shape = var_4797, x = key_33_cast_fp16)[name = tensor("op_4798_cast_fp16")]; + tensor mh_w_65_transpose_x_0 = const()[name = tensor("mh_w_65_transpose_x_0"), val = tensor(true)]; + tensor mh_w_65_transpose_y_0 = const()[name = tensor("mh_w_65_transpose_y_0"), val = tensor(false)]; + tensor mh_w_65_cast_fp16 = matmul(transpose_x = mh_w_65_transpose_x_0, transpose_y = mh_w_65_transpose_y_0, x = var_4794_cast_fp16, y = var_4798_cast_fp16)[name = tensor("mh_w_65_cast_fp16")]; + tensor mh_w_67_cast_fp16 = add(x = mh_w_65_cast_fp16, y = qk_mask_33_cast_fp16)[name = tensor("mh_w_67_cast_fp16")]; + tensor var_4802_cast_fp16 = softmax(axis = var_4643, x = mh_w_67_cast_fp16)[name = tensor("op_4802_cast_fp16")]; + tensor var_4803 = const()[name = tensor("op_4803"), val = tensor([1, 8, 128, 188])]; + tensor var_4804_cast_fp16 = reshape(shape = var_4803, x = value_33_cast_fp16)[name = tensor("op_4804_cast_fp16")]; + tensor attn_33_transpose_x_0 = const()[name = tensor("attn_33_transpose_x_0"), val = tensor(false)]; + tensor attn_33_transpose_y_0 = const()[name = tensor("attn_33_transpose_y_0"), val = tensor(true)]; + tensor attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = var_4804_cast_fp16, y = var_4802_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4807 = const()[name = tensor("op_4807"), val = tensor([1, 1024, 1, 188])]; + tensor input_439_cast_fp16 = reshape(shape = var_4807, x = attn_33_cast_fp16)[name = tensor("input_439_cast_fp16")]; + tensor obj_69_pad_type_0 = const()[name = tensor("obj_69_pad_type_0"), val = tensor("valid")]; + tensor obj_69_strides_0 = const()[name = tensor("obj_69_strides_0"), val = tensor([1, 1])]; + tensor obj_69_pad_0 = const()[name = tensor("obj_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_69_dilations_0 = const()[name = tensor("obj_69_dilations_0"), val = tensor([1, 1])]; + tensor obj_69_groups_0 = const()[name = tensor("obj_69_groups_0"), val = tensor(1)]; + tensor layers_16_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(315596864))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316383360))), name = tensor("layers_16_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_69_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_69_dilations_0, groups = obj_69_groups_0, pad = obj_69_pad_0, pad_type = obj_69_pad_type_0, strides = obj_69_strides_0, weight = layers_16_self_attn_o_proj_weight_to_fp16_palettized, x = input_439_cast_fp16)[name = tensor("obj_69_cast_fp16")]; + tensor inputs_165_cast_fp16 = add(x = inputs_163_cast_fp16, y = obj_69_cast_fp16)[name = tensor("inputs_165_cast_fp16")]; + tensor out_165_axes_0 = const()[name = tensor("out_165_axes_0"), val = tensor([1])]; + tensor var_4825_to_fp16 = const()[name = tensor("op_4825_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_165_cast_fp16 = layer_norm(axes = out_165_axes_0, epsilon = var_4825_to_fp16, x = inputs_165_cast_fp16)[name = tensor("out_165_cast_fp16")]; + tensor input_441_gamma_0_to_fp16 = const()[name = tensor("input_441_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316383552)))]; + tensor input_441_beta_0_to_fp16 = const()[name = tensor("input_441_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316385664)))]; + tensor input_441_epsilon_0_to_fp16 = const()[name = tensor("input_441_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_441_cast_fp16 = batch_norm(beta = input_441_beta_0_to_fp16, epsilon = input_441_epsilon_0_to_fp16, gamma = input_441_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor("input_441_cast_fp16")]; + tensor input_443_pad_type_0 = const()[name = tensor("input_443_pad_type_0"), val = tensor("valid")]; + tensor input_443_strides_0 = const()[name = tensor("input_443_strides_0"), val = tensor([1, 1])]; + tensor input_443_pad_0 = const()[name = tensor("input_443_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_443_dilations_0 = const()[name = tensor("input_443_dilations_0"), val = tensor([1, 1])]; + tensor input_443_groups_0 = const()[name = tensor("input_443_groups_0"), val = tensor(1)]; + tensor layers_16_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(316387776))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317960704))), name = tensor("layers_16_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_443_cast_fp16 = conv(dilations = input_443_dilations_0, groups = input_443_groups_0, pad = input_443_pad_0, pad_type = input_443_pad_type_0, strides = input_443_strides_0, weight = layers_16_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_441_cast_fp16)[name = tensor("input_443_cast_fp16")]; + tensor input_445_split_num_splits_0 = const()[name = tensor("input_445_split_num_splits_0"), val = tensor(2)]; + tensor input_445_split_axis_0 = const()[name = tensor("input_445_split_axis_0"), val = tensor(1)]; + tensor input_445_split_cast_fp16_0, tensor input_445_split_cast_fp16_1 = split(axis = input_445_split_axis_0, num_splits = input_445_split_num_splits_0, x = input_443_cast_fp16)[name = tensor("input_445_split_cast_fp16")]; + tensor input_445_split_1_sigmoid_cast_fp16 = sigmoid(x = input_445_split_cast_fp16_1)[name = tensor("input_445_split_1_sigmoid_cast_fp16")]; + tensor input_445_cast_fp16 = mul(x = input_445_split_cast_fp16_0, y = input_445_split_1_sigmoid_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor input_447_pad_type_0 = const()[name = tensor("input_447_pad_type_0"), val = tensor("custom")]; + tensor input_447_pad_0 = const()[name = tensor("input_447_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_447_groups_0 = const()[name = tensor("input_447_groups_0"), val = tensor(1024)]; + tensor input_447_strides_0 = const()[name = tensor("input_447_strides_0"), val = tensor([1, 1])]; + tensor input_447_dilations_0 = const()[name = tensor("input_447_dilations_0"), val = tensor([1, 1])]; + tensor const_300_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317960896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317967872))), name = tensor("const_300_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_301_to_fp16 = const()[name = tensor("const_301_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317968064)))]; + tensor input_449_cast_fp16 = conv(bias = const_301_to_fp16, dilations = input_447_dilations_0, groups = input_447_groups_0, pad = input_447_pad_0, pad_type = input_447_pad_type_0, strides = input_447_strides_0, weight = const_300_to_fp16_palettized, x = input_445_cast_fp16)[name = tensor("input_449_cast_fp16")]; + tensor input_451_cast_fp16 = silu(x = input_449_cast_fp16)[name = tensor("input_451_cast_fp16")]; + tensor x_101_pad_type_0 = const()[name = tensor("x_101_pad_type_0"), val = tensor("valid")]; + tensor x_101_strides_0 = const()[name = tensor("x_101_strides_0"), val = tensor([1, 1])]; + tensor x_101_pad_0 = const()[name = tensor("x_101_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_101_dilations_0 = const()[name = tensor("x_101_dilations_0"), val = tensor([1, 1])]; + tensor x_101_groups_0 = const()[name = tensor("x_101_groups_0"), val = tensor(1)]; + tensor layers_16_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(317970176))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318756672))), name = tensor("layers_16_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_101_cast_fp16 = conv(dilations = x_101_dilations_0, groups = x_101_groups_0, pad = x_101_pad_0, pad_type = x_101_pad_type_0, strides = x_101_strides_0, weight = layers_16_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_451_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor inputs_167_cast_fp16 = add(x = inputs_165_cast_fp16, y = x_101_cast_fp16)[name = tensor("inputs_167_cast_fp16")]; + tensor out_167_axes_0 = const()[name = tensor("out_167_axes_0"), val = tensor([1])]; + tensor var_4873_to_fp16 = const()[name = tensor("op_4873_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_167_cast_fp16 = layer_norm(axes = out_167_axes_0, epsilon = var_4873_to_fp16, x = inputs_167_cast_fp16)[name = tensor("out_167_cast_fp16")]; + tensor input_453_gamma_0_to_fp16 = const()[name = tensor("input_453_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318756864)))]; + tensor input_453_beta_0_to_fp16 = const()[name = tensor("input_453_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318758976)))]; + tensor input_453_epsilon_0_to_fp16 = const()[name = tensor("input_453_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_453_cast_fp16 = batch_norm(beta = input_453_beta_0_to_fp16, epsilon = input_453_epsilon_0_to_fp16, gamma = input_453_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor input_455_pad_type_0 = const()[name = tensor("input_455_pad_type_0"), val = tensor("valid")]; + tensor input_455_strides_0 = const()[name = tensor("input_455_strides_0"), val = tensor([1, 1])]; + tensor input_455_pad_0 = const()[name = tensor("input_455_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_455_dilations_0 = const()[name = tensor("input_455_dilations_0"), val = tensor([1, 1])]; + tensor input_455_groups_0 = const()[name = tensor("input_455_groups_0"), val = tensor(1)]; + tensor layers_16_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(318761088))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321906880))), name = tensor("layers_16_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_455_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_455_dilations_0, groups = input_455_groups_0, pad = input_455_pad_0, pad_type = input_455_pad_type_0, strides = input_455_strides_0, weight = layers_16_feed_forward2_fc1_weight_to_fp16_palettized, x = input_453_cast_fp16)[name = tensor("input_455_cast_fp16")]; + tensor input_457_cast_fp16 = silu(x = input_455_cast_fp16)[name = tensor("input_457_cast_fp16")]; + tensor x_103_pad_type_0 = const()[name = tensor("x_103_pad_type_0"), val = tensor("valid")]; + tensor x_103_strides_0 = const()[name = tensor("x_103_strides_0"), val = tensor([1, 1])]; + tensor x_103_pad_0 = const()[name = tensor("x_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_103_dilations_0 = const()[name = tensor("x_103_dilations_0"), val = tensor([1, 1])]; + tensor x_103_groups_0 = const()[name = tensor("x_103_groups_0"), val = tensor(1)]; + tensor op_4901_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(321907072))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325052864))), name = tensor("op_4901_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4901_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = op_4901_weight_0_to_fp16_palettized, x = input_457_cast_fp16)[name = tensor("op_4901_cast_fp16")]; + tensor inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = var_4901_cast_fp16)[name = tensor("inputs_169_cast_fp16")]; + tensor out_169_axes_0 = const()[name = tensor("out_169_axes_0"), val = tensor([1])]; + tensor var_4911_to_fp16 = const()[name = tensor("op_4911_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_169_cast_fp16 = layer_norm(axes = out_169_axes_0, epsilon = var_4911_to_fp16, x = inputs_169_cast_fp16)[name = tensor("out_169_cast_fp16")]; + tensor inputs_171_gamma_0_to_fp16 = const()[name = tensor("inputs_171_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325053056)))]; + tensor inputs_171_beta_0_to_fp16 = const()[name = tensor("inputs_171_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325055168)))]; + tensor inputs_171_epsilon_0_to_fp16 = const()[name = tensor("inputs_171_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_171_cast_fp16 = batch_norm(beta = inputs_171_beta_0_to_fp16, epsilon = inputs_171_epsilon_0_to_fp16, gamma = inputs_171_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor("inputs_171_cast_fp16")]; + tensor var_4925 = const()[name = tensor("op_4925"), val = tensor(3)]; + tensor out_171_axes_0 = const()[name = tensor("out_171_axes_0"), val = tensor([1])]; + tensor var_4956_to_fp16 = const()[name = tensor("op_4956_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_171_cast_fp16 = layer_norm(axes = out_171_axes_0, epsilon = var_4956_to_fp16, x = inputs_171_cast_fp16)[name = tensor("out_171_cast_fp16")]; + tensor input_459_gamma_0_to_fp16 = const()[name = tensor("input_459_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325057280)))]; + tensor input_459_beta_0_to_fp16 = const()[name = tensor("input_459_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325059392)))]; + tensor input_459_epsilon_0_to_fp16 = const()[name = tensor("input_459_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_459_cast_fp16 = batch_norm(beta = input_459_beta_0_to_fp16, epsilon = input_459_epsilon_0_to_fp16, gamma = input_459_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor("input_459_cast_fp16")]; + tensor input_461_pad_type_0 = const()[name = tensor("input_461_pad_type_0"), val = tensor("valid")]; + tensor input_461_strides_0 = const()[name = tensor("input_461_strides_0"), val = tensor([1, 1])]; + tensor input_461_pad_0 = const()[name = tensor("input_461_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_461_dilations_0 = const()[name = tensor("input_461_dilations_0"), val = tensor([1, 1])]; + tensor input_461_groups_0 = const()[name = tensor("input_461_groups_0"), val = tensor(1)]; + tensor layers_17_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(325061504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328207296))), name = tensor("layers_17_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_461_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_461_dilations_0, groups = input_461_groups_0, pad = input_461_pad_0, pad_type = input_461_pad_type_0, strides = input_461_strides_0, weight = layers_17_feed_forward1_fc1_weight_to_fp16_palettized, x = input_459_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor input_463_cast_fp16 = silu(x = input_461_cast_fp16)[name = tensor("input_463_cast_fp16")]; + tensor x_105_pad_type_0 = const()[name = tensor("x_105_pad_type_0"), val = tensor("valid")]; + tensor x_105_strides_0 = const()[name = tensor("x_105_strides_0"), val = tensor([1, 1])]; + tensor x_105_pad_0 = const()[name = tensor("x_105_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_105_dilations_0 = const()[name = tensor("x_105_dilations_0"), val = tensor([1, 1])]; + tensor x_105_groups_0 = const()[name = tensor("x_105_groups_0"), val = tensor(1)]; + tensor op_4984_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328207488))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331353280))), name = tensor("op_4984_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_4984_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_105_dilations_0, groups = x_105_groups_0, pad = x_105_pad_0, pad_type = x_105_pad_type_0, strides = x_105_strides_0, weight = op_4984_weight_0_to_fp16_palettized, x = input_463_cast_fp16)[name = tensor("op_4984_cast_fp16")]; + tensor inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = var_4984_cast_fp16)[name = tensor("inputs_173_cast_fp16")]; + tensor out_173_axes_0 = const()[name = tensor("out_173_axes_0"), val = tensor([1])]; + tensor var_4994_to_fp16 = const()[name = tensor("op_4994_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_173_cast_fp16 = layer_norm(axes = out_173_axes_0, epsilon = var_4994_to_fp16, x = inputs_173_cast_fp16)[name = tensor("out_173_cast_fp16")]; + tensor obj_71_gamma_0_to_fp16 = const()[name = tensor("obj_71_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331353472)))]; + tensor obj_71_beta_0_to_fp16 = const()[name = tensor("obj_71_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331355584)))]; + tensor obj_71_epsilon_0_to_fp16 = const()[name = tensor("obj_71_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_71_cast_fp16 = batch_norm(beta = obj_71_beta_0_to_fp16, epsilon = obj_71_epsilon_0_to_fp16, gamma = obj_71_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_173_cast_fp16)[name = tensor("obj_71_cast_fp16")]; + tensor query_69_pad_type_0 = const()[name = tensor("query_69_pad_type_0"), val = tensor("valid")]; + tensor query_69_strides_0 = const()[name = tensor("query_69_strides_0"), val = tensor([1, 1])]; + tensor query_69_pad_0 = const()[name = tensor("query_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_69_dilations_0 = const()[name = tensor("query_69_dilations_0"), val = tensor([1, 1])]; + tensor query_69_groups_0 = const()[name = tensor("query_69_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(331357696))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332144192))), name = tensor("layers_17_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_69_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_69_dilations_0, groups = query_69_groups_0, pad = query_69_pad_0, pad_type = query_69_pad_type_0, strides = query_69_strides_0, weight = layers_17_self_attn_q_proj_weight_to_fp16_palettized, x = obj_71_cast_fp16)[name = tensor("query_69_cast_fp16")]; + tensor key_35_pad_type_0 = const()[name = tensor("key_35_pad_type_0"), val = tensor("valid")]; + tensor key_35_strides_0 = const()[name = tensor("key_35_strides_0"), val = tensor([1, 1])]; + tensor key_35_pad_0 = const()[name = tensor("key_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_35_dilations_0 = const()[name = tensor("key_35_dilations_0"), val = tensor([1, 1])]; + tensor key_35_groups_0 = const()[name = tensor("key_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332144384))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332930880))), name = tensor("layers_17_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_35_cast_fp16 = conv(dilations = key_35_dilations_0, groups = key_35_groups_0, pad = key_35_pad_0, pad_type = key_35_pad_type_0, strides = key_35_strides_0, weight = layers_17_self_attn_k_proj_weight_to_fp16_palettized, x = obj_71_cast_fp16)[name = tensor("key_35_cast_fp16")]; + tensor value_35_pad_type_0 = const()[name = tensor("value_35_pad_type_0"), val = tensor("valid")]; + tensor value_35_strides_0 = const()[name = tensor("value_35_strides_0"), val = tensor([1, 1])]; + tensor value_35_pad_0 = const()[name = tensor("value_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_35_dilations_0 = const()[name = tensor("value_35_dilations_0"), val = tensor([1, 1])]; + tensor value_35_groups_0 = const()[name = tensor("value_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332931072))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333717568))), name = tensor("layers_17_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_35_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_35_dilations_0, groups = value_35_groups_0, pad = value_35_pad_0, pad_type = value_35_pad_type_0, strides = value_35_strides_0, weight = layers_17_self_attn_v_proj_weight_to_fp16_palettized, x = obj_71_cast_fp16)[name = tensor("value_35_cast_fp16")]; + tensor var_5032_to_fp16 = const()[name = tensor("op_5032_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333717760)))]; + tensor query_71_cast_fp16 = add(x = query_69_cast_fp16, y = var_5032_to_fp16)[name = tensor("query_71_cast_fp16")]; + tensor var_5035_to_fp16 = const()[name = tensor("op_5035_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333719872)))]; + tensor q_with_bias_v_35_cast_fp16 = add(x = query_69_cast_fp16, y = var_5035_to_fp16)[name = tensor("q_with_bias_v_35_cast_fp16")]; + tensor p_35_pad_type_0 = const()[name = tensor("p_35_pad_type_0"), val = tensor("valid")]; + tensor p_35_strides_0 = const()[name = tensor("p_35_strides_0"), val = tensor([1, 1])]; + tensor p_35_pad_0 = const()[name = tensor("p_35_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_35_dilations_0 = const()[name = tensor("p_35_dilations_0"), val = tensor([1, 1])]; + tensor p_35_groups_0 = const()[name = tensor("p_35_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(333721984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334508480))), name = tensor("layers_17_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_35_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_35_dilations_0, groups = p_35_groups_0, pad = p_35_pad_0, pad_type = p_35_pad_type_0, strides = p_35_strides_0, weight = layers_17_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_35_cast_fp16")]; + tensor var_5046 = const()[name = tensor("op_5046"), val = tensor([1, 8, 128, 188])]; + tensor var_5047_cast_fp16 = reshape(shape = var_5046, x = q_with_bias_v_35_cast_fp16)[name = tensor("op_5047_cast_fp16")]; + tensor var_5048 = const()[name = tensor("op_5048"), val = tensor([1, 8, 128, -1])]; + tensor var_5049_cast_fp16 = reshape(shape = var_5048, x = p_35_cast_fp16)[name = tensor("op_5049_cast_fp16")]; + tensor matrix_bd_137_transpose_x_0 = const()[name = tensor("matrix_bd_137_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_137_transpose_y_0 = const()[name = tensor("matrix_bd_137_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_137_cast_fp16 = matmul(transpose_x = matrix_bd_137_transpose_x_0, transpose_y = matrix_bd_137_transpose_y_0, x = var_5047_cast_fp16, y = var_5049_cast_fp16)[name = tensor("matrix_bd_137_cast_fp16")]; + tensor matrix_bd_139_pad_0 = const()[name = tensor("matrix_bd_139_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_139_mode_0 = const()[name = tensor("matrix_bd_139_mode_0"), val = tensor("constant")]; + tensor const_197_to_fp16 = const()[name = tensor("const_197_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_139_cast_fp16 = pad(constant_val = const_197_to_fp16, mode = matrix_bd_139_mode_0, pad = matrix_bd_139_pad_0, x = matrix_bd_137_cast_fp16)[name = tensor("matrix_bd_139_cast_fp16")]; + tensor var_5058 = const()[name = tensor("op_5058"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_141_cast_fp16 = reshape(shape = var_5058, x = matrix_bd_139_cast_fp16)[name = tensor("matrix_bd_141_cast_fp16")]; + tensor var_5062_begin_0 = const()[name = tensor("op_5062_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5062_end_0 = const()[name = tensor("op_5062_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5062_end_mask_0 = const()[name = tensor("op_5062_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5062_cast_fp16 = slice_by_index(begin = var_5062_begin_0, end = var_5062_end_0, end_mask = var_5062_end_mask_0, x = matrix_bd_141_cast_fp16)[name = tensor("op_5062_cast_fp16")]; + tensor var_5063 = const()[name = tensor("op_5063"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_143_cast_fp16 = reshape(shape = var_5063, x = var_5062_cast_fp16)[name = tensor("matrix_bd_143_cast_fp16")]; + tensor var_5068_begin_0 = const()[name = tensor("op_5068_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5068_end_0 = const()[name = tensor("op_5068_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5068_end_mask_0 = const()[name = tensor("op_5068_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5068_cast_fp16 = slice_by_index(begin = var_5068_begin_0, end = var_5068_end_0, end_mask = var_5068_end_mask_0, x = matrix_bd_143_cast_fp16)[name = tensor("op_5068_cast_fp16")]; + tensor var_5069_to_fp16 = const()[name = tensor("op_5069_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_35_cast_fp16 = mul(x = var_5068_cast_fp16, y = var_5069_to_fp16)[name = tensor("qk_mask_35_cast_fp16")]; + tensor var_5073 = const()[name = tensor("op_5073"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_5073, x = query_71_cast_fp16)[name = tensor("mh_q_35_cast_fp16")]; + tensor var_5075_to_fp16 = const()[name = tensor("op_5075_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5076_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_5075_to_fp16)[name = tensor("op_5076_cast_fp16")]; + tensor var_5079 = const()[name = tensor("op_5079"), val = tensor([1, 8, 128, 188])]; + tensor var_5080_cast_fp16 = reshape(shape = var_5079, x = key_35_cast_fp16)[name = tensor("op_5080_cast_fp16")]; + tensor mh_w_69_transpose_x_0 = const()[name = tensor("mh_w_69_transpose_x_0"), val = tensor(true)]; + tensor mh_w_69_transpose_y_0 = const()[name = tensor("mh_w_69_transpose_y_0"), val = tensor(false)]; + tensor mh_w_69_cast_fp16 = matmul(transpose_x = mh_w_69_transpose_x_0, transpose_y = mh_w_69_transpose_y_0, x = var_5076_cast_fp16, y = var_5080_cast_fp16)[name = tensor("mh_w_69_cast_fp16")]; + tensor mh_w_71_cast_fp16 = add(x = mh_w_69_cast_fp16, y = qk_mask_35_cast_fp16)[name = tensor("mh_w_71_cast_fp16")]; + tensor var_5084_cast_fp16 = softmax(axis = var_4925, x = mh_w_71_cast_fp16)[name = tensor("op_5084_cast_fp16")]; + tensor var_5085 = const()[name = tensor("op_5085"), val = tensor([1, 8, 128, 188])]; + tensor var_5086_cast_fp16 = reshape(shape = var_5085, x = value_35_cast_fp16)[name = tensor("op_5086_cast_fp16")]; + tensor attn_35_transpose_x_0 = const()[name = tensor("attn_35_transpose_x_0"), val = tensor(false)]; + tensor attn_35_transpose_y_0 = const()[name = tensor("attn_35_transpose_y_0"), val = tensor(true)]; + tensor attn_35_cast_fp16 = matmul(transpose_x = attn_35_transpose_x_0, transpose_y = attn_35_transpose_y_0, x = var_5086_cast_fp16, y = var_5084_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_5089 = const()[name = tensor("op_5089"), val = tensor([1, 1024, 1, 188])]; + tensor input_465_cast_fp16 = reshape(shape = var_5089, x = attn_35_cast_fp16)[name = tensor("input_465_cast_fp16")]; + tensor obj_73_pad_type_0 = const()[name = tensor("obj_73_pad_type_0"), val = tensor("valid")]; + tensor obj_73_strides_0 = const()[name = tensor("obj_73_strides_0"), val = tensor([1, 1])]; + tensor obj_73_pad_0 = const()[name = tensor("obj_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_73_dilations_0 = const()[name = tensor("obj_73_dilations_0"), val = tensor([1, 1])]; + tensor obj_73_groups_0 = const()[name = tensor("obj_73_groups_0"), val = tensor(1)]; + tensor layers_17_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334508672))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335295168))), name = tensor("layers_17_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_73_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_73_dilations_0, groups = obj_73_groups_0, pad = obj_73_pad_0, pad_type = obj_73_pad_type_0, strides = obj_73_strides_0, weight = layers_17_self_attn_o_proj_weight_to_fp16_palettized, x = input_465_cast_fp16)[name = tensor("obj_73_cast_fp16")]; + tensor inputs_175_cast_fp16 = add(x = inputs_173_cast_fp16, y = obj_73_cast_fp16)[name = tensor("inputs_175_cast_fp16")]; + tensor out_175_axes_0 = const()[name = tensor("out_175_axes_0"), val = tensor([1])]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_175_cast_fp16 = layer_norm(axes = out_175_axes_0, epsilon = var_5107_to_fp16, x = inputs_175_cast_fp16)[name = tensor("out_175_cast_fp16")]; + tensor input_467_gamma_0_to_fp16 = const()[name = tensor("input_467_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335295360)))]; + tensor input_467_beta_0_to_fp16 = const()[name = tensor("input_467_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335297472)))]; + tensor input_467_epsilon_0_to_fp16 = const()[name = tensor("input_467_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_467_cast_fp16 = batch_norm(beta = input_467_beta_0_to_fp16, epsilon = input_467_epsilon_0_to_fp16, gamma = input_467_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor("input_467_cast_fp16")]; + tensor input_469_pad_type_0 = const()[name = tensor("input_469_pad_type_0"), val = tensor("valid")]; + tensor input_469_strides_0 = const()[name = tensor("input_469_strides_0"), val = tensor([1, 1])]; + tensor input_469_pad_0 = const()[name = tensor("input_469_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_469_dilations_0 = const()[name = tensor("input_469_dilations_0"), val = tensor([1, 1])]; + tensor input_469_groups_0 = const()[name = tensor("input_469_groups_0"), val = tensor(1)]; + tensor layers_17_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(335299584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336872512))), name = tensor("layers_17_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_469_cast_fp16 = conv(dilations = input_469_dilations_0, groups = input_469_groups_0, pad = input_469_pad_0, pad_type = input_469_pad_type_0, strides = input_469_strides_0, weight = layers_17_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_467_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor input_471_split_num_splits_0 = const()[name = tensor("input_471_split_num_splits_0"), val = tensor(2)]; + tensor input_471_split_axis_0 = const()[name = tensor("input_471_split_axis_0"), val = tensor(1)]; + tensor input_471_split_cast_fp16_0, tensor input_471_split_cast_fp16_1 = split(axis = input_471_split_axis_0, num_splits = input_471_split_num_splits_0, x = input_469_cast_fp16)[name = tensor("input_471_split_cast_fp16")]; + tensor input_471_split_1_sigmoid_cast_fp16 = sigmoid(x = input_471_split_cast_fp16_1)[name = tensor("input_471_split_1_sigmoid_cast_fp16")]; + tensor input_471_cast_fp16 = mul(x = input_471_split_cast_fp16_0, y = input_471_split_1_sigmoid_cast_fp16)[name = tensor("input_471_cast_fp16")]; + tensor input_473_pad_type_0 = const()[name = tensor("input_473_pad_type_0"), val = tensor("custom")]; + tensor input_473_pad_0 = const()[name = tensor("input_473_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_473_groups_0 = const()[name = tensor("input_473_groups_0"), val = tensor(1024)]; + tensor input_473_strides_0 = const()[name = tensor("input_473_strides_0"), val = tensor([1, 1])]; + tensor input_473_dilations_0 = const()[name = tensor("input_473_dilations_0"), val = tensor([1, 1])]; + tensor const_302_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336872704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336879680))), name = tensor("const_302_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_303_to_fp16 = const()[name = tensor("const_303_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336879872)))]; + tensor input_475_cast_fp16 = conv(bias = const_303_to_fp16, dilations = input_473_dilations_0, groups = input_473_groups_0, pad = input_473_pad_0, pad_type = input_473_pad_type_0, strides = input_473_strides_0, weight = const_302_to_fp16_palettized, x = input_471_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor input_477_cast_fp16 = silu(x = input_475_cast_fp16)[name = tensor("input_477_cast_fp16")]; + tensor x_107_pad_type_0 = const()[name = tensor("x_107_pad_type_0"), val = tensor("valid")]; + tensor x_107_strides_0 = const()[name = tensor("x_107_strides_0"), val = tensor([1, 1])]; + tensor x_107_pad_0 = const()[name = tensor("x_107_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_107_dilations_0 = const()[name = tensor("x_107_dilations_0"), val = tensor([1, 1])]; + tensor x_107_groups_0 = const()[name = tensor("x_107_groups_0"), val = tensor(1)]; + tensor layers_17_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336881984))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337668480))), name = tensor("layers_17_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_107_cast_fp16 = conv(dilations = x_107_dilations_0, groups = x_107_groups_0, pad = x_107_pad_0, pad_type = x_107_pad_type_0, strides = x_107_strides_0, weight = layers_17_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_477_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor inputs_177_cast_fp16 = add(x = inputs_175_cast_fp16, y = x_107_cast_fp16)[name = tensor("inputs_177_cast_fp16")]; + tensor out_177_axes_0 = const()[name = tensor("out_177_axes_0"), val = tensor([1])]; + tensor var_5155_to_fp16 = const()[name = tensor("op_5155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_177_cast_fp16 = layer_norm(axes = out_177_axes_0, epsilon = var_5155_to_fp16, x = inputs_177_cast_fp16)[name = tensor("out_177_cast_fp16")]; + tensor input_479_gamma_0_to_fp16 = const()[name = tensor("input_479_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337668672)))]; + tensor input_479_beta_0_to_fp16 = const()[name = tensor("input_479_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337670784)))]; + tensor input_479_epsilon_0_to_fp16 = const()[name = tensor("input_479_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_479_cast_fp16 = batch_norm(beta = input_479_beta_0_to_fp16, epsilon = input_479_epsilon_0_to_fp16, gamma = input_479_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor("input_479_cast_fp16")]; + tensor input_481_pad_type_0 = const()[name = tensor("input_481_pad_type_0"), val = tensor("valid")]; + tensor input_481_strides_0 = const()[name = tensor("input_481_strides_0"), val = tensor([1, 1])]; + tensor input_481_pad_0 = const()[name = tensor("input_481_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_481_dilations_0 = const()[name = tensor("input_481_dilations_0"), val = tensor([1, 1])]; + tensor input_481_groups_0 = const()[name = tensor("input_481_groups_0"), val = tensor(1)]; + tensor layers_17_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(337672896))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340818688))), name = tensor("layers_17_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_481_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_481_dilations_0, groups = input_481_groups_0, pad = input_481_pad_0, pad_type = input_481_pad_type_0, strides = input_481_strides_0, weight = layers_17_feed_forward2_fc1_weight_to_fp16_palettized, x = input_479_cast_fp16)[name = tensor("input_481_cast_fp16")]; + tensor input_483_cast_fp16 = silu(x = input_481_cast_fp16)[name = tensor("input_483_cast_fp16")]; + tensor x_109_pad_type_0 = const()[name = tensor("x_109_pad_type_0"), val = tensor("valid")]; + tensor x_109_strides_0 = const()[name = tensor("x_109_strides_0"), val = tensor([1, 1])]; + tensor x_109_pad_0 = const()[name = tensor("x_109_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_109_dilations_0 = const()[name = tensor("x_109_dilations_0"), val = tensor([1, 1])]; + tensor x_109_groups_0 = const()[name = tensor("x_109_groups_0"), val = tensor(1)]; + tensor op_5183_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(340818880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343964672))), name = tensor("op_5183_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_5183_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, 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 = op_5183_weight_0_to_fp16_palettized, x = input_483_cast_fp16)[name = tensor("op_5183_cast_fp16")]; + tensor inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = var_5183_cast_fp16)[name = tensor("inputs_179_cast_fp16")]; + tensor out_179_axes_0 = const()[name = tensor("out_179_axes_0"), val = tensor([1])]; + tensor var_5193_to_fp16 = const()[name = tensor("op_5193_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_179_cast_fp16 = layer_norm(axes = out_179_axes_0, epsilon = var_5193_to_fp16, x = inputs_179_cast_fp16)[name = tensor("out_179_cast_fp16")]; + tensor inputs_181_gamma_0_to_fp16 = const()[name = tensor("inputs_181_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343964864)))]; + tensor inputs_181_beta_0_to_fp16 = const()[name = tensor("inputs_181_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343966976)))]; + tensor inputs_181_epsilon_0_to_fp16 = const()[name = tensor("inputs_181_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_181_cast_fp16 = batch_norm(beta = inputs_181_beta_0_to_fp16, epsilon = inputs_181_epsilon_0_to_fp16, gamma = inputs_181_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor("inputs_181_cast_fp16")]; + tensor var_5207 = const()[name = tensor("op_5207"), val = tensor(3)]; + tensor out_181_axes_0 = const()[name = tensor("out_181_axes_0"), val = tensor([1])]; + tensor var_5238_to_fp16 = const()[name = tensor("op_5238_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_181_cast_fp16 = layer_norm(axes = out_181_axes_0, epsilon = var_5238_to_fp16, x = inputs_181_cast_fp16)[name = tensor("out_181_cast_fp16")]; + tensor input_485_gamma_0_to_fp16 = const()[name = tensor("input_485_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343969088)))]; + tensor input_485_beta_0_to_fp16 = const()[name = tensor("input_485_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343971200)))]; + tensor input_485_epsilon_0_to_fp16 = const()[name = tensor("input_485_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_485_cast_fp16 = batch_norm(beta = input_485_beta_0_to_fp16, epsilon = input_485_epsilon_0_to_fp16, gamma = input_485_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor("input_485_cast_fp16")]; + tensor input_487_pad_type_0 = const()[name = tensor("input_487_pad_type_0"), val = tensor("valid")]; + tensor input_487_strides_0 = const()[name = tensor("input_487_strides_0"), val = tensor([1, 1])]; + tensor input_487_pad_0 = const()[name = tensor("input_487_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_487_dilations_0 = const()[name = tensor("input_487_dilations_0"), val = tensor([1, 1])]; + tensor input_487_groups_0 = const()[name = tensor("input_487_groups_0"), val = tensor(1)]; + tensor layers_18_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343973312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347119104))), name = tensor("layers_18_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_487_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_487_dilations_0, groups = input_487_groups_0, pad = input_487_pad_0, pad_type = input_487_pad_type_0, strides = input_487_strides_0, weight = layers_18_feed_forward1_fc1_weight_to_fp16_palettized, x = input_485_cast_fp16)[name = tensor("input_487_cast_fp16")]; + tensor input_489_cast_fp16 = silu(x = input_487_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor x_111_pad_type_0 = const()[name = tensor("x_111_pad_type_0"), val = tensor("valid")]; + tensor x_111_strides_0 = const()[name = tensor("x_111_strides_0"), val = tensor([1, 1])]; + tensor x_111_pad_0 = const()[name = tensor("x_111_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_111_dilations_0 = const()[name = tensor("x_111_dilations_0"), val = tensor([1, 1])]; + tensor x_111_groups_0 = const()[name = tensor("x_111_groups_0"), val = tensor(1)]; + tensor op_5266_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(347119296))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350265088))), name = tensor("op_5266_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_5266_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_111_dilations_0, groups = x_111_groups_0, pad = x_111_pad_0, pad_type = x_111_pad_type_0, strides = x_111_strides_0, weight = op_5266_weight_0_to_fp16_palettized, x = input_489_cast_fp16)[name = tensor("op_5266_cast_fp16")]; + tensor inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = var_5266_cast_fp16)[name = tensor("inputs_183_cast_fp16")]; + tensor out_183_axes_0 = const()[name = tensor("out_183_axes_0"), val = tensor([1])]; + tensor var_5276_to_fp16 = const()[name = tensor("op_5276_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_183_cast_fp16 = layer_norm(axes = out_183_axes_0, epsilon = var_5276_to_fp16, x = inputs_183_cast_fp16)[name = tensor("out_183_cast_fp16")]; + tensor obj_75_gamma_0_to_fp16 = const()[name = tensor("obj_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350265280)))]; + tensor obj_75_beta_0_to_fp16 = const()[name = tensor("obj_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350267392)))]; + tensor obj_75_epsilon_0_to_fp16 = const()[name = tensor("obj_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_75_cast_fp16 = batch_norm(beta = obj_75_beta_0_to_fp16, epsilon = obj_75_epsilon_0_to_fp16, gamma = obj_75_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_183_cast_fp16)[name = tensor("obj_75_cast_fp16")]; + tensor query_73_pad_type_0 = const()[name = tensor("query_73_pad_type_0"), val = tensor("valid")]; + tensor query_73_strides_0 = const()[name = tensor("query_73_strides_0"), val = tensor([1, 1])]; + tensor query_73_pad_0 = const()[name = tensor("query_73_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_73_dilations_0 = const()[name = tensor("query_73_dilations_0"), val = tensor([1, 1])]; + tensor query_73_groups_0 = const()[name = tensor("query_73_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(350269504))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351056000))), name = tensor("layers_18_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_73_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_73_dilations_0, groups = query_73_groups_0, pad = query_73_pad_0, pad_type = query_73_pad_type_0, strides = query_73_strides_0, weight = layers_18_self_attn_q_proj_weight_to_fp16_palettized, x = obj_75_cast_fp16)[name = tensor("query_73_cast_fp16")]; + tensor key_37_pad_type_0 = const()[name = tensor("key_37_pad_type_0"), val = tensor("valid")]; + tensor key_37_strides_0 = const()[name = tensor("key_37_strides_0"), val = tensor([1, 1])]; + tensor key_37_pad_0 = const()[name = tensor("key_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_37_dilations_0 = const()[name = tensor("key_37_dilations_0"), val = tensor([1, 1])]; + tensor key_37_groups_0 = const()[name = tensor("key_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351056192))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351842688))), name = tensor("layers_18_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_37_cast_fp16 = conv(dilations = key_37_dilations_0, groups = key_37_groups_0, pad = key_37_pad_0, pad_type = key_37_pad_type_0, strides = key_37_strides_0, weight = layers_18_self_attn_k_proj_weight_to_fp16_palettized, x = obj_75_cast_fp16)[name = tensor("key_37_cast_fp16")]; + tensor value_37_pad_type_0 = const()[name = tensor("value_37_pad_type_0"), val = tensor("valid")]; + tensor value_37_strides_0 = const()[name = tensor("value_37_strides_0"), val = tensor([1, 1])]; + tensor value_37_pad_0 = const()[name = tensor("value_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_37_dilations_0 = const()[name = tensor("value_37_dilations_0"), val = tensor([1, 1])]; + tensor value_37_groups_0 = const()[name = tensor("value_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(351842880))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352629376))), name = tensor("layers_18_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_37_dilations_0, groups = value_37_groups_0, pad = value_37_pad_0, pad_type = value_37_pad_type_0, strides = value_37_strides_0, weight = layers_18_self_attn_v_proj_weight_to_fp16_palettized, x = obj_75_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_5314_to_fp16 = const()[name = tensor("op_5314_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352629568)))]; + tensor query_75_cast_fp16 = add(x = query_73_cast_fp16, y = var_5314_to_fp16)[name = tensor("query_75_cast_fp16")]; + tensor var_5317_to_fp16 = const()[name = tensor("op_5317_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352631680)))]; + tensor q_with_bias_v_37_cast_fp16 = add(x = query_73_cast_fp16, y = var_5317_to_fp16)[name = tensor("q_with_bias_v_37_cast_fp16")]; + tensor p_37_pad_type_0 = const()[name = tensor("p_37_pad_type_0"), val = tensor("valid")]; + tensor p_37_strides_0 = const()[name = tensor("p_37_strides_0"), val = tensor([1, 1])]; + tensor p_37_pad_0 = const()[name = tensor("p_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_37_dilations_0 = const()[name = tensor("p_37_dilations_0"), val = tensor([1, 1])]; + tensor p_37_groups_0 = const()[name = tensor("p_37_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(352633792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353420288))), name = tensor("layers_18_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_37_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_37_dilations_0, groups = p_37_groups_0, pad = p_37_pad_0, pad_type = p_37_pad_type_0, strides = p_37_strides_0, weight = layers_18_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_37_cast_fp16")]; + tensor var_5328 = const()[name = tensor("op_5328"), val = tensor([1, 8, 128, 188])]; + tensor var_5329_cast_fp16 = reshape(shape = var_5328, x = q_with_bias_v_37_cast_fp16)[name = tensor("op_5329_cast_fp16")]; + tensor var_5330 = const()[name = tensor("op_5330"), val = tensor([1, 8, 128, -1])]; + tensor var_5331_cast_fp16 = reshape(shape = var_5330, x = p_37_cast_fp16)[name = tensor("op_5331_cast_fp16")]; + tensor matrix_bd_145_transpose_x_0 = const()[name = tensor("matrix_bd_145_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_145_transpose_y_0 = const()[name = tensor("matrix_bd_145_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_145_cast_fp16 = matmul(transpose_x = matrix_bd_145_transpose_x_0, transpose_y = matrix_bd_145_transpose_y_0, x = var_5329_cast_fp16, y = var_5331_cast_fp16)[name = tensor("matrix_bd_145_cast_fp16")]; + tensor matrix_bd_147_pad_0 = const()[name = tensor("matrix_bd_147_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_147_mode_0 = const()[name = tensor("matrix_bd_147_mode_0"), val = tensor("constant")]; + tensor const_208_to_fp16 = const()[name = tensor("const_208_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_147_cast_fp16 = pad(constant_val = const_208_to_fp16, mode = matrix_bd_147_mode_0, pad = matrix_bd_147_pad_0, x = matrix_bd_145_cast_fp16)[name = tensor("matrix_bd_147_cast_fp16")]; + tensor var_5340 = const()[name = tensor("op_5340"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_149_cast_fp16 = reshape(shape = var_5340, x = matrix_bd_147_cast_fp16)[name = tensor("matrix_bd_149_cast_fp16")]; + tensor var_5344_begin_0 = const()[name = tensor("op_5344_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5344_end_0 = const()[name = tensor("op_5344_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5344_end_mask_0 = const()[name = tensor("op_5344_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, x = matrix_bd_149_cast_fp16)[name = tensor("op_5344_cast_fp16")]; + tensor var_5345 = const()[name = tensor("op_5345"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_151_cast_fp16 = reshape(shape = var_5345, x = var_5344_cast_fp16)[name = tensor("matrix_bd_151_cast_fp16")]; + tensor var_5350_begin_0 = const()[name = tensor("op_5350_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5350_end_0 = const()[name = tensor("op_5350_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5350_end_mask_0 = const()[name = tensor("op_5350_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5350_cast_fp16 = slice_by_index(begin = var_5350_begin_0, end = var_5350_end_0, end_mask = var_5350_end_mask_0, x = matrix_bd_151_cast_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor var_5351_to_fp16 = const()[name = tensor("op_5351_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_37_cast_fp16 = mul(x = var_5350_cast_fp16, y = var_5351_to_fp16)[name = tensor("qk_mask_37_cast_fp16")]; + tensor var_5355 = const()[name = tensor("op_5355"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_5355, x = query_75_cast_fp16)[name = tensor("mh_q_37_cast_fp16")]; + tensor var_5357_to_fp16 = const()[name = tensor("op_5357_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5358_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_5357_to_fp16)[name = tensor("op_5358_cast_fp16")]; + tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 8, 128, 188])]; + tensor var_5362_cast_fp16 = reshape(shape = var_5361, x = key_37_cast_fp16)[name = tensor("op_5362_cast_fp16")]; + tensor mh_w_73_transpose_x_0 = const()[name = tensor("mh_w_73_transpose_x_0"), val = tensor(true)]; + tensor mh_w_73_transpose_y_0 = const()[name = tensor("mh_w_73_transpose_y_0"), val = tensor(false)]; + tensor mh_w_73_cast_fp16 = matmul(transpose_x = mh_w_73_transpose_x_0, transpose_y = mh_w_73_transpose_y_0, x = var_5358_cast_fp16, y = var_5362_cast_fp16)[name = tensor("mh_w_73_cast_fp16")]; + tensor mh_w_75_cast_fp16 = add(x = mh_w_73_cast_fp16, y = qk_mask_37_cast_fp16)[name = tensor("mh_w_75_cast_fp16")]; + tensor var_5366_cast_fp16 = softmax(axis = var_5207, x = mh_w_75_cast_fp16)[name = tensor("op_5366_cast_fp16")]; + tensor var_5367 = const()[name = tensor("op_5367"), val = tensor([1, 8, 128, 188])]; + tensor var_5368_cast_fp16 = reshape(shape = var_5367, x = value_37_cast_fp16)[name = tensor("op_5368_cast_fp16")]; + tensor attn_37_transpose_x_0 = const()[name = tensor("attn_37_transpose_x_0"), val = tensor(false)]; + tensor attn_37_transpose_y_0 = const()[name = tensor("attn_37_transpose_y_0"), val = tensor(true)]; + tensor attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = var_5368_cast_fp16, y = var_5366_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_5371 = const()[name = tensor("op_5371"), val = tensor([1, 1024, 1, 188])]; + tensor input_491_cast_fp16 = reshape(shape = var_5371, x = attn_37_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor obj_77_pad_type_0 = const()[name = tensor("obj_77_pad_type_0"), val = tensor("valid")]; + tensor obj_77_strides_0 = const()[name = tensor("obj_77_strides_0"), val = tensor([1, 1])]; + tensor obj_77_pad_0 = const()[name = tensor("obj_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_77_dilations_0 = const()[name = tensor("obj_77_dilations_0"), val = tensor([1, 1])]; + tensor obj_77_groups_0 = const()[name = tensor("obj_77_groups_0"), val = tensor(1)]; + tensor layers_18_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353420480))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354206976))), name = tensor("layers_18_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_77_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_77_dilations_0, groups = obj_77_groups_0, pad = obj_77_pad_0, pad_type = obj_77_pad_type_0, strides = obj_77_strides_0, weight = layers_18_self_attn_o_proj_weight_to_fp16_palettized, x = input_491_cast_fp16)[name = tensor("obj_77_cast_fp16")]; + tensor inputs_185_cast_fp16 = add(x = inputs_183_cast_fp16, y = obj_77_cast_fp16)[name = tensor("inputs_185_cast_fp16")]; + tensor out_185_axes_0 = const()[name = tensor("out_185_axes_0"), val = tensor([1])]; + tensor var_5389_to_fp16 = const()[name = tensor("op_5389_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_185_cast_fp16 = layer_norm(axes = out_185_axes_0, epsilon = var_5389_to_fp16, x = inputs_185_cast_fp16)[name = tensor("out_185_cast_fp16")]; + tensor input_493_gamma_0_to_fp16 = const()[name = tensor("input_493_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354207168)))]; + tensor input_493_beta_0_to_fp16 = const()[name = tensor("input_493_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354209280)))]; + tensor input_493_epsilon_0_to_fp16 = const()[name = tensor("input_493_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_493_cast_fp16 = batch_norm(beta = input_493_beta_0_to_fp16, epsilon = input_493_epsilon_0_to_fp16, gamma = input_493_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor("input_493_cast_fp16")]; + tensor input_495_pad_type_0 = const()[name = tensor("input_495_pad_type_0"), val = tensor("valid")]; + tensor input_495_strides_0 = const()[name = tensor("input_495_strides_0"), val = tensor([1, 1])]; + tensor input_495_pad_0 = const()[name = tensor("input_495_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_495_dilations_0 = const()[name = tensor("input_495_dilations_0"), val = tensor([1, 1])]; + tensor input_495_groups_0 = const()[name = tensor("input_495_groups_0"), val = tensor(1)]; + tensor layers_18_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(354211392))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355784320))), name = tensor("layers_18_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_495_cast_fp16 = conv(dilations = input_495_dilations_0, groups = input_495_groups_0, pad = input_495_pad_0, pad_type = input_495_pad_type_0, strides = input_495_strides_0, weight = layers_18_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_493_cast_fp16)[name = tensor("input_495_cast_fp16")]; + tensor input_497_split_num_splits_0 = const()[name = tensor("input_497_split_num_splits_0"), val = tensor(2)]; + tensor input_497_split_axis_0 = const()[name = tensor("input_497_split_axis_0"), val = tensor(1)]; + tensor input_497_split_cast_fp16_0, tensor input_497_split_cast_fp16_1 = split(axis = input_497_split_axis_0, num_splits = input_497_split_num_splits_0, x = input_495_cast_fp16)[name = tensor("input_497_split_cast_fp16")]; + tensor input_497_split_1_sigmoid_cast_fp16 = sigmoid(x = input_497_split_cast_fp16_1)[name = tensor("input_497_split_1_sigmoid_cast_fp16")]; + tensor input_497_cast_fp16 = mul(x = input_497_split_cast_fp16_0, y = input_497_split_1_sigmoid_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor input_499_pad_type_0 = const()[name = tensor("input_499_pad_type_0"), val = tensor("custom")]; + tensor input_499_pad_0 = const()[name = tensor("input_499_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_499_groups_0 = const()[name = tensor("input_499_groups_0"), val = tensor(1024)]; + tensor input_499_strides_0 = const()[name = tensor("input_499_strides_0"), val = tensor([1, 1])]; + tensor input_499_dilations_0 = const()[name = tensor("input_499_dilations_0"), val = tensor([1, 1])]; + tensor const_304_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355784512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355791488))), name = tensor("const_304_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_305_to_fp16 = const()[name = tensor("const_305_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355791680)))]; + tensor input_501_cast_fp16 = conv(bias = const_305_to_fp16, dilations = input_499_dilations_0, groups = input_499_groups_0, pad = input_499_pad_0, pad_type = input_499_pad_type_0, strides = input_499_strides_0, weight = const_304_to_fp16_palettized, x = input_497_cast_fp16)[name = tensor("input_501_cast_fp16")]; + tensor input_503_cast_fp16 = silu(x = input_501_cast_fp16)[name = tensor("input_503_cast_fp16")]; + tensor x_113_pad_type_0 = const()[name = tensor("x_113_pad_type_0"), val = tensor("valid")]; + tensor x_113_strides_0 = const()[name = tensor("x_113_strides_0"), val = tensor([1, 1])]; + tensor x_113_pad_0 = const()[name = tensor("x_113_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_113_dilations_0 = const()[name = tensor("x_113_dilations_0"), val = tensor([1, 1])]; + tensor x_113_groups_0 = const()[name = tensor("x_113_groups_0"), val = tensor(1)]; + tensor layers_18_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(355793792))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356580288))), name = tensor("layers_18_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_113_cast_fp16 = conv(dilations = x_113_dilations_0, groups = x_113_groups_0, pad = x_113_pad_0, pad_type = x_113_pad_type_0, strides = x_113_strides_0, weight = layers_18_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_503_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor inputs_187_cast_fp16 = add(x = inputs_185_cast_fp16, y = x_113_cast_fp16)[name = tensor("inputs_187_cast_fp16")]; + tensor out_187_axes_0 = const()[name = tensor("out_187_axes_0"), val = tensor([1])]; + tensor var_5437_to_fp16 = const()[name = tensor("op_5437_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_187_cast_fp16 = layer_norm(axes = out_187_axes_0, epsilon = var_5437_to_fp16, x = inputs_187_cast_fp16)[name = tensor("out_187_cast_fp16")]; + tensor input_505_gamma_0_to_fp16 = const()[name = tensor("input_505_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356580480)))]; + tensor input_505_beta_0_to_fp16 = const()[name = tensor("input_505_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356582592)))]; + tensor input_505_epsilon_0_to_fp16 = const()[name = tensor("input_505_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_505_cast_fp16 = batch_norm(beta = input_505_beta_0_to_fp16, epsilon = input_505_epsilon_0_to_fp16, gamma = input_505_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor input_507_pad_type_0 = const()[name = tensor("input_507_pad_type_0"), val = tensor("valid")]; + tensor input_507_strides_0 = const()[name = tensor("input_507_strides_0"), val = tensor([1, 1])]; + tensor input_507_pad_0 = const()[name = tensor("input_507_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_507_dilations_0 = const()[name = tensor("input_507_dilations_0"), val = tensor([1, 1])]; + tensor input_507_groups_0 = const()[name = tensor("input_507_groups_0"), val = tensor(1)]; + tensor layers_18_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(356584704))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359730496))), name = tensor("layers_18_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_507_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_507_dilations_0, groups = input_507_groups_0, pad = input_507_pad_0, pad_type = input_507_pad_type_0, strides = input_507_strides_0, weight = layers_18_feed_forward2_fc1_weight_to_fp16_palettized, x = input_505_cast_fp16)[name = tensor("input_507_cast_fp16")]; + tensor input_509_cast_fp16 = silu(x = input_507_cast_fp16)[name = tensor("input_509_cast_fp16")]; + tensor x_115_pad_type_0 = const()[name = tensor("x_115_pad_type_0"), val = tensor("valid")]; + tensor x_115_strides_0 = const()[name = tensor("x_115_strides_0"), val = tensor([1, 1])]; + tensor x_115_pad_0 = const()[name = tensor("x_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_115_dilations_0 = const()[name = tensor("x_115_dilations_0"), val = tensor([1, 1])]; + tensor x_115_groups_0 = const()[name = tensor("x_115_groups_0"), val = tensor(1)]; + tensor op_5465_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359730688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362876480))), name = tensor("op_5465_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_5465_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_115_dilations_0, groups = x_115_groups_0, pad = x_115_pad_0, pad_type = x_115_pad_type_0, strides = x_115_strides_0, weight = op_5465_weight_0_to_fp16_palettized, x = input_509_cast_fp16)[name = tensor("op_5465_cast_fp16")]; + tensor inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = var_5465_cast_fp16)[name = tensor("inputs_189_cast_fp16")]; + tensor out_189_axes_0 = const()[name = tensor("out_189_axes_0"), val = tensor([1])]; + tensor var_5475_to_fp16 = const()[name = tensor("op_5475_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_189_cast_fp16 = layer_norm(axes = out_189_axes_0, epsilon = var_5475_to_fp16, x = inputs_189_cast_fp16)[name = tensor("out_189_cast_fp16")]; + tensor inputs_191_gamma_0_to_fp16 = const()[name = tensor("inputs_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362876672)))]; + tensor inputs_191_beta_0_to_fp16 = const()[name = tensor("inputs_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362878784)))]; + tensor inputs_191_epsilon_0_to_fp16 = const()[name = tensor("inputs_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_191_cast_fp16 = batch_norm(beta = inputs_191_beta_0_to_fp16, epsilon = inputs_191_epsilon_0_to_fp16, gamma = inputs_191_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor("inputs_191_cast_fp16")]; + tensor var_5489 = const()[name = tensor("op_5489"), val = tensor(3)]; + tensor out_191_axes_0 = const()[name = tensor("out_191_axes_0"), val = tensor([1])]; + tensor var_5520_to_fp16 = const()[name = tensor("op_5520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_191_cast_fp16 = layer_norm(axes = out_191_axes_0, epsilon = var_5520_to_fp16, x = inputs_191_cast_fp16)[name = tensor("out_191_cast_fp16")]; + tensor input_511_gamma_0_to_fp16 = const()[name = tensor("input_511_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362880896)))]; + tensor input_511_beta_0_to_fp16 = const()[name = tensor("input_511_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362883008)))]; + tensor input_511_epsilon_0_to_fp16 = const()[name = tensor("input_511_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_511_cast_fp16 = batch_norm(beta = input_511_beta_0_to_fp16, epsilon = input_511_epsilon_0_to_fp16, gamma = input_511_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor("input_511_cast_fp16")]; + tensor input_513_pad_type_0 = const()[name = tensor("input_513_pad_type_0"), val = tensor("valid")]; + tensor input_513_strides_0 = const()[name = tensor("input_513_strides_0"), val = tensor([1, 1])]; + tensor input_513_pad_0 = const()[name = tensor("input_513_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_513_dilations_0 = const()[name = tensor("input_513_dilations_0"), val = tensor([1, 1])]; + tensor input_513_groups_0 = const()[name = tensor("input_513_groups_0"), val = tensor(1)]; + tensor layers_19_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362885120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366030912))), name = tensor("layers_19_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_513_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_513_dilations_0, groups = input_513_groups_0, pad = input_513_pad_0, pad_type = input_513_pad_type_0, strides = input_513_strides_0, weight = layers_19_feed_forward1_fc1_weight_to_fp16_palettized, x = input_511_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor input_515_cast_fp16 = silu(x = input_513_cast_fp16)[name = tensor("input_515_cast_fp16")]; + tensor x_117_pad_type_0 = const()[name = tensor("x_117_pad_type_0"), val = tensor("valid")]; + tensor x_117_strides_0 = const()[name = tensor("x_117_strides_0"), val = tensor([1, 1])]; + tensor x_117_pad_0 = const()[name = tensor("x_117_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_117_dilations_0 = const()[name = tensor("x_117_dilations_0"), val = tensor([1, 1])]; + tensor x_117_groups_0 = const()[name = tensor("x_117_groups_0"), val = tensor(1)]; + tensor op_5548_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(366031104))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369176896))), name = tensor("op_5548_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_5548_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_117_dilations_0, groups = x_117_groups_0, pad = x_117_pad_0, pad_type = x_117_pad_type_0, strides = x_117_strides_0, weight = op_5548_weight_0_to_fp16_palettized, x = input_515_cast_fp16)[name = tensor("op_5548_cast_fp16")]; + tensor inputs_193_cast_fp16 = add(x = inputs_191_cast_fp16, y = var_5548_cast_fp16)[name = tensor("inputs_193_cast_fp16")]; + tensor out_193_axes_0 = const()[name = tensor("out_193_axes_0"), val = tensor([1])]; + tensor var_5558_to_fp16 = const()[name = tensor("op_5558_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_193_cast_fp16 = layer_norm(axes = out_193_axes_0, epsilon = var_5558_to_fp16, x = inputs_193_cast_fp16)[name = tensor("out_193_cast_fp16")]; + tensor obj_79_gamma_0_to_fp16 = const()[name = tensor("obj_79_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369177088)))]; + tensor obj_79_beta_0_to_fp16 = const()[name = tensor("obj_79_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369179200)))]; + tensor obj_79_epsilon_0_to_fp16 = const()[name = tensor("obj_79_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_79_cast_fp16 = batch_norm(beta = obj_79_beta_0_to_fp16, epsilon = obj_79_epsilon_0_to_fp16, gamma = obj_79_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_193_cast_fp16)[name = tensor("obj_79_cast_fp16")]; + tensor query_77_pad_type_0 = const()[name = tensor("query_77_pad_type_0"), val = tensor("valid")]; + tensor query_77_strides_0 = const()[name = tensor("query_77_strides_0"), val = tensor([1, 1])]; + tensor query_77_pad_0 = const()[name = tensor("query_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_77_dilations_0 = const()[name = tensor("query_77_dilations_0"), val = tensor([1, 1])]; + tensor query_77_groups_0 = const()[name = tensor("query_77_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369181312))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369967808))), name = tensor("layers_19_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_77_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_77_dilations_0, groups = query_77_groups_0, pad = query_77_pad_0, pad_type = query_77_pad_type_0, strides = query_77_strides_0, weight = layers_19_self_attn_q_proj_weight_to_fp16_palettized, x = obj_79_cast_fp16)[name = tensor("query_77_cast_fp16")]; + tensor key_39_pad_type_0 = const()[name = tensor("key_39_pad_type_0"), val = tensor("valid")]; + tensor key_39_strides_0 = const()[name = tensor("key_39_strides_0"), val = tensor([1, 1])]; + tensor key_39_pad_0 = const()[name = tensor("key_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_39_dilations_0 = const()[name = tensor("key_39_dilations_0"), val = tensor([1, 1])]; + tensor key_39_groups_0 = const()[name = tensor("key_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(369968000))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370754496))), name = tensor("layers_19_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_39_cast_fp16 = conv(dilations = key_39_dilations_0, groups = key_39_groups_0, pad = key_39_pad_0, pad_type = key_39_pad_type_0, strides = key_39_strides_0, weight = layers_19_self_attn_k_proj_weight_to_fp16_palettized, x = obj_79_cast_fp16)[name = tensor("key_39_cast_fp16")]; + tensor value_39_pad_type_0 = const()[name = tensor("value_39_pad_type_0"), val = tensor("valid")]; + tensor value_39_strides_0 = const()[name = tensor("value_39_strides_0"), val = tensor([1, 1])]; + tensor value_39_pad_0 = const()[name = tensor("value_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_39_dilations_0 = const()[name = tensor("value_39_dilations_0"), val = tensor([1, 1])]; + tensor value_39_groups_0 = const()[name = tensor("value_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370754688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371541184))), name = tensor("layers_19_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_39_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_39_dilations_0, groups = value_39_groups_0, pad = value_39_pad_0, pad_type = value_39_pad_type_0, strides = value_39_strides_0, weight = layers_19_self_attn_v_proj_weight_to_fp16_palettized, x = obj_79_cast_fp16)[name = tensor("value_39_cast_fp16")]; + tensor var_5596_to_fp16 = const()[name = tensor("op_5596_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371541376)))]; + tensor query_79_cast_fp16 = add(x = query_77_cast_fp16, y = var_5596_to_fp16)[name = tensor("query_79_cast_fp16")]; + tensor var_5599_to_fp16 = const()[name = tensor("op_5599_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371543488)))]; + tensor q_with_bias_v_39_cast_fp16 = add(x = query_77_cast_fp16, y = var_5599_to_fp16)[name = tensor("q_with_bias_v_39_cast_fp16")]; + tensor p_39_pad_type_0 = const()[name = tensor("p_39_pad_type_0"), val = tensor("valid")]; + tensor p_39_strides_0 = const()[name = tensor("p_39_strides_0"), val = tensor([1, 1])]; + tensor p_39_pad_0 = const()[name = tensor("p_39_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_39_dilations_0 = const()[name = tensor("p_39_dilations_0"), val = tensor([1, 1])]; + tensor p_39_groups_0 = const()[name = tensor("p_39_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(371545600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372332096))), name = tensor("layers_19_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_39_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_39_dilations_0, groups = p_39_groups_0, pad = p_39_pad_0, pad_type = p_39_pad_type_0, strides = p_39_strides_0, weight = layers_19_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_39_cast_fp16")]; + tensor var_5610 = const()[name = tensor("op_5610"), val = tensor([1, 8, 128, 188])]; + tensor var_5611_cast_fp16 = reshape(shape = var_5610, x = q_with_bias_v_39_cast_fp16)[name = tensor("op_5611_cast_fp16")]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor([1, 8, 128, -1])]; + tensor var_5613_cast_fp16 = reshape(shape = var_5612, x = p_39_cast_fp16)[name = tensor("op_5613_cast_fp16")]; + tensor matrix_bd_153_transpose_x_0 = const()[name = tensor("matrix_bd_153_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_153_transpose_y_0 = const()[name = tensor("matrix_bd_153_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_153_cast_fp16 = matmul(transpose_x = matrix_bd_153_transpose_x_0, transpose_y = matrix_bd_153_transpose_y_0, x = var_5611_cast_fp16, y = var_5613_cast_fp16)[name = tensor("matrix_bd_153_cast_fp16")]; + tensor matrix_bd_155_pad_0 = const()[name = tensor("matrix_bd_155_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_155_mode_0 = const()[name = tensor("matrix_bd_155_mode_0"), val = tensor("constant")]; + tensor const_219_to_fp16 = const()[name = tensor("const_219_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_155_cast_fp16 = pad(constant_val = const_219_to_fp16, mode = matrix_bd_155_mode_0, pad = matrix_bd_155_pad_0, x = matrix_bd_153_cast_fp16)[name = tensor("matrix_bd_155_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_157_cast_fp16 = reshape(shape = var_5622, x = matrix_bd_155_cast_fp16)[name = tensor("matrix_bd_157_cast_fp16")]; + tensor var_5626_begin_0 = const()[name = tensor("op_5626_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5626_end_0 = const()[name = tensor("op_5626_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5626_end_mask_0 = const()[name = tensor("op_5626_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5626_cast_fp16 = slice_by_index(begin = var_5626_begin_0, end = var_5626_end_0, end_mask = var_5626_end_mask_0, x = matrix_bd_157_cast_fp16)[name = tensor("op_5626_cast_fp16")]; + tensor var_5627 = const()[name = tensor("op_5627"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_159_cast_fp16 = reshape(shape = var_5627, x = var_5626_cast_fp16)[name = tensor("matrix_bd_159_cast_fp16")]; + tensor var_5632_begin_0 = const()[name = tensor("op_5632_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5632_end_0 = const()[name = tensor("op_5632_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5632_end_mask_0 = const()[name = tensor("op_5632_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5632_cast_fp16 = slice_by_index(begin = var_5632_begin_0, end = var_5632_end_0, end_mask = var_5632_end_mask_0, x = matrix_bd_159_cast_fp16)[name = tensor("op_5632_cast_fp16")]; + tensor var_5633_to_fp16 = const()[name = tensor("op_5633_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_39_cast_fp16 = mul(x = var_5632_cast_fp16, y = var_5633_to_fp16)[name = tensor("qk_mask_39_cast_fp16")]; + tensor var_5637 = const()[name = tensor("op_5637"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_5637, x = query_79_cast_fp16)[name = tensor("mh_q_39_cast_fp16")]; + tensor var_5639_to_fp16 = const()[name = tensor("op_5639_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5640_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_5639_to_fp16)[name = tensor("op_5640_cast_fp16")]; + tensor var_5643 = const()[name = tensor("op_5643"), val = tensor([1, 8, 128, 188])]; + tensor var_5644_cast_fp16 = reshape(shape = var_5643, x = key_39_cast_fp16)[name = tensor("op_5644_cast_fp16")]; + tensor mh_w_77_transpose_x_0 = const()[name = tensor("mh_w_77_transpose_x_0"), val = tensor(true)]; + tensor mh_w_77_transpose_y_0 = const()[name = tensor("mh_w_77_transpose_y_0"), val = tensor(false)]; + tensor mh_w_77_cast_fp16 = matmul(transpose_x = mh_w_77_transpose_x_0, transpose_y = mh_w_77_transpose_y_0, x = var_5640_cast_fp16, y = var_5644_cast_fp16)[name = tensor("mh_w_77_cast_fp16")]; + tensor mh_w_79_cast_fp16 = add(x = mh_w_77_cast_fp16, y = qk_mask_39_cast_fp16)[name = tensor("mh_w_79_cast_fp16")]; + tensor var_5648_cast_fp16 = softmax(axis = var_5489, x = mh_w_79_cast_fp16)[name = tensor("op_5648_cast_fp16")]; + tensor var_5649 = const()[name = tensor("op_5649"), val = tensor([1, 8, 128, 188])]; + tensor var_5650_cast_fp16 = reshape(shape = var_5649, x = value_39_cast_fp16)[name = tensor("op_5650_cast_fp16")]; + tensor attn_39_transpose_x_0 = const()[name = tensor("attn_39_transpose_x_0"), val = tensor(false)]; + tensor attn_39_transpose_y_0 = const()[name = tensor("attn_39_transpose_y_0"), val = tensor(true)]; + tensor attn_39_cast_fp16 = matmul(transpose_x = attn_39_transpose_x_0, transpose_y = attn_39_transpose_y_0, x = var_5650_cast_fp16, y = var_5648_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_5653 = const()[name = tensor("op_5653"), val = tensor([1, 1024, 1, 188])]; + tensor input_517_cast_fp16 = reshape(shape = var_5653, x = attn_39_cast_fp16)[name = tensor("input_517_cast_fp16")]; + tensor obj_81_pad_type_0 = const()[name = tensor("obj_81_pad_type_0"), val = tensor("valid")]; + tensor obj_81_strides_0 = const()[name = tensor("obj_81_strides_0"), val = tensor([1, 1])]; + tensor obj_81_pad_0 = const()[name = tensor("obj_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_81_dilations_0 = const()[name = tensor("obj_81_dilations_0"), val = tensor([1, 1])]; + tensor obj_81_groups_0 = const()[name = tensor("obj_81_groups_0"), val = tensor(1)]; + tensor layers_19_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(372332288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373118784))), name = tensor("layers_19_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_81_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_81_dilations_0, groups = obj_81_groups_0, pad = obj_81_pad_0, pad_type = obj_81_pad_type_0, strides = obj_81_strides_0, weight = layers_19_self_attn_o_proj_weight_to_fp16_palettized, x = input_517_cast_fp16)[name = tensor("obj_81_cast_fp16")]; + tensor inputs_195_cast_fp16 = add(x = inputs_193_cast_fp16, y = obj_81_cast_fp16)[name = tensor("inputs_195_cast_fp16")]; + tensor out_195_axes_0 = const()[name = tensor("out_195_axes_0"), val = tensor([1])]; + tensor var_5671_to_fp16 = const()[name = tensor("op_5671_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_195_cast_fp16 = layer_norm(axes = out_195_axes_0, epsilon = var_5671_to_fp16, x = inputs_195_cast_fp16)[name = tensor("out_195_cast_fp16")]; + tensor input_519_gamma_0_to_fp16 = const()[name = tensor("input_519_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373118976)))]; + tensor input_519_beta_0_to_fp16 = const()[name = tensor("input_519_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373121088)))]; + tensor input_519_epsilon_0_to_fp16 = const()[name = tensor("input_519_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_519_cast_fp16 = batch_norm(beta = input_519_beta_0_to_fp16, epsilon = input_519_epsilon_0_to_fp16, gamma = input_519_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_195_cast_fp16)[name = tensor("input_519_cast_fp16")]; + tensor input_521_pad_type_0 = const()[name = tensor("input_521_pad_type_0"), val = tensor("valid")]; + tensor input_521_strides_0 = const()[name = tensor("input_521_strides_0"), val = tensor([1, 1])]; + tensor input_521_pad_0 = const()[name = tensor("input_521_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_521_dilations_0 = const()[name = tensor("input_521_dilations_0"), val = tensor([1, 1])]; + tensor input_521_groups_0 = const()[name = tensor("input_521_groups_0"), val = tensor(1)]; + tensor layers_19_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(373123200))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374696128))), name = tensor("layers_19_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_521_cast_fp16 = conv(dilations = input_521_dilations_0, groups = input_521_groups_0, pad = input_521_pad_0, pad_type = input_521_pad_type_0, strides = input_521_strides_0, weight = layers_19_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_519_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor input_523_split_num_splits_0 = const()[name = tensor("input_523_split_num_splits_0"), val = tensor(2)]; + tensor input_523_split_axis_0 = const()[name = tensor("input_523_split_axis_0"), val = tensor(1)]; + tensor input_523_split_cast_fp16_0, tensor input_523_split_cast_fp16_1 = split(axis = input_523_split_axis_0, num_splits = input_523_split_num_splits_0, x = input_521_cast_fp16)[name = tensor("input_523_split_cast_fp16")]; + tensor input_523_split_1_sigmoid_cast_fp16 = sigmoid(x = input_523_split_cast_fp16_1)[name = tensor("input_523_split_1_sigmoid_cast_fp16")]; + tensor input_523_cast_fp16 = mul(x = input_523_split_cast_fp16_0, y = input_523_split_1_sigmoid_cast_fp16)[name = tensor("input_523_cast_fp16")]; + tensor input_525_pad_type_0 = const()[name = tensor("input_525_pad_type_0"), val = tensor("custom")]; + tensor input_525_pad_0 = const()[name = tensor("input_525_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_525_groups_0 = const()[name = tensor("input_525_groups_0"), val = tensor(1024)]; + tensor input_525_strides_0 = const()[name = tensor("input_525_strides_0"), val = tensor([1, 1])]; + tensor input_525_dilations_0 = const()[name = tensor("input_525_dilations_0"), val = tensor([1, 1])]; + tensor const_306_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374696320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374703296))), name = tensor("const_306_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_307_to_fp16 = const()[name = tensor("const_307_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374703488)))]; + tensor input_527_cast_fp16 = conv(bias = const_307_to_fp16, dilations = input_525_dilations_0, groups = input_525_groups_0, pad = input_525_pad_0, pad_type = input_525_pad_type_0, strides = input_525_strides_0, weight = const_306_to_fp16_palettized, x = input_523_cast_fp16)[name = tensor("input_527_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = input_527_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor x_119_pad_type_0 = const()[name = tensor("x_119_pad_type_0"), val = tensor("valid")]; + tensor x_119_strides_0 = const()[name = tensor("x_119_strides_0"), val = tensor([1, 1])]; + tensor x_119_pad_0 = const()[name = tensor("x_119_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_119_dilations_0 = const()[name = tensor("x_119_dilations_0"), val = tensor([1, 1])]; + tensor x_119_groups_0 = const()[name = tensor("x_119_groups_0"), val = tensor(1)]; + tensor layers_19_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(374705600))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375492096))), name = tensor("layers_19_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_119_cast_fp16 = conv(dilations = x_119_dilations_0, groups = x_119_groups_0, pad = x_119_pad_0, pad_type = x_119_pad_type_0, strides = x_119_strides_0, weight = layers_19_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_529_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor inputs_197_cast_fp16 = add(x = inputs_195_cast_fp16, y = x_119_cast_fp16)[name = tensor("inputs_197_cast_fp16")]; + tensor out_197_axes_0 = const()[name = tensor("out_197_axes_0"), val = tensor([1])]; + tensor var_5719_to_fp16 = const()[name = tensor("op_5719_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_197_cast_fp16 = layer_norm(axes = out_197_axes_0, epsilon = var_5719_to_fp16, x = inputs_197_cast_fp16)[name = tensor("out_197_cast_fp16")]; + tensor input_531_gamma_0_to_fp16 = const()[name = tensor("input_531_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375492288)))]; + tensor input_531_beta_0_to_fp16 = const()[name = tensor("input_531_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375494400)))]; + tensor input_531_epsilon_0_to_fp16 = const()[name = tensor("input_531_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_531_cast_fp16 = batch_norm(beta = input_531_beta_0_to_fp16, epsilon = input_531_epsilon_0_to_fp16, gamma = input_531_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_197_cast_fp16)[name = tensor("input_531_cast_fp16")]; + tensor input_533_pad_type_0 = const()[name = tensor("input_533_pad_type_0"), val = tensor("valid")]; + tensor input_533_strides_0 = const()[name = tensor("input_533_strides_0"), val = tensor([1, 1])]; + tensor input_533_pad_0 = const()[name = tensor("input_533_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_533_dilations_0 = const()[name = tensor("input_533_dilations_0"), val = tensor([1, 1])]; + tensor input_533_groups_0 = const()[name = tensor("input_533_groups_0"), val = tensor(1)]; + tensor layers_19_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(375496512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378642304))), name = tensor("layers_19_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_533_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_533_dilations_0, groups = input_533_groups_0, pad = input_533_pad_0, pad_type = input_533_pad_type_0, strides = input_533_strides_0, weight = layers_19_feed_forward2_fc1_weight_to_fp16_palettized, x = input_531_cast_fp16)[name = tensor("input_533_cast_fp16")]; + tensor input_535_cast_fp16 = silu(x = input_533_cast_fp16)[name = tensor("input_535_cast_fp16")]; + tensor x_121_pad_type_0 = const()[name = tensor("x_121_pad_type_0"), val = tensor("valid")]; + tensor x_121_strides_0 = const()[name = tensor("x_121_strides_0"), val = tensor([1, 1])]; + tensor x_121_pad_0 = const()[name = tensor("x_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_121_dilations_0 = const()[name = tensor("x_121_dilations_0"), val = tensor([1, 1])]; + tensor x_121_groups_0 = const()[name = tensor("x_121_groups_0"), val = tensor(1)]; + tensor op_5747_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378642496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381788288))), name = tensor("op_5747_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_5747_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_121_dilations_0, groups = x_121_groups_0, pad = x_121_pad_0, pad_type = x_121_pad_type_0, strides = x_121_strides_0, weight = op_5747_weight_0_to_fp16_palettized, x = input_535_cast_fp16)[name = tensor("op_5747_cast_fp16")]; + tensor inputs_199_cast_fp16 = add(x = inputs_197_cast_fp16, y = var_5747_cast_fp16)[name = tensor("inputs_199_cast_fp16")]; + tensor out_199_axes_0 = const()[name = tensor("out_199_axes_0"), val = tensor([1])]; + tensor var_5757_to_fp16 = const()[name = tensor("op_5757_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_199_cast_fp16 = layer_norm(axes = out_199_axes_0, epsilon = var_5757_to_fp16, x = inputs_199_cast_fp16)[name = tensor("out_199_cast_fp16")]; + tensor inputs_201_gamma_0_to_fp16 = const()[name = tensor("inputs_201_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381788480)))]; + tensor inputs_201_beta_0_to_fp16 = const()[name = tensor("inputs_201_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381790592)))]; + tensor inputs_201_epsilon_0_to_fp16 = const()[name = tensor("inputs_201_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_201_cast_fp16 = batch_norm(beta = inputs_201_beta_0_to_fp16, epsilon = inputs_201_epsilon_0_to_fp16, gamma = inputs_201_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_199_cast_fp16)[name = tensor("inputs_201_cast_fp16")]; + tensor var_5771 = const()[name = tensor("op_5771"), val = tensor(3)]; + tensor out_201_axes_0 = const()[name = tensor("out_201_axes_0"), val = tensor([1])]; + tensor var_5802_to_fp16 = const()[name = tensor("op_5802_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_201_cast_fp16 = layer_norm(axes = out_201_axes_0, epsilon = var_5802_to_fp16, x = inputs_201_cast_fp16)[name = tensor("out_201_cast_fp16")]; + tensor input_537_gamma_0_to_fp16 = const()[name = tensor("input_537_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381792704)))]; + tensor input_537_beta_0_to_fp16 = const()[name = tensor("input_537_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381794816)))]; + tensor input_537_epsilon_0_to_fp16 = const()[name = tensor("input_537_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_537_cast_fp16 = batch_norm(beta = input_537_beta_0_to_fp16, epsilon = input_537_epsilon_0_to_fp16, gamma = input_537_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_201_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor input_539_pad_type_0 = const()[name = tensor("input_539_pad_type_0"), val = tensor("valid")]; + tensor input_539_strides_0 = const()[name = tensor("input_539_strides_0"), val = tensor([1, 1])]; + tensor input_539_pad_0 = const()[name = tensor("input_539_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_539_dilations_0 = const()[name = tensor("input_539_dilations_0"), val = tensor([1, 1])]; + tensor input_539_groups_0 = const()[name = tensor("input_539_groups_0"), val = tensor(1)]; + tensor layers_20_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(381796928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384942720))), name = tensor("layers_20_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_539_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_539_dilations_0, groups = input_539_groups_0, pad = input_539_pad_0, pad_type = input_539_pad_type_0, strides = input_539_strides_0, weight = layers_20_feed_forward1_fc1_weight_to_fp16_palettized, x = input_537_cast_fp16)[name = tensor("input_539_cast_fp16")]; + tensor input_541_cast_fp16 = silu(x = input_539_cast_fp16)[name = tensor("input_541_cast_fp16")]; + tensor x_123_pad_type_0 = const()[name = tensor("x_123_pad_type_0"), val = tensor("valid")]; + tensor x_123_strides_0 = const()[name = tensor("x_123_strides_0"), val = tensor([1, 1])]; + tensor x_123_pad_0 = const()[name = tensor("x_123_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_123_dilations_0 = const()[name = tensor("x_123_dilations_0"), val = tensor([1, 1])]; + tensor x_123_groups_0 = const()[name = tensor("x_123_groups_0"), val = tensor(1)]; + tensor op_5830_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384942912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388088704))), name = tensor("op_5830_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_5830_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_123_dilations_0, groups = x_123_groups_0, pad = x_123_pad_0, pad_type = x_123_pad_type_0, strides = x_123_strides_0, weight = op_5830_weight_0_to_fp16_palettized, x = input_541_cast_fp16)[name = tensor("op_5830_cast_fp16")]; + tensor inputs_203_cast_fp16 = add(x = inputs_201_cast_fp16, y = var_5830_cast_fp16)[name = tensor("inputs_203_cast_fp16")]; + tensor out_203_axes_0 = const()[name = tensor("out_203_axes_0"), val = tensor([1])]; + tensor var_5840_to_fp16 = const()[name = tensor("op_5840_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_203_cast_fp16 = layer_norm(axes = out_203_axes_0, epsilon = var_5840_to_fp16, x = inputs_203_cast_fp16)[name = tensor("out_203_cast_fp16")]; + tensor obj_83_gamma_0_to_fp16 = const()[name = tensor("obj_83_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388088896)))]; + tensor obj_83_beta_0_to_fp16 = const()[name = tensor("obj_83_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388091008)))]; + tensor obj_83_epsilon_0_to_fp16 = const()[name = tensor("obj_83_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_83_cast_fp16 = batch_norm(beta = obj_83_beta_0_to_fp16, epsilon = obj_83_epsilon_0_to_fp16, gamma = obj_83_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_203_cast_fp16)[name = tensor("obj_83_cast_fp16")]; + tensor query_81_pad_type_0 = const()[name = tensor("query_81_pad_type_0"), val = tensor("valid")]; + tensor query_81_strides_0 = const()[name = tensor("query_81_strides_0"), val = tensor([1, 1])]; + tensor query_81_pad_0 = const()[name = tensor("query_81_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_81_dilations_0 = const()[name = tensor("query_81_dilations_0"), val = tensor([1, 1])]; + tensor query_81_groups_0 = const()[name = tensor("query_81_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388093120))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388879616))), name = tensor("layers_20_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_81_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_81_dilations_0, groups = query_81_groups_0, pad = query_81_pad_0, pad_type = query_81_pad_type_0, strides = query_81_strides_0, weight = layers_20_self_attn_q_proj_weight_to_fp16_palettized, x = obj_83_cast_fp16)[name = tensor("query_81_cast_fp16")]; + tensor key_41_pad_type_0 = const()[name = tensor("key_41_pad_type_0"), val = tensor("valid")]; + tensor key_41_strides_0 = const()[name = tensor("key_41_strides_0"), val = tensor([1, 1])]; + tensor key_41_pad_0 = const()[name = tensor("key_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_41_dilations_0 = const()[name = tensor("key_41_dilations_0"), val = tensor([1, 1])]; + tensor key_41_groups_0 = const()[name = tensor("key_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(388879808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389666304))), name = tensor("layers_20_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_41_cast_fp16 = conv(dilations = key_41_dilations_0, groups = key_41_groups_0, pad = key_41_pad_0, pad_type = key_41_pad_type_0, strides = key_41_strides_0, weight = layers_20_self_attn_k_proj_weight_to_fp16_palettized, x = obj_83_cast_fp16)[name = tensor("key_41_cast_fp16")]; + tensor value_41_pad_type_0 = const()[name = tensor("value_41_pad_type_0"), val = tensor("valid")]; + tensor value_41_strides_0 = const()[name = tensor("value_41_strides_0"), val = tensor([1, 1])]; + tensor value_41_pad_0 = const()[name = tensor("value_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_41_dilations_0 = const()[name = tensor("value_41_dilations_0"), val = tensor([1, 1])]; + tensor value_41_groups_0 = const()[name = tensor("value_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389666496))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390452992))), name = tensor("layers_20_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_41_dilations_0, groups = value_41_groups_0, pad = value_41_pad_0, pad_type = value_41_pad_type_0, strides = value_41_strides_0, weight = layers_20_self_attn_v_proj_weight_to_fp16_palettized, x = obj_83_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_5878_to_fp16 = const()[name = tensor("op_5878_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390453184)))]; + tensor query_83_cast_fp16 = add(x = query_81_cast_fp16, y = var_5878_to_fp16)[name = tensor("query_83_cast_fp16")]; + tensor var_5881_to_fp16 = const()[name = tensor("op_5881_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390455296)))]; + tensor q_with_bias_v_41_cast_fp16 = add(x = query_81_cast_fp16, y = var_5881_to_fp16)[name = tensor("q_with_bias_v_41_cast_fp16")]; + tensor p_41_pad_type_0 = const()[name = tensor("p_41_pad_type_0"), val = tensor("valid")]; + tensor p_41_strides_0 = const()[name = tensor("p_41_strides_0"), val = tensor([1, 1])]; + tensor p_41_pad_0 = const()[name = tensor("p_41_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_41_dilations_0 = const()[name = tensor("p_41_dilations_0"), val = tensor([1, 1])]; + tensor p_41_groups_0 = const()[name = tensor("p_41_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(390457408))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391243904))), name = tensor("layers_20_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_41_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_41_dilations_0, groups = p_41_groups_0, pad = p_41_pad_0, pad_type = p_41_pad_type_0, strides = p_41_strides_0, weight = layers_20_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_41_cast_fp16")]; + tensor var_5892 = const()[name = tensor("op_5892"), val = tensor([1, 8, 128, 188])]; + tensor var_5893_cast_fp16 = reshape(shape = var_5892, x = q_with_bias_v_41_cast_fp16)[name = tensor("op_5893_cast_fp16")]; + tensor var_5894 = const()[name = tensor("op_5894"), val = tensor([1, 8, 128, -1])]; + tensor var_5895_cast_fp16 = reshape(shape = var_5894, x = p_41_cast_fp16)[name = tensor("op_5895_cast_fp16")]; + tensor matrix_bd_161_transpose_x_0 = const()[name = tensor("matrix_bd_161_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_161_transpose_y_0 = const()[name = tensor("matrix_bd_161_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_161_cast_fp16 = matmul(transpose_x = matrix_bd_161_transpose_x_0, transpose_y = matrix_bd_161_transpose_y_0, x = var_5893_cast_fp16, y = var_5895_cast_fp16)[name = tensor("matrix_bd_161_cast_fp16")]; + tensor matrix_bd_163_pad_0 = const()[name = tensor("matrix_bd_163_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_163_mode_0 = const()[name = tensor("matrix_bd_163_mode_0"), val = tensor("constant")]; + tensor const_230_to_fp16 = const()[name = tensor("const_230_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_163_cast_fp16 = pad(constant_val = const_230_to_fp16, mode = matrix_bd_163_mode_0, pad = matrix_bd_163_pad_0, x = matrix_bd_161_cast_fp16)[name = tensor("matrix_bd_163_cast_fp16")]; + tensor var_5904 = const()[name = tensor("op_5904"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_165_cast_fp16 = reshape(shape = var_5904, x = matrix_bd_163_cast_fp16)[name = tensor("matrix_bd_165_cast_fp16")]; + tensor var_5908_begin_0 = const()[name = tensor("op_5908_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5908_end_0 = const()[name = tensor("op_5908_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5908_end_mask_0 = const()[name = tensor("op_5908_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5908_cast_fp16 = slice_by_index(begin = var_5908_begin_0, end = var_5908_end_0, end_mask = var_5908_end_mask_0, x = matrix_bd_165_cast_fp16)[name = tensor("op_5908_cast_fp16")]; + tensor var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_167_cast_fp16 = reshape(shape = var_5909, x = var_5908_cast_fp16)[name = tensor("matrix_bd_167_cast_fp16")]; + tensor var_5914_begin_0 = const()[name = tensor("op_5914_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5914_end_0 = const()[name = tensor("op_5914_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5914_end_mask_0 = const()[name = tensor("op_5914_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5914_cast_fp16 = slice_by_index(begin = var_5914_begin_0, end = var_5914_end_0, end_mask = var_5914_end_mask_0, x = matrix_bd_167_cast_fp16)[name = tensor("op_5914_cast_fp16")]; + tensor var_5915_to_fp16 = const()[name = tensor("op_5915_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_41_cast_fp16 = mul(x = var_5914_cast_fp16, y = var_5915_to_fp16)[name = tensor("qk_mask_41_cast_fp16")]; + tensor var_5919 = const()[name = tensor("op_5919"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_5919, x = query_83_cast_fp16)[name = tensor("mh_q_41_cast_fp16")]; + tensor var_5921_to_fp16 = const()[name = tensor("op_5921_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5922_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_5921_to_fp16)[name = tensor("op_5922_cast_fp16")]; + tensor var_5925 = const()[name = tensor("op_5925"), val = tensor([1, 8, 128, 188])]; + tensor var_5926_cast_fp16 = reshape(shape = var_5925, x = key_41_cast_fp16)[name = tensor("op_5926_cast_fp16")]; + tensor mh_w_81_transpose_x_0 = const()[name = tensor("mh_w_81_transpose_x_0"), val = tensor(true)]; + tensor mh_w_81_transpose_y_0 = const()[name = tensor("mh_w_81_transpose_y_0"), val = tensor(false)]; + tensor mh_w_81_cast_fp16 = matmul(transpose_x = mh_w_81_transpose_x_0, transpose_y = mh_w_81_transpose_y_0, x = var_5922_cast_fp16, y = var_5926_cast_fp16)[name = tensor("mh_w_81_cast_fp16")]; + tensor mh_w_83_cast_fp16 = add(x = mh_w_81_cast_fp16, y = qk_mask_41_cast_fp16)[name = tensor("mh_w_83_cast_fp16")]; + tensor var_5930_cast_fp16 = softmax(axis = var_5771, x = mh_w_83_cast_fp16)[name = tensor("op_5930_cast_fp16")]; + tensor var_5931 = const()[name = tensor("op_5931"), val = tensor([1, 8, 128, 188])]; + tensor var_5932_cast_fp16 = reshape(shape = var_5931, x = value_41_cast_fp16)[name = tensor("op_5932_cast_fp16")]; + tensor attn_41_transpose_x_0 = const()[name = tensor("attn_41_transpose_x_0"), val = tensor(false)]; + tensor attn_41_transpose_y_0 = const()[name = tensor("attn_41_transpose_y_0"), val = tensor(true)]; + tensor attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = var_5932_cast_fp16, y = var_5930_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5935 = const()[name = tensor("op_5935"), val = tensor([1, 1024, 1, 188])]; + tensor input_543_cast_fp16 = reshape(shape = var_5935, x = attn_41_cast_fp16)[name = tensor("input_543_cast_fp16")]; + tensor obj_85_pad_type_0 = const()[name = tensor("obj_85_pad_type_0"), val = tensor("valid")]; + tensor obj_85_strides_0 = const()[name = tensor("obj_85_strides_0"), val = tensor([1, 1])]; + tensor obj_85_pad_0 = const()[name = tensor("obj_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_85_dilations_0 = const()[name = tensor("obj_85_dilations_0"), val = tensor([1, 1])]; + tensor obj_85_groups_0 = const()[name = tensor("obj_85_groups_0"), val = tensor(1)]; + tensor layers_20_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(391244096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392030592))), name = tensor("layers_20_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_85_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_85_dilations_0, groups = obj_85_groups_0, pad = obj_85_pad_0, pad_type = obj_85_pad_type_0, strides = obj_85_strides_0, weight = layers_20_self_attn_o_proj_weight_to_fp16_palettized, x = input_543_cast_fp16)[name = tensor("obj_85_cast_fp16")]; + tensor inputs_205_cast_fp16 = add(x = inputs_203_cast_fp16, y = obj_85_cast_fp16)[name = tensor("inputs_205_cast_fp16")]; + tensor out_205_axes_0 = const()[name = tensor("out_205_axes_0"), val = tensor([1])]; + tensor var_5953_to_fp16 = const()[name = tensor("op_5953_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_205_cast_fp16 = layer_norm(axes = out_205_axes_0, epsilon = var_5953_to_fp16, x = inputs_205_cast_fp16)[name = tensor("out_205_cast_fp16")]; + tensor input_545_gamma_0_to_fp16 = const()[name = tensor("input_545_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392030784)))]; + tensor input_545_beta_0_to_fp16 = const()[name = tensor("input_545_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392032896)))]; + tensor input_545_epsilon_0_to_fp16 = const()[name = tensor("input_545_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_545_cast_fp16 = batch_norm(beta = input_545_beta_0_to_fp16, epsilon = input_545_epsilon_0_to_fp16, gamma = input_545_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_205_cast_fp16)[name = tensor("input_545_cast_fp16")]; + tensor input_547_pad_type_0 = const()[name = tensor("input_547_pad_type_0"), val = tensor("valid")]; + tensor input_547_strides_0 = const()[name = tensor("input_547_strides_0"), val = tensor([1, 1])]; + tensor input_547_pad_0 = const()[name = tensor("input_547_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_547_dilations_0 = const()[name = tensor("input_547_dilations_0"), val = tensor([1, 1])]; + tensor input_547_groups_0 = const()[name = tensor("input_547_groups_0"), val = tensor(1)]; + tensor layers_20_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(392035008))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393607936))), name = tensor("layers_20_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_547_cast_fp16 = conv(dilations = input_547_dilations_0, groups = input_547_groups_0, pad = input_547_pad_0, pad_type = input_547_pad_type_0, strides = input_547_strides_0, weight = layers_20_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_545_cast_fp16)[name = tensor("input_547_cast_fp16")]; + tensor input_549_split_num_splits_0 = const()[name = tensor("input_549_split_num_splits_0"), val = tensor(2)]; + tensor input_549_split_axis_0 = const()[name = tensor("input_549_split_axis_0"), val = tensor(1)]; + tensor input_549_split_cast_fp16_0, tensor input_549_split_cast_fp16_1 = split(axis = input_549_split_axis_0, num_splits = input_549_split_num_splits_0, x = input_547_cast_fp16)[name = tensor("input_549_split_cast_fp16")]; + tensor input_549_split_1_sigmoid_cast_fp16 = sigmoid(x = input_549_split_cast_fp16_1)[name = tensor("input_549_split_1_sigmoid_cast_fp16")]; + tensor input_549_cast_fp16 = mul(x = input_549_split_cast_fp16_0, y = input_549_split_1_sigmoid_cast_fp16)[name = tensor("input_549_cast_fp16")]; + tensor input_551_pad_type_0 = const()[name = tensor("input_551_pad_type_0"), val = tensor("custom")]; + tensor input_551_pad_0 = const()[name = tensor("input_551_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_551_groups_0 = const()[name = tensor("input_551_groups_0"), val = tensor(1024)]; + tensor input_551_strides_0 = const()[name = tensor("input_551_strides_0"), val = tensor([1, 1])]; + tensor input_551_dilations_0 = const()[name = tensor("input_551_dilations_0"), val = tensor([1, 1])]; + tensor const_308_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393608128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393615104))), name = tensor("const_308_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_309_to_fp16 = const()[name = tensor("const_309_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393615296)))]; + tensor input_553_cast_fp16 = conv(bias = const_309_to_fp16, dilations = input_551_dilations_0, groups = input_551_groups_0, pad = input_551_pad_0, pad_type = input_551_pad_type_0, strides = input_551_strides_0, weight = const_308_to_fp16_palettized, x = input_549_cast_fp16)[name = tensor("input_553_cast_fp16")]; + tensor input_555_cast_fp16 = silu(x = input_553_cast_fp16)[name = tensor("input_555_cast_fp16")]; + tensor x_125_pad_type_0 = const()[name = tensor("x_125_pad_type_0"), val = tensor("valid")]; + tensor x_125_strides_0 = const()[name = tensor("x_125_strides_0"), val = tensor([1, 1])]; + tensor x_125_pad_0 = const()[name = tensor("x_125_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_125_dilations_0 = const()[name = tensor("x_125_dilations_0"), val = tensor([1, 1])]; + tensor x_125_groups_0 = const()[name = tensor("x_125_groups_0"), val = tensor(1)]; + tensor layers_20_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393617408))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394403904))), name = tensor("layers_20_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_125_cast_fp16 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = layers_20_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_555_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor inputs_207_cast_fp16 = add(x = inputs_205_cast_fp16, y = x_125_cast_fp16)[name = tensor("inputs_207_cast_fp16")]; + tensor out_207_axes_0 = const()[name = tensor("out_207_axes_0"), val = tensor([1])]; + tensor var_6001_to_fp16 = const()[name = tensor("op_6001_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_207_cast_fp16 = layer_norm(axes = out_207_axes_0, epsilon = var_6001_to_fp16, x = inputs_207_cast_fp16)[name = tensor("out_207_cast_fp16")]; + tensor input_557_gamma_0_to_fp16 = const()[name = tensor("input_557_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394404096)))]; + tensor input_557_beta_0_to_fp16 = const()[name = tensor("input_557_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394406208)))]; + tensor input_557_epsilon_0_to_fp16 = const()[name = tensor("input_557_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_557_cast_fp16 = batch_norm(beta = input_557_beta_0_to_fp16, epsilon = input_557_epsilon_0_to_fp16, gamma = input_557_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_207_cast_fp16)[name = tensor("input_557_cast_fp16")]; + tensor input_559_pad_type_0 = const()[name = tensor("input_559_pad_type_0"), val = tensor("valid")]; + tensor input_559_strides_0 = const()[name = tensor("input_559_strides_0"), val = tensor([1, 1])]; + tensor input_559_pad_0 = const()[name = tensor("input_559_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_559_dilations_0 = const()[name = tensor("input_559_dilations_0"), val = tensor([1, 1])]; + tensor input_559_groups_0 = const()[name = tensor("input_559_groups_0"), val = tensor(1)]; + tensor layers_20_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394408320))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397554112))), name = tensor("layers_20_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_559_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_559_dilations_0, groups = input_559_groups_0, pad = input_559_pad_0, pad_type = input_559_pad_type_0, strides = input_559_strides_0, weight = layers_20_feed_forward2_fc1_weight_to_fp16_palettized, x = input_557_cast_fp16)[name = tensor("input_559_cast_fp16")]; + tensor input_561_cast_fp16 = silu(x = input_559_cast_fp16)[name = tensor("input_561_cast_fp16")]; + tensor x_127_pad_type_0 = const()[name = tensor("x_127_pad_type_0"), val = tensor("valid")]; + tensor x_127_strides_0 = const()[name = tensor("x_127_strides_0"), val = tensor([1, 1])]; + tensor x_127_pad_0 = const()[name = tensor("x_127_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_127_dilations_0 = const()[name = tensor("x_127_dilations_0"), val = tensor([1, 1])]; + tensor x_127_groups_0 = const()[name = tensor("x_127_groups_0"), val = tensor(1)]; + tensor op_6029_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(397554304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400700096))), name = tensor("op_6029_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_6029_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_127_dilations_0, groups = x_127_groups_0, pad = x_127_pad_0, pad_type = x_127_pad_type_0, strides = x_127_strides_0, weight = op_6029_weight_0_to_fp16_palettized, x = input_561_cast_fp16)[name = tensor("op_6029_cast_fp16")]; + tensor inputs_209_cast_fp16 = add(x = inputs_207_cast_fp16, y = var_6029_cast_fp16)[name = tensor("inputs_209_cast_fp16")]; + tensor out_209_axes_0 = const()[name = tensor("out_209_axes_0"), val = tensor([1])]; + tensor var_6039_to_fp16 = const()[name = tensor("op_6039_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_209_cast_fp16 = layer_norm(axes = out_209_axes_0, epsilon = var_6039_to_fp16, x = inputs_209_cast_fp16)[name = tensor("out_209_cast_fp16")]; + tensor inputs_211_gamma_0_to_fp16 = const()[name = tensor("inputs_211_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400700288)))]; + tensor inputs_211_beta_0_to_fp16 = const()[name = tensor("inputs_211_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400702400)))]; + tensor inputs_211_epsilon_0_to_fp16 = const()[name = tensor("inputs_211_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_211_cast_fp16 = batch_norm(beta = inputs_211_beta_0_to_fp16, epsilon = inputs_211_epsilon_0_to_fp16, gamma = inputs_211_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_209_cast_fp16)[name = tensor("inputs_211_cast_fp16")]; + tensor var_6053 = const()[name = tensor("op_6053"), val = tensor(3)]; + tensor out_211_axes_0 = const()[name = tensor("out_211_axes_0"), val = tensor([1])]; + tensor var_6084_to_fp16 = const()[name = tensor("op_6084_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_211_cast_fp16 = layer_norm(axes = out_211_axes_0, epsilon = var_6084_to_fp16, x = inputs_211_cast_fp16)[name = tensor("out_211_cast_fp16")]; + tensor input_563_gamma_0_to_fp16 = const()[name = tensor("input_563_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400704512)))]; + tensor input_563_beta_0_to_fp16 = const()[name = tensor("input_563_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400706624)))]; + tensor input_563_epsilon_0_to_fp16 = const()[name = tensor("input_563_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_563_cast_fp16 = batch_norm(beta = input_563_beta_0_to_fp16, epsilon = input_563_epsilon_0_to_fp16, gamma = input_563_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_211_cast_fp16)[name = tensor("input_563_cast_fp16")]; + tensor input_565_pad_type_0 = const()[name = tensor("input_565_pad_type_0"), val = tensor("valid")]; + tensor input_565_strides_0 = const()[name = tensor("input_565_strides_0"), val = tensor([1, 1])]; + tensor input_565_pad_0 = const()[name = tensor("input_565_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_565_dilations_0 = const()[name = tensor("input_565_dilations_0"), val = tensor([1, 1])]; + tensor input_565_groups_0 = const()[name = tensor("input_565_groups_0"), val = tensor(1)]; + tensor layers_21_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(400708736))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403854528))), name = tensor("layers_21_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_565_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_565_dilations_0, groups = input_565_groups_0, pad = input_565_pad_0, pad_type = input_565_pad_type_0, strides = input_565_strides_0, weight = layers_21_feed_forward1_fc1_weight_to_fp16_palettized, x = input_563_cast_fp16)[name = tensor("input_565_cast_fp16")]; + tensor input_567_cast_fp16 = silu(x = input_565_cast_fp16)[name = tensor("input_567_cast_fp16")]; + tensor x_129_pad_type_0 = const()[name = tensor("x_129_pad_type_0"), val = tensor("valid")]; + tensor x_129_strides_0 = const()[name = tensor("x_129_strides_0"), val = tensor([1, 1])]; + tensor x_129_pad_0 = const()[name = tensor("x_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_129_dilations_0 = const()[name = tensor("x_129_dilations_0"), val = tensor([1, 1])]; + tensor x_129_groups_0 = const()[name = tensor("x_129_groups_0"), val = tensor(1)]; + tensor op_6112_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403854720))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407000512))), name = tensor("op_6112_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_6112_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = op_6112_weight_0_to_fp16_palettized, x = input_567_cast_fp16)[name = tensor("op_6112_cast_fp16")]; + tensor inputs_213_cast_fp16 = add(x = inputs_211_cast_fp16, y = var_6112_cast_fp16)[name = tensor("inputs_213_cast_fp16")]; + tensor out_213_axes_0 = const()[name = tensor("out_213_axes_0"), val = tensor([1])]; + tensor var_6122_to_fp16 = const()[name = tensor("op_6122_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_213_cast_fp16 = layer_norm(axes = out_213_axes_0, epsilon = var_6122_to_fp16, x = inputs_213_cast_fp16)[name = tensor("out_213_cast_fp16")]; + tensor obj_87_gamma_0_to_fp16 = const()[name = tensor("obj_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407000704)))]; + tensor obj_87_beta_0_to_fp16 = const()[name = tensor("obj_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407002816)))]; + tensor obj_87_epsilon_0_to_fp16 = const()[name = tensor("obj_87_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_87_cast_fp16 = batch_norm(beta = obj_87_beta_0_to_fp16, epsilon = obj_87_epsilon_0_to_fp16, gamma = obj_87_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_213_cast_fp16)[name = tensor("obj_87_cast_fp16")]; + tensor query_85_pad_type_0 = const()[name = tensor("query_85_pad_type_0"), val = tensor("valid")]; + tensor query_85_strides_0 = const()[name = tensor("query_85_strides_0"), val = tensor([1, 1])]; + tensor query_85_pad_0 = const()[name = tensor("query_85_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_85_dilations_0 = const()[name = tensor("query_85_dilations_0"), val = tensor([1, 1])]; + tensor query_85_groups_0 = const()[name = tensor("query_85_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407004928))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407791424))), name = tensor("layers_21_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_85_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_85_dilations_0, groups = query_85_groups_0, pad = query_85_pad_0, pad_type = query_85_pad_type_0, strides = query_85_strides_0, weight = layers_21_self_attn_q_proj_weight_to_fp16_palettized, x = obj_87_cast_fp16)[name = tensor("query_85_cast_fp16")]; + tensor key_43_pad_type_0 = const()[name = tensor("key_43_pad_type_0"), val = tensor("valid")]; + tensor key_43_strides_0 = const()[name = tensor("key_43_strides_0"), val = tensor([1, 1])]; + tensor key_43_pad_0 = const()[name = tensor("key_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_43_dilations_0 = const()[name = tensor("key_43_dilations_0"), val = tensor([1, 1])]; + tensor key_43_groups_0 = const()[name = tensor("key_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(407791616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408578112))), name = tensor("layers_21_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_43_cast_fp16 = conv(dilations = key_43_dilations_0, groups = key_43_groups_0, pad = key_43_pad_0, pad_type = key_43_pad_type_0, strides = key_43_strides_0, weight = layers_21_self_attn_k_proj_weight_to_fp16_palettized, x = obj_87_cast_fp16)[name = tensor("key_43_cast_fp16")]; + tensor value_43_pad_type_0 = const()[name = tensor("value_43_pad_type_0"), val = tensor("valid")]; + tensor value_43_strides_0 = const()[name = tensor("value_43_strides_0"), val = tensor([1, 1])]; + tensor value_43_pad_0 = const()[name = tensor("value_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_43_dilations_0 = const()[name = tensor("value_43_dilations_0"), val = tensor([1, 1])]; + tensor value_43_groups_0 = const()[name = tensor("value_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(408578304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409364800))), name = tensor("layers_21_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_43_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_43_dilations_0, groups = value_43_groups_0, pad = value_43_pad_0, pad_type = value_43_pad_type_0, strides = value_43_strides_0, weight = layers_21_self_attn_v_proj_weight_to_fp16_palettized, x = obj_87_cast_fp16)[name = tensor("value_43_cast_fp16")]; + tensor var_6160_to_fp16 = const()[name = tensor("op_6160_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409364992)))]; + tensor query_87_cast_fp16 = add(x = query_85_cast_fp16, y = var_6160_to_fp16)[name = tensor("query_87_cast_fp16")]; + tensor var_6163_to_fp16 = const()[name = tensor("op_6163_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409367104)))]; + tensor q_with_bias_v_43_cast_fp16 = add(x = query_85_cast_fp16, y = var_6163_to_fp16)[name = tensor("q_with_bias_v_43_cast_fp16")]; + tensor p_43_pad_type_0 = const()[name = tensor("p_43_pad_type_0"), val = tensor("valid")]; + tensor p_43_strides_0 = const()[name = tensor("p_43_strides_0"), val = tensor([1, 1])]; + tensor p_43_pad_0 = const()[name = tensor("p_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_43_dilations_0 = const()[name = tensor("p_43_dilations_0"), val = tensor([1, 1])]; + tensor p_43_groups_0 = const()[name = tensor("p_43_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(409369216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410155712))), name = tensor("layers_21_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_43_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_43_dilations_0, groups = p_43_groups_0, pad = p_43_pad_0, pad_type = p_43_pad_type_0, strides = p_43_strides_0, weight = layers_21_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_43_cast_fp16")]; + tensor var_6174 = const()[name = tensor("op_6174"), val = tensor([1, 8, 128, 188])]; + tensor var_6175_cast_fp16 = reshape(shape = var_6174, x = q_with_bias_v_43_cast_fp16)[name = tensor("op_6175_cast_fp16")]; + tensor var_6176 = const()[name = tensor("op_6176"), val = tensor([1, 8, 128, -1])]; + tensor var_6177_cast_fp16 = reshape(shape = var_6176, x = p_43_cast_fp16)[name = tensor("op_6177_cast_fp16")]; + tensor matrix_bd_169_transpose_x_0 = const()[name = tensor("matrix_bd_169_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_169_transpose_y_0 = const()[name = tensor("matrix_bd_169_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_169_cast_fp16 = matmul(transpose_x = matrix_bd_169_transpose_x_0, transpose_y = matrix_bd_169_transpose_y_0, x = var_6175_cast_fp16, y = var_6177_cast_fp16)[name = tensor("matrix_bd_169_cast_fp16")]; + tensor matrix_bd_171_pad_0 = const()[name = tensor("matrix_bd_171_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_171_mode_0 = const()[name = tensor("matrix_bd_171_mode_0"), val = tensor("constant")]; + tensor const_241_to_fp16 = const()[name = tensor("const_241_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_171_cast_fp16 = pad(constant_val = const_241_to_fp16, mode = matrix_bd_171_mode_0, pad = matrix_bd_171_pad_0, x = matrix_bd_169_cast_fp16)[name = tensor("matrix_bd_171_cast_fp16")]; + tensor var_6186 = const()[name = tensor("op_6186"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_173_cast_fp16 = reshape(shape = var_6186, x = matrix_bd_171_cast_fp16)[name = tensor("matrix_bd_173_cast_fp16")]; + tensor var_6190_begin_0 = const()[name = tensor("op_6190_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_6190_end_0 = const()[name = tensor("op_6190_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_6190_end_mask_0 = const()[name = tensor("op_6190_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6190_cast_fp16 = slice_by_index(begin = var_6190_begin_0, end = var_6190_end_0, end_mask = var_6190_end_mask_0, x = matrix_bd_173_cast_fp16)[name = tensor("op_6190_cast_fp16")]; + tensor var_6191 = const()[name = tensor("op_6191"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_175_cast_fp16 = reshape(shape = var_6191, x = var_6190_cast_fp16)[name = tensor("matrix_bd_175_cast_fp16")]; + tensor var_6196_begin_0 = const()[name = tensor("op_6196_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6196_end_0 = const()[name = tensor("op_6196_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_6196_end_mask_0 = const()[name = tensor("op_6196_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6196_cast_fp16 = slice_by_index(begin = var_6196_begin_0, end = var_6196_end_0, end_mask = var_6196_end_mask_0, x = matrix_bd_175_cast_fp16)[name = tensor("op_6196_cast_fp16")]; + tensor var_6197_to_fp16 = const()[name = tensor("op_6197_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_43_cast_fp16 = mul(x = var_6196_cast_fp16, y = var_6197_to_fp16)[name = tensor("qk_mask_43_cast_fp16")]; + tensor var_6201 = const()[name = tensor("op_6201"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_6201, x = query_87_cast_fp16)[name = tensor("mh_q_43_cast_fp16")]; + tensor var_6203_to_fp16 = const()[name = tensor("op_6203_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_6204_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_6203_to_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6207 = const()[name = tensor("op_6207"), val = tensor([1, 8, 128, 188])]; + tensor var_6208_cast_fp16 = reshape(shape = var_6207, x = key_43_cast_fp16)[name = tensor("op_6208_cast_fp16")]; + tensor mh_w_85_transpose_x_0 = const()[name = tensor("mh_w_85_transpose_x_0"), val = tensor(true)]; + tensor mh_w_85_transpose_y_0 = const()[name = tensor("mh_w_85_transpose_y_0"), val = tensor(false)]; + tensor mh_w_85_cast_fp16 = matmul(transpose_x = mh_w_85_transpose_x_0, transpose_y = mh_w_85_transpose_y_0, x = var_6204_cast_fp16, y = var_6208_cast_fp16)[name = tensor("mh_w_85_cast_fp16")]; + tensor mh_w_87_cast_fp16 = add(x = mh_w_85_cast_fp16, y = qk_mask_43_cast_fp16)[name = tensor("mh_w_87_cast_fp16")]; + tensor var_6212_cast_fp16 = softmax(axis = var_6053, x = mh_w_87_cast_fp16)[name = tensor("op_6212_cast_fp16")]; + tensor var_6213 = const()[name = tensor("op_6213"), val = tensor([1, 8, 128, 188])]; + tensor var_6214_cast_fp16 = reshape(shape = var_6213, x = value_43_cast_fp16)[name = tensor("op_6214_cast_fp16")]; + tensor attn_43_transpose_x_0 = const()[name = tensor("attn_43_transpose_x_0"), val = tensor(false)]; + tensor attn_43_transpose_y_0 = const()[name = tensor("attn_43_transpose_y_0"), val = tensor(true)]; + tensor attn_43_cast_fp16 = matmul(transpose_x = attn_43_transpose_x_0, transpose_y = attn_43_transpose_y_0, x = var_6214_cast_fp16, y = var_6212_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_6217 = const()[name = tensor("op_6217"), val = tensor([1, 1024, 1, 188])]; + tensor input_569_cast_fp16 = reshape(shape = var_6217, x = attn_43_cast_fp16)[name = tensor("input_569_cast_fp16")]; + tensor obj_89_pad_type_0 = const()[name = tensor("obj_89_pad_type_0"), val = tensor("valid")]; + tensor obj_89_strides_0 = const()[name = tensor("obj_89_strides_0"), val = tensor([1, 1])]; + tensor obj_89_pad_0 = const()[name = tensor("obj_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_89_dilations_0 = const()[name = tensor("obj_89_dilations_0"), val = tensor([1, 1])]; + tensor obj_89_groups_0 = const()[name = tensor("obj_89_groups_0"), val = tensor(1)]; + tensor layers_21_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410155904))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410942400))), name = tensor("layers_21_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_89_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_89_dilations_0, groups = obj_89_groups_0, pad = obj_89_pad_0, pad_type = obj_89_pad_type_0, strides = obj_89_strides_0, weight = layers_21_self_attn_o_proj_weight_to_fp16_palettized, x = input_569_cast_fp16)[name = tensor("obj_89_cast_fp16")]; + tensor inputs_215_cast_fp16 = add(x = inputs_213_cast_fp16, y = obj_89_cast_fp16)[name = tensor("inputs_215_cast_fp16")]; + tensor out_215_axes_0 = const()[name = tensor("out_215_axes_0"), val = tensor([1])]; + tensor var_6235_to_fp16 = const()[name = tensor("op_6235_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_215_cast_fp16 = layer_norm(axes = out_215_axes_0, epsilon = var_6235_to_fp16, x = inputs_215_cast_fp16)[name = tensor("out_215_cast_fp16")]; + tensor input_571_gamma_0_to_fp16 = const()[name = tensor("input_571_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410942592)))]; + tensor input_571_beta_0_to_fp16 = const()[name = tensor("input_571_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410944704)))]; + tensor input_571_epsilon_0_to_fp16 = const()[name = tensor("input_571_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_571_cast_fp16 = batch_norm(beta = input_571_beta_0_to_fp16, epsilon = input_571_epsilon_0_to_fp16, gamma = input_571_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_215_cast_fp16)[name = tensor("input_571_cast_fp16")]; + tensor input_573_pad_type_0 = const()[name = tensor("input_573_pad_type_0"), val = tensor("valid")]; + tensor input_573_strides_0 = const()[name = tensor("input_573_strides_0"), val = tensor([1, 1])]; + tensor input_573_pad_0 = const()[name = tensor("input_573_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_573_dilations_0 = const()[name = tensor("input_573_dilations_0"), val = tensor([1, 1])]; + tensor input_573_groups_0 = const()[name = tensor("input_573_groups_0"), val = tensor(1)]; + tensor layers_21_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410946816))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412519744))), name = tensor("layers_21_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_573_cast_fp16 = conv(dilations = input_573_dilations_0, groups = input_573_groups_0, pad = input_573_pad_0, pad_type = input_573_pad_type_0, strides = input_573_strides_0, weight = layers_21_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_571_cast_fp16)[name = tensor("input_573_cast_fp16")]; + tensor input_575_split_num_splits_0 = const()[name = tensor("input_575_split_num_splits_0"), val = tensor(2)]; + tensor input_575_split_axis_0 = const()[name = tensor("input_575_split_axis_0"), val = tensor(1)]; + tensor input_575_split_cast_fp16_0, tensor input_575_split_cast_fp16_1 = split(axis = input_575_split_axis_0, num_splits = input_575_split_num_splits_0, x = input_573_cast_fp16)[name = tensor("input_575_split_cast_fp16")]; + tensor input_575_split_1_sigmoid_cast_fp16 = sigmoid(x = input_575_split_cast_fp16_1)[name = tensor("input_575_split_1_sigmoid_cast_fp16")]; + tensor input_575_cast_fp16 = mul(x = input_575_split_cast_fp16_0, y = input_575_split_1_sigmoid_cast_fp16)[name = tensor("input_575_cast_fp16")]; + tensor input_577_pad_type_0 = const()[name = tensor("input_577_pad_type_0"), val = tensor("custom")]; + tensor input_577_pad_0 = const()[name = tensor("input_577_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_577_groups_0 = const()[name = tensor("input_577_groups_0"), val = tensor(1024)]; + tensor input_577_strides_0 = const()[name = tensor("input_577_strides_0"), val = tensor([1, 1])]; + tensor input_577_dilations_0 = const()[name = tensor("input_577_dilations_0"), val = tensor([1, 1])]; + tensor const_310_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412519936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412526912))), name = tensor("const_310_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_311_to_fp16 = const()[name = tensor("const_311_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412527104)))]; + tensor input_579_cast_fp16 = conv(bias = const_311_to_fp16, dilations = input_577_dilations_0, groups = input_577_groups_0, pad = input_577_pad_0, pad_type = input_577_pad_type_0, strides = input_577_strides_0, weight = const_310_to_fp16_palettized, x = input_575_cast_fp16)[name = tensor("input_579_cast_fp16")]; + tensor input_581_cast_fp16 = silu(x = input_579_cast_fp16)[name = tensor("input_581_cast_fp16")]; + tensor x_131_pad_type_0 = const()[name = tensor("x_131_pad_type_0"), val = tensor("valid")]; + tensor x_131_strides_0 = const()[name = tensor("x_131_strides_0"), val = tensor([1, 1])]; + tensor x_131_pad_0 = const()[name = tensor("x_131_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_131_dilations_0 = const()[name = tensor("x_131_dilations_0"), val = tensor([1, 1])]; + tensor x_131_groups_0 = const()[name = tensor("x_131_groups_0"), val = tensor(1)]; + tensor layers_21_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412529216))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413315712))), name = tensor("layers_21_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor 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 = layers_21_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_581_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor inputs_217_cast_fp16 = add(x = inputs_215_cast_fp16, y = x_131_cast_fp16)[name = tensor("inputs_217_cast_fp16")]; + tensor out_217_axes_0 = const()[name = tensor("out_217_axes_0"), val = tensor([1])]; + tensor var_6283_to_fp16 = const()[name = tensor("op_6283_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_217_cast_fp16 = layer_norm(axes = out_217_axes_0, epsilon = var_6283_to_fp16, x = inputs_217_cast_fp16)[name = tensor("out_217_cast_fp16")]; + tensor input_583_gamma_0_to_fp16 = const()[name = tensor("input_583_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413315904)))]; + tensor input_583_beta_0_to_fp16 = const()[name = tensor("input_583_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413318016)))]; + tensor input_583_epsilon_0_to_fp16 = const()[name = tensor("input_583_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_583_cast_fp16 = batch_norm(beta = input_583_beta_0_to_fp16, epsilon = input_583_epsilon_0_to_fp16, gamma = input_583_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_217_cast_fp16)[name = tensor("input_583_cast_fp16")]; + tensor input_585_pad_type_0 = const()[name = tensor("input_585_pad_type_0"), val = tensor("valid")]; + tensor input_585_strides_0 = const()[name = tensor("input_585_strides_0"), val = tensor([1, 1])]; + tensor input_585_pad_0 = const()[name = tensor("input_585_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_585_dilations_0 = const()[name = tensor("input_585_dilations_0"), val = tensor([1, 1])]; + tensor input_585_groups_0 = const()[name = tensor("input_585_groups_0"), val = tensor(1)]; + tensor layers_21_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(413320128))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416465920))), name = tensor("layers_21_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_585_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_585_dilations_0, groups = input_585_groups_0, pad = input_585_pad_0, pad_type = input_585_pad_type_0, strides = input_585_strides_0, weight = layers_21_feed_forward2_fc1_weight_to_fp16_palettized, x = input_583_cast_fp16)[name = tensor("input_585_cast_fp16")]; + tensor input_587_cast_fp16 = silu(x = input_585_cast_fp16)[name = tensor("input_587_cast_fp16")]; + tensor x_133_pad_type_0 = const()[name = tensor("x_133_pad_type_0"), val = tensor("valid")]; + tensor x_133_strides_0 = const()[name = tensor("x_133_strides_0"), val = tensor([1, 1])]; + tensor x_133_pad_0 = const()[name = tensor("x_133_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_133_dilations_0 = const()[name = tensor("x_133_dilations_0"), val = tensor([1, 1])]; + tensor x_133_groups_0 = const()[name = tensor("x_133_groups_0"), val = tensor(1)]; + tensor op_6311_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(416466112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419611904))), name = tensor("op_6311_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_6311_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_133_dilations_0, groups = x_133_groups_0, pad = x_133_pad_0, pad_type = x_133_pad_type_0, strides = x_133_strides_0, weight = op_6311_weight_0_to_fp16_palettized, x = input_587_cast_fp16)[name = tensor("op_6311_cast_fp16")]; + tensor inputs_219_cast_fp16 = add(x = inputs_217_cast_fp16, y = var_6311_cast_fp16)[name = tensor("inputs_219_cast_fp16")]; + tensor out_219_axes_0 = const()[name = tensor("out_219_axes_0"), val = tensor([1])]; + tensor var_6321_to_fp16 = const()[name = tensor("op_6321_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_219_cast_fp16 = layer_norm(axes = out_219_axes_0, epsilon = var_6321_to_fp16, x = inputs_219_cast_fp16)[name = tensor("out_219_cast_fp16")]; + tensor inputs_221_gamma_0_to_fp16 = const()[name = tensor("inputs_221_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419612096)))]; + tensor inputs_221_beta_0_to_fp16 = const()[name = tensor("inputs_221_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419614208)))]; + tensor inputs_221_epsilon_0_to_fp16 = const()[name = tensor("inputs_221_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_221_cast_fp16 = batch_norm(beta = inputs_221_beta_0_to_fp16, epsilon = inputs_221_epsilon_0_to_fp16, gamma = inputs_221_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_219_cast_fp16)[name = tensor("inputs_221_cast_fp16")]; + tensor var_6335 = const()[name = tensor("op_6335"), val = tensor(3)]; + tensor out_221_axes_0 = const()[name = tensor("out_221_axes_0"), val = tensor([1])]; + tensor var_6366_to_fp16 = const()[name = tensor("op_6366_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_221_cast_fp16 = layer_norm(axes = out_221_axes_0, epsilon = var_6366_to_fp16, x = inputs_221_cast_fp16)[name = tensor("out_221_cast_fp16")]; + tensor input_589_gamma_0_to_fp16 = const()[name = tensor("input_589_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419616320)))]; + tensor input_589_beta_0_to_fp16 = const()[name = tensor("input_589_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419618432)))]; + tensor input_589_epsilon_0_to_fp16 = const()[name = tensor("input_589_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_589_cast_fp16 = batch_norm(beta = input_589_beta_0_to_fp16, epsilon = input_589_epsilon_0_to_fp16, gamma = input_589_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_221_cast_fp16)[name = tensor("input_589_cast_fp16")]; + tensor input_591_pad_type_0 = const()[name = tensor("input_591_pad_type_0"), val = tensor("valid")]; + tensor input_591_strides_0 = const()[name = tensor("input_591_strides_0"), val = tensor([1, 1])]; + tensor input_591_pad_0 = const()[name = tensor("input_591_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_591_dilations_0 = const()[name = tensor("input_591_dilations_0"), val = tensor([1, 1])]; + tensor input_591_groups_0 = const()[name = tensor("input_591_groups_0"), val = tensor(1)]; + tensor layers_22_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(419620544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422766336))), name = tensor("layers_22_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_591_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_591_dilations_0, groups = input_591_groups_0, pad = input_591_pad_0, pad_type = input_591_pad_type_0, strides = input_591_strides_0, weight = layers_22_feed_forward1_fc1_weight_to_fp16_palettized, x = input_589_cast_fp16)[name = tensor("input_591_cast_fp16")]; + tensor input_593_cast_fp16 = silu(x = input_591_cast_fp16)[name = tensor("input_593_cast_fp16")]; + tensor x_135_pad_type_0 = const()[name = tensor("x_135_pad_type_0"), val = tensor("valid")]; + tensor x_135_strides_0 = const()[name = tensor("x_135_strides_0"), val = tensor([1, 1])]; + tensor x_135_pad_0 = const()[name = tensor("x_135_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_135_dilations_0 = const()[name = tensor("x_135_dilations_0"), val = tensor([1, 1])]; + tensor x_135_groups_0 = const()[name = tensor("x_135_groups_0"), val = tensor(1)]; + tensor op_6394_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(422766528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425912320))), name = tensor("op_6394_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_6394_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_135_dilations_0, groups = x_135_groups_0, pad = x_135_pad_0, pad_type = x_135_pad_type_0, strides = x_135_strides_0, weight = op_6394_weight_0_to_fp16_palettized, x = input_593_cast_fp16)[name = tensor("op_6394_cast_fp16")]; + tensor inputs_223_cast_fp16 = add(x = inputs_221_cast_fp16, y = var_6394_cast_fp16)[name = tensor("inputs_223_cast_fp16")]; + tensor out_223_axes_0 = const()[name = tensor("out_223_axes_0"), val = tensor([1])]; + tensor var_6404_to_fp16 = const()[name = tensor("op_6404_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_223_cast_fp16 = layer_norm(axes = out_223_axes_0, epsilon = var_6404_to_fp16, x = inputs_223_cast_fp16)[name = tensor("out_223_cast_fp16")]; + tensor obj_91_gamma_0_to_fp16 = const()[name = tensor("obj_91_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425912512)))]; + tensor obj_91_beta_0_to_fp16 = const()[name = tensor("obj_91_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425914624)))]; + tensor obj_91_epsilon_0_to_fp16 = const()[name = tensor("obj_91_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_91_cast_fp16 = batch_norm(beta = obj_91_beta_0_to_fp16, epsilon = obj_91_epsilon_0_to_fp16, gamma = obj_91_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_223_cast_fp16)[name = tensor("obj_91_cast_fp16")]; + tensor query_89_pad_type_0 = const()[name = tensor("query_89_pad_type_0"), val = tensor("valid")]; + tensor query_89_strides_0 = const()[name = tensor("query_89_strides_0"), val = tensor([1, 1])]; + tensor query_89_pad_0 = const()[name = tensor("query_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_89_dilations_0 = const()[name = tensor("query_89_dilations_0"), val = tensor([1, 1])]; + tensor query_89_groups_0 = const()[name = tensor("query_89_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(425916736))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426703232))), name = tensor("layers_22_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_89_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_89_dilations_0, groups = query_89_groups_0, pad = query_89_pad_0, pad_type = query_89_pad_type_0, strides = query_89_strides_0, weight = layers_22_self_attn_q_proj_weight_to_fp16_palettized, x = obj_91_cast_fp16)[name = tensor("query_89_cast_fp16")]; + tensor key_45_pad_type_0 = const()[name = tensor("key_45_pad_type_0"), val = tensor("valid")]; + tensor key_45_strides_0 = const()[name = tensor("key_45_strides_0"), val = tensor([1, 1])]; + tensor key_45_pad_0 = const()[name = tensor("key_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_45_dilations_0 = const()[name = tensor("key_45_dilations_0"), val = tensor([1, 1])]; + tensor key_45_groups_0 = const()[name = tensor("key_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426703424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427489920))), name = tensor("layers_22_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_45_cast_fp16 = conv(dilations = key_45_dilations_0, groups = key_45_groups_0, pad = key_45_pad_0, pad_type = key_45_pad_type_0, strides = key_45_strides_0, weight = layers_22_self_attn_k_proj_weight_to_fp16_palettized, x = obj_91_cast_fp16)[name = tensor("key_45_cast_fp16")]; + tensor value_45_pad_type_0 = const()[name = tensor("value_45_pad_type_0"), val = tensor("valid")]; + tensor value_45_strides_0 = const()[name = tensor("value_45_strides_0"), val = tensor([1, 1])]; + tensor value_45_pad_0 = const()[name = tensor("value_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_45_dilations_0 = const()[name = tensor("value_45_dilations_0"), val = tensor([1, 1])]; + tensor value_45_groups_0 = const()[name = tensor("value_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(427490112))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428276608))), name = tensor("layers_22_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_45_dilations_0, groups = value_45_groups_0, pad = value_45_pad_0, pad_type = value_45_pad_type_0, strides = value_45_strides_0, weight = layers_22_self_attn_v_proj_weight_to_fp16_palettized, x = obj_91_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_6442_to_fp16 = const()[name = tensor("op_6442_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428276800)))]; + tensor query_91_cast_fp16 = add(x = query_89_cast_fp16, y = var_6442_to_fp16)[name = tensor("query_91_cast_fp16")]; + tensor var_6445_to_fp16 = const()[name = tensor("op_6445_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428278912)))]; + tensor q_with_bias_v_45_cast_fp16 = add(x = query_89_cast_fp16, y = var_6445_to_fp16)[name = tensor("q_with_bias_v_45_cast_fp16")]; + tensor p_45_pad_type_0 = const()[name = tensor("p_45_pad_type_0"), val = tensor("valid")]; + tensor p_45_strides_0 = const()[name = tensor("p_45_strides_0"), val = tensor([1, 1])]; + tensor p_45_pad_0 = const()[name = tensor("p_45_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_45_dilations_0 = const()[name = tensor("p_45_dilations_0"), val = tensor([1, 1])]; + tensor p_45_groups_0 = const()[name = tensor("p_45_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(428281024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429067520))), name = tensor("layers_22_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_45_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_45_dilations_0, groups = p_45_groups_0, pad = p_45_pad_0, pad_type = p_45_pad_type_0, strides = p_45_strides_0, weight = layers_22_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_45_cast_fp16")]; + tensor var_6456 = const()[name = tensor("op_6456"), val = tensor([1, 8, 128, 188])]; + tensor var_6457_cast_fp16 = reshape(shape = var_6456, x = q_with_bias_v_45_cast_fp16)[name = tensor("op_6457_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([1, 8, 128, -1])]; + tensor var_6459_cast_fp16 = reshape(shape = var_6458, x = p_45_cast_fp16)[name = tensor("op_6459_cast_fp16")]; + tensor matrix_bd_177_transpose_x_0 = const()[name = tensor("matrix_bd_177_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_177_transpose_y_0 = const()[name = tensor("matrix_bd_177_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_177_cast_fp16 = matmul(transpose_x = matrix_bd_177_transpose_x_0, transpose_y = matrix_bd_177_transpose_y_0, x = var_6457_cast_fp16, y = var_6459_cast_fp16)[name = tensor("matrix_bd_177_cast_fp16")]; + tensor matrix_bd_179_pad_0 = const()[name = tensor("matrix_bd_179_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_179_mode_0 = const()[name = tensor("matrix_bd_179_mode_0"), val = tensor("constant")]; + tensor const_252_to_fp16 = const()[name = tensor("const_252_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_179_cast_fp16 = pad(constant_val = const_252_to_fp16, mode = matrix_bd_179_mode_0, pad = matrix_bd_179_pad_0, x = matrix_bd_177_cast_fp16)[name = tensor("matrix_bd_179_cast_fp16")]; + tensor var_6468 = const()[name = tensor("op_6468"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_181_cast_fp16 = reshape(shape = var_6468, x = matrix_bd_179_cast_fp16)[name = tensor("matrix_bd_181_cast_fp16")]; + tensor var_6472_begin_0 = const()[name = tensor("op_6472_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_6472_end_0 = const()[name = tensor("op_6472_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_6472_end_mask_0 = const()[name = tensor("op_6472_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6472_cast_fp16 = slice_by_index(begin = var_6472_begin_0, end = var_6472_end_0, end_mask = var_6472_end_mask_0, x = matrix_bd_181_cast_fp16)[name = tensor("op_6472_cast_fp16")]; + tensor var_6473 = const()[name = tensor("op_6473"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_183_cast_fp16 = reshape(shape = var_6473, x = var_6472_cast_fp16)[name = tensor("matrix_bd_183_cast_fp16")]; + tensor var_6478_begin_0 = const()[name = tensor("op_6478_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6478_end_0 = const()[name = tensor("op_6478_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_6478_end_mask_0 = const()[name = tensor("op_6478_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6478_cast_fp16 = slice_by_index(begin = var_6478_begin_0, end = var_6478_end_0, end_mask = var_6478_end_mask_0, x = matrix_bd_183_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6479_to_fp16 = const()[name = tensor("op_6479_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_45_cast_fp16 = mul(x = var_6478_cast_fp16, y = var_6479_to_fp16)[name = tensor("qk_mask_45_cast_fp16")]; + tensor var_6483 = const()[name = tensor("op_6483"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_6483, x = query_91_cast_fp16)[name = tensor("mh_q_45_cast_fp16")]; + tensor var_6485_to_fp16 = const()[name = tensor("op_6485_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_6486_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_6485_to_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6489 = const()[name = tensor("op_6489"), val = tensor([1, 8, 128, 188])]; + tensor var_6490_cast_fp16 = reshape(shape = var_6489, x = key_45_cast_fp16)[name = tensor("op_6490_cast_fp16")]; + tensor mh_w_89_transpose_x_0 = const()[name = tensor("mh_w_89_transpose_x_0"), val = tensor(true)]; + tensor mh_w_89_transpose_y_0 = const()[name = tensor("mh_w_89_transpose_y_0"), val = tensor(false)]; + tensor mh_w_89_cast_fp16 = matmul(transpose_x = mh_w_89_transpose_x_0, transpose_y = mh_w_89_transpose_y_0, x = var_6486_cast_fp16, y = var_6490_cast_fp16)[name = tensor("mh_w_89_cast_fp16")]; + tensor mh_w_91_cast_fp16 = add(x = mh_w_89_cast_fp16, y = qk_mask_45_cast_fp16)[name = tensor("mh_w_91_cast_fp16")]; + tensor var_6494_cast_fp16 = softmax(axis = var_6335, x = mh_w_91_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor var_6495 = const()[name = tensor("op_6495"), val = tensor([1, 8, 128, 188])]; + tensor var_6496_cast_fp16 = reshape(shape = var_6495, x = value_45_cast_fp16)[name = tensor("op_6496_cast_fp16")]; + tensor attn_45_transpose_x_0 = const()[name = tensor("attn_45_transpose_x_0"), val = tensor(false)]; + tensor attn_45_transpose_y_0 = const()[name = tensor("attn_45_transpose_y_0"), val = tensor(true)]; + tensor attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = var_6496_cast_fp16, y = var_6494_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_6499 = const()[name = tensor("op_6499"), val = tensor([1, 1024, 1, 188])]; + tensor input_595_cast_fp16 = reshape(shape = var_6499, x = attn_45_cast_fp16)[name = tensor("input_595_cast_fp16")]; + tensor obj_93_pad_type_0 = const()[name = tensor("obj_93_pad_type_0"), val = tensor("valid")]; + tensor obj_93_strides_0 = const()[name = tensor("obj_93_strides_0"), val = tensor([1, 1])]; + tensor obj_93_pad_0 = const()[name = tensor("obj_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_93_dilations_0 = const()[name = tensor("obj_93_dilations_0"), val = tensor([1, 1])]; + tensor obj_93_groups_0 = const()[name = tensor("obj_93_groups_0"), val = tensor(1)]; + tensor layers_22_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429067712))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429854208))), name = tensor("layers_22_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_93_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_93_dilations_0, groups = obj_93_groups_0, pad = obj_93_pad_0, pad_type = obj_93_pad_type_0, strides = obj_93_strides_0, weight = layers_22_self_attn_o_proj_weight_to_fp16_palettized, x = input_595_cast_fp16)[name = tensor("obj_93_cast_fp16")]; + tensor inputs_225_cast_fp16 = add(x = inputs_223_cast_fp16, y = obj_93_cast_fp16)[name = tensor("inputs_225_cast_fp16")]; + tensor out_225_axes_0 = const()[name = tensor("out_225_axes_0"), val = tensor([1])]; + tensor var_6517_to_fp16 = const()[name = tensor("op_6517_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_225_cast_fp16 = layer_norm(axes = out_225_axes_0, epsilon = var_6517_to_fp16, x = inputs_225_cast_fp16)[name = tensor("out_225_cast_fp16")]; + tensor input_597_gamma_0_to_fp16 = const()[name = tensor("input_597_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429854400)))]; + tensor input_597_beta_0_to_fp16 = const()[name = tensor("input_597_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429856512)))]; + tensor input_597_epsilon_0_to_fp16 = const()[name = tensor("input_597_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_597_cast_fp16 = batch_norm(beta = input_597_beta_0_to_fp16, epsilon = input_597_epsilon_0_to_fp16, gamma = input_597_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_225_cast_fp16)[name = tensor("input_597_cast_fp16")]; + tensor input_599_pad_type_0 = const()[name = tensor("input_599_pad_type_0"), val = tensor("valid")]; + tensor input_599_strides_0 = const()[name = tensor("input_599_strides_0"), val = tensor([1, 1])]; + tensor input_599_pad_0 = const()[name = tensor("input_599_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_599_dilations_0 = const()[name = tensor("input_599_dilations_0"), val = tensor([1, 1])]; + tensor input_599_groups_0 = const()[name = tensor("input_599_groups_0"), val = tensor(1)]; + tensor layers_22_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429858624))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431431552))), name = tensor("layers_22_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_599_cast_fp16 = conv(dilations = input_599_dilations_0, groups = input_599_groups_0, pad = input_599_pad_0, pad_type = input_599_pad_type_0, strides = input_599_strides_0, weight = layers_22_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_597_cast_fp16)[name = tensor("input_599_cast_fp16")]; + tensor input_601_split_num_splits_0 = const()[name = tensor("input_601_split_num_splits_0"), val = tensor(2)]; + tensor input_601_split_axis_0 = const()[name = tensor("input_601_split_axis_0"), val = tensor(1)]; + tensor input_601_split_cast_fp16_0, tensor input_601_split_cast_fp16_1 = split(axis = input_601_split_axis_0, num_splits = input_601_split_num_splits_0, x = input_599_cast_fp16)[name = tensor("input_601_split_cast_fp16")]; + tensor input_601_split_1_sigmoid_cast_fp16 = sigmoid(x = input_601_split_cast_fp16_1)[name = tensor("input_601_split_1_sigmoid_cast_fp16")]; + tensor input_601_cast_fp16 = mul(x = input_601_split_cast_fp16_0, y = input_601_split_1_sigmoid_cast_fp16)[name = tensor("input_601_cast_fp16")]; + tensor input_603_pad_type_0 = const()[name = tensor("input_603_pad_type_0"), val = tensor("custom")]; + tensor input_603_pad_0 = const()[name = tensor("input_603_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_603_groups_0 = const()[name = tensor("input_603_groups_0"), val = tensor(1024)]; + tensor input_603_strides_0 = const()[name = tensor("input_603_strides_0"), val = tensor([1, 1])]; + tensor input_603_dilations_0 = const()[name = tensor("input_603_dilations_0"), val = tensor([1, 1])]; + tensor const_312_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431431744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431438720))), name = tensor("const_312_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_313_to_fp16 = const()[name = tensor("const_313_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431438912)))]; + tensor input_605_cast_fp16 = conv(bias = const_313_to_fp16, dilations = input_603_dilations_0, groups = input_603_groups_0, pad = input_603_pad_0, pad_type = input_603_pad_type_0, strides = input_603_strides_0, weight = const_312_to_fp16_palettized, x = input_601_cast_fp16)[name = tensor("input_605_cast_fp16")]; + tensor input_607_cast_fp16 = silu(x = input_605_cast_fp16)[name = tensor("input_607_cast_fp16")]; + tensor x_137_pad_type_0 = const()[name = tensor("x_137_pad_type_0"), val = tensor("valid")]; + tensor x_137_strides_0 = const()[name = tensor("x_137_strides_0"), val = tensor([1, 1])]; + tensor x_137_pad_0 = const()[name = tensor("x_137_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_137_dilations_0 = const()[name = tensor("x_137_dilations_0"), val = tensor([1, 1])]; + tensor x_137_groups_0 = const()[name = tensor("x_137_groups_0"), val = tensor(1)]; + tensor layers_22_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431441024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(432227520))), name = tensor("layers_22_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_137_cast_fp16 = conv(dilations = x_137_dilations_0, groups = x_137_groups_0, pad = x_137_pad_0, pad_type = x_137_pad_type_0, strides = x_137_strides_0, weight = layers_22_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_607_cast_fp16)[name = tensor("x_137_cast_fp16")]; + tensor inputs_227_cast_fp16 = add(x = inputs_225_cast_fp16, y = x_137_cast_fp16)[name = tensor("inputs_227_cast_fp16")]; + tensor out_227_axes_0 = const()[name = tensor("out_227_axes_0"), val = tensor([1])]; + tensor var_6565_to_fp16 = const()[name = tensor("op_6565_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_227_cast_fp16 = layer_norm(axes = out_227_axes_0, epsilon = var_6565_to_fp16, x = inputs_227_cast_fp16)[name = tensor("out_227_cast_fp16")]; + tensor input_609_gamma_0_to_fp16 = const()[name = tensor("input_609_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(432227712)))]; + tensor input_609_beta_0_to_fp16 = const()[name = tensor("input_609_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(432229824)))]; + tensor input_609_epsilon_0_to_fp16 = const()[name = tensor("input_609_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_609_cast_fp16 = batch_norm(beta = input_609_beta_0_to_fp16, epsilon = input_609_epsilon_0_to_fp16, gamma = input_609_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_227_cast_fp16)[name = tensor("input_609_cast_fp16")]; + tensor input_611_pad_type_0 = const()[name = tensor("input_611_pad_type_0"), val = tensor("valid")]; + tensor input_611_strides_0 = const()[name = tensor("input_611_strides_0"), val = tensor([1, 1])]; + tensor input_611_pad_0 = const()[name = tensor("input_611_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_611_dilations_0 = const()[name = tensor("input_611_dilations_0"), val = tensor([1, 1])]; + tensor input_611_groups_0 = const()[name = tensor("input_611_groups_0"), val = tensor(1)]; + tensor layers_22_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(432231936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435377728))), name = tensor("layers_22_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_611_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_611_dilations_0, groups = input_611_groups_0, pad = input_611_pad_0, pad_type = input_611_pad_type_0, strides = input_611_strides_0, weight = layers_22_feed_forward2_fc1_weight_to_fp16_palettized, x = input_609_cast_fp16)[name = tensor("input_611_cast_fp16")]; + tensor input_613_cast_fp16 = silu(x = input_611_cast_fp16)[name = tensor("input_613_cast_fp16")]; + tensor x_139_pad_type_0 = const()[name = tensor("x_139_pad_type_0"), val = tensor("valid")]; + tensor x_139_strides_0 = const()[name = tensor("x_139_strides_0"), val = tensor([1, 1])]; + tensor x_139_pad_0 = const()[name = tensor("x_139_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_139_dilations_0 = const()[name = tensor("x_139_dilations_0"), val = tensor([1, 1])]; + tensor x_139_groups_0 = const()[name = tensor("x_139_groups_0"), val = tensor(1)]; + tensor op_6593_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435377920))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438523712))), name = tensor("op_6593_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_6593_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_139_dilations_0, groups = x_139_groups_0, pad = x_139_pad_0, pad_type = x_139_pad_type_0, strides = x_139_strides_0, weight = op_6593_weight_0_to_fp16_palettized, x = input_613_cast_fp16)[name = tensor("op_6593_cast_fp16")]; + tensor inputs_229_cast_fp16 = add(x = inputs_227_cast_fp16, y = var_6593_cast_fp16)[name = tensor("inputs_229_cast_fp16")]; + tensor out_229_axes_0 = const()[name = tensor("out_229_axes_0"), val = tensor([1])]; + tensor var_6603_to_fp16 = const()[name = tensor("op_6603_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_229_cast_fp16 = layer_norm(axes = out_229_axes_0, epsilon = var_6603_to_fp16, x = inputs_229_cast_fp16)[name = tensor("out_229_cast_fp16")]; + tensor inputs_231_gamma_0_to_fp16 = const()[name = tensor("inputs_231_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438523904)))]; + tensor inputs_231_beta_0_to_fp16 = const()[name = tensor("inputs_231_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438526016)))]; + tensor inputs_231_epsilon_0_to_fp16 = const()[name = tensor("inputs_231_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor inputs_231_cast_fp16 = batch_norm(beta = inputs_231_beta_0_to_fp16, epsilon = inputs_231_epsilon_0_to_fp16, gamma = inputs_231_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_229_cast_fp16)[name = tensor("inputs_231_cast_fp16")]; + tensor var_6617 = const()[name = tensor("op_6617"), val = tensor(3)]; + tensor out_231_axes_0 = const()[name = tensor("out_231_axes_0"), val = tensor([1])]; + tensor var_6648_to_fp16 = const()[name = tensor("op_6648_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_231_cast_fp16 = layer_norm(axes = out_231_axes_0, epsilon = var_6648_to_fp16, x = inputs_231_cast_fp16)[name = tensor("out_231_cast_fp16")]; + tensor input_615_gamma_0_to_fp16 = const()[name = tensor("input_615_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438528128)))]; + tensor input_615_beta_0_to_fp16 = const()[name = tensor("input_615_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438530240)))]; + tensor input_615_epsilon_0_to_fp16 = const()[name = tensor("input_615_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_615_cast_fp16 = batch_norm(beta = input_615_beta_0_to_fp16, epsilon = input_615_epsilon_0_to_fp16, gamma = input_615_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_231_cast_fp16)[name = tensor("input_615_cast_fp16")]; + tensor input_617_pad_type_0 = const()[name = tensor("input_617_pad_type_0"), val = tensor("valid")]; + tensor input_617_strides_0 = const()[name = tensor("input_617_strides_0"), val = tensor([1, 1])]; + tensor input_617_pad_0 = const()[name = tensor("input_617_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_617_dilations_0 = const()[name = tensor("input_617_dilations_0"), val = tensor([1, 1])]; + tensor input_617_groups_0 = const()[name = tensor("input_617_groups_0"), val = tensor(1)]; + tensor layers_23_feed_forward1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(438532352))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441678144))), name = tensor("layers_23_feed_forward1_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_617_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_617_dilations_0, groups = input_617_groups_0, pad = input_617_pad_0, pad_type = input_617_pad_type_0, strides = input_617_strides_0, weight = layers_23_feed_forward1_fc1_weight_to_fp16_palettized, x = input_615_cast_fp16)[name = tensor("input_617_cast_fp16")]; + tensor input_619_cast_fp16 = silu(x = input_617_cast_fp16)[name = tensor("input_619_cast_fp16")]; + tensor x_141_pad_type_0 = const()[name = tensor("x_141_pad_type_0"), val = tensor("valid")]; + tensor x_141_strides_0 = const()[name = tensor("x_141_strides_0"), val = tensor([1, 1])]; + tensor x_141_pad_0 = const()[name = tensor("x_141_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_141_dilations_0 = const()[name = tensor("x_141_dilations_0"), val = tensor([1, 1])]; + tensor x_141_groups_0 = const()[name = tensor("x_141_groups_0"), val = tensor(1)]; + tensor op_6676_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(441678336))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444824128))), name = tensor("op_6676_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_6676_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_141_dilations_0, groups = x_141_groups_0, pad = x_141_pad_0, pad_type = x_141_pad_type_0, strides = x_141_strides_0, weight = op_6676_weight_0_to_fp16_palettized, x = input_619_cast_fp16)[name = tensor("op_6676_cast_fp16")]; + tensor inputs_233_cast_fp16 = add(x = inputs_231_cast_fp16, y = var_6676_cast_fp16)[name = tensor("inputs_233_cast_fp16")]; + tensor out_233_axes_0 = const()[name = tensor("out_233_axes_0"), val = tensor([1])]; + tensor var_6686_to_fp16 = const()[name = tensor("op_6686_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_233_cast_fp16 = layer_norm(axes = out_233_axes_0, epsilon = var_6686_to_fp16, x = inputs_233_cast_fp16)[name = tensor("out_233_cast_fp16")]; + tensor obj_95_gamma_0_to_fp16 = const()[name = tensor("obj_95_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444824320)))]; + tensor obj_95_beta_0_to_fp16 = const()[name = tensor("obj_95_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444826432)))]; + tensor obj_95_epsilon_0_to_fp16 = const()[name = tensor("obj_95_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor obj_95_cast_fp16 = batch_norm(beta = obj_95_beta_0_to_fp16, epsilon = obj_95_epsilon_0_to_fp16, gamma = obj_95_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_233_cast_fp16)[name = tensor("obj_95_cast_fp16")]; + tensor query_93_pad_type_0 = const()[name = tensor("query_93_pad_type_0"), val = tensor("valid")]; + tensor query_93_strides_0 = const()[name = tensor("query_93_strides_0"), val = tensor([1, 1])]; + tensor query_93_pad_0 = const()[name = tensor("query_93_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor query_93_dilations_0 = const()[name = tensor("query_93_dilations_0"), val = tensor([1, 1])]; + tensor query_93_groups_0 = const()[name = tensor("query_93_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(444828544))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445615040))), name = tensor("layers_23_self_attn_q_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor query_93_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = query_93_dilations_0, groups = query_93_groups_0, pad = query_93_pad_0, pad_type = query_93_pad_type_0, strides = query_93_strides_0, weight = layers_23_self_attn_q_proj_weight_to_fp16_palettized, x = obj_95_cast_fp16)[name = tensor("query_93_cast_fp16")]; + tensor key_pad_type_0 = const()[name = tensor("key_pad_type_0"), val = tensor("valid")]; + tensor key_strides_0 = const()[name = tensor("key_strides_0"), val = tensor([1, 1])]; + tensor key_pad_0 = const()[name = tensor("key_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor key_dilations_0 = const()[name = tensor("key_dilations_0"), val = tensor([1, 1])]; + tensor key_groups_0 = const()[name = tensor("key_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445615232))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446401728))), name = tensor("layers_23_self_attn_k_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor key_cast_fp16 = conv(dilations = key_dilations_0, groups = key_groups_0, pad = key_pad_0, pad_type = key_pad_type_0, strides = key_strides_0, weight = layers_23_self_attn_k_proj_weight_to_fp16_palettized, x = obj_95_cast_fp16)[name = tensor("key_cast_fp16")]; + tensor value_pad_type_0 = const()[name = tensor("value_pad_type_0"), val = tensor("valid")]; + tensor value_strides_0 = const()[name = tensor("value_strides_0"), val = tensor([1, 1])]; + tensor value_pad_0 = const()[name = tensor("value_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor value_dilations_0 = const()[name = tensor("value_dilations_0"), val = tensor([1, 1])]; + tensor value_groups_0 = const()[name = tensor("value_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446401920))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447188416))), name = tensor("layers_23_self_attn_v_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor value_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = value_dilations_0, groups = value_groups_0, pad = value_pad_0, pad_type = value_pad_type_0, strides = value_strides_0, weight = layers_23_self_attn_v_proj_weight_to_fp16_palettized, x = obj_95_cast_fp16)[name = tensor("value_cast_fp16")]; + tensor var_6724_to_fp16 = const()[name = tensor("op_6724_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447188608)))]; + tensor query_cast_fp16 = add(x = query_93_cast_fp16, y = var_6724_to_fp16)[name = tensor("query_cast_fp16")]; + tensor var_6727_to_fp16 = const()[name = tensor("op_6727_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447190720)))]; + tensor q_with_bias_v_cast_fp16 = add(x = query_93_cast_fp16, y = var_6727_to_fp16)[name = tensor("q_with_bias_v_cast_fp16")]; + tensor p_pad_type_0 = const()[name = tensor("p_pad_type_0"), val = tensor("valid")]; + tensor p_strides_0 = const()[name = tensor("p_strides_0"), val = tensor([1, 1])]; + tensor p_pad_0 = const()[name = tensor("p_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor p_dilations_0 = const()[name = tensor("p_dilations_0"), val = tensor([1, 1])]; + tensor p_groups_0 = const()[name = tensor("p_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_linear_pos_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447192832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447979328))), name = tensor("layers_23_self_attn_linear_pos_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor p_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = p_dilations_0, groups = p_groups_0, pad = p_pad_0, pad_type = p_pad_type_0, strides = p_strides_0, weight = layers_23_self_attn_linear_pos_weight_to_fp16_palettized, x = obj_3_cast_fp16)[name = tensor("p_cast_fp16")]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 8, 128, 188])]; + tensor var_6739_cast_fp16 = reshape(shape = var_6738, x = q_with_bias_v_cast_fp16)[name = tensor("op_6739_cast_fp16")]; + tensor var_6740 = const()[name = tensor("op_6740"), val = tensor([1, 8, 128, -1])]; + tensor var_6741_cast_fp16 = reshape(shape = var_6740, x = p_cast_fp16)[name = tensor("op_6741_cast_fp16")]; + tensor matrix_bd_185_transpose_x_0 = const()[name = tensor("matrix_bd_185_transpose_x_0"), val = tensor(true)]; + tensor matrix_bd_185_transpose_y_0 = const()[name = tensor("matrix_bd_185_transpose_y_0"), val = tensor(false)]; + tensor matrix_bd_185_cast_fp16 = matmul(transpose_x = matrix_bd_185_transpose_x_0, transpose_y = matrix_bd_185_transpose_y_0, x = var_6739_cast_fp16, y = var_6741_cast_fp16)[name = tensor("matrix_bd_185_cast_fp16")]; + tensor matrix_bd_187_pad_0 = const()[name = tensor("matrix_bd_187_pad_0"), val = tensor([0, 0, 0, 0, 0, 0, 1, 0])]; + tensor matrix_bd_187_mode_0 = const()[name = tensor("matrix_bd_187_mode_0"), val = tensor("constant")]; + tensor const_263_to_fp16 = const()[name = tensor("const_263_to_fp16"), val = tensor(0x0p+0)]; + tensor matrix_bd_187_cast_fp16 = pad(constant_val = const_263_to_fp16, mode = matrix_bd_187_mode_0, pad = matrix_bd_187_pad_0, x = matrix_bd_185_cast_fp16)[name = tensor("matrix_bd_187_cast_fp16")]; + tensor var_6750 = const()[name = tensor("op_6750"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_189_cast_fp16 = reshape(shape = var_6750, x = matrix_bd_187_cast_fp16)[name = tensor("matrix_bd_189_cast_fp16")]; + tensor var_6754_begin_0 = const()[name = tensor("op_6754_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_6754_end_0 = const()[name = tensor("op_6754_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_6754_end_mask_0 = const()[name = tensor("op_6754_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6754_cast_fp16 = slice_by_index(begin = var_6754_begin_0, end = var_6754_end_0, end_mask = var_6754_end_mask_0, x = matrix_bd_189_cast_fp16)[name = tensor("op_6754_cast_fp16")]; + tensor var_6755 = const()[name = tensor("op_6755"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_cast_fp16 = reshape(shape = var_6755, x = var_6754_cast_fp16)[name = tensor("matrix_bd_cast_fp16")]; + tensor var_6760_begin_0 = const()[name = tensor("op_6760_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6760_end_0 = const()[name = tensor("op_6760_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_6760_end_mask_0 = const()[name = tensor("op_6760_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6760_cast_fp16 = slice_by_index(begin = var_6760_begin_0, end = var_6760_end_0, end_mask = var_6760_end_mask_0, x = matrix_bd_cast_fp16)[name = tensor("op_6760_cast_fp16")]; + tensor var_6761_to_fp16 = const()[name = tensor("op_6761_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_cast_fp16 = mul(x = var_6760_cast_fp16, y = var_6761_to_fp16)[name = tensor("qk_mask_cast_fp16")]; + tensor var_6765 = const()[name = tensor("op_6765"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_cast_fp16 = reshape(shape = var_6765, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_6767_to_fp16 = const()[name = tensor("op_6767_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_6768_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_6767_to_fp16)[name = tensor("op_6768_cast_fp16")]; + tensor var_6771 = const()[name = tensor("op_6771"), val = tensor([1, 8, 128, 188])]; + tensor var_6772_cast_fp16 = reshape(shape = var_6771, x = key_cast_fp16)[name = tensor("op_6772_cast_fp16")]; + tensor mh_w_93_transpose_x_0 = const()[name = tensor("mh_w_93_transpose_x_0"), val = tensor(true)]; + tensor mh_w_93_transpose_y_0 = const()[name = tensor("mh_w_93_transpose_y_0"), val = tensor(false)]; + tensor mh_w_93_cast_fp16 = matmul(transpose_x = mh_w_93_transpose_x_0, transpose_y = mh_w_93_transpose_y_0, x = var_6768_cast_fp16, y = var_6772_cast_fp16)[name = tensor("mh_w_93_cast_fp16")]; + tensor mh_w_cast_fp16 = add(x = mh_w_93_cast_fp16, y = qk_mask_cast_fp16)[name = tensor("mh_w_cast_fp16")]; + tensor var_6776_cast_fp16 = softmax(axis = var_6617, x = mh_w_cast_fp16)[name = tensor("op_6776_cast_fp16")]; + tensor var_6777 = const()[name = tensor("op_6777"), val = tensor([1, 8, 128, 188])]; + tensor var_6778_cast_fp16 = reshape(shape = var_6777, x = value_cast_fp16)[name = tensor("op_6778_cast_fp16")]; + tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; + tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; + tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_6778_cast_fp16, y = var_6776_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_6781 = const()[name = tensor("op_6781"), val = tensor([1, 1024, 1, 188])]; + tensor input_621_cast_fp16 = reshape(shape = var_6781, x = attn_cast_fp16)[name = tensor("input_621_cast_fp16")]; + tensor obj_pad_type_0 = const()[name = tensor("obj_pad_type_0"), val = tensor("valid")]; + tensor obj_strides_0 = const()[name = tensor("obj_strides_0"), val = tensor([1, 1])]; + tensor obj_pad_0 = const()[name = tensor("obj_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor obj_dilations_0 = const()[name = tensor("obj_dilations_0"), val = tensor([1, 1])]; + tensor obj_groups_0 = const()[name = tensor("obj_groups_0"), val = tensor(1)]; + tensor layers_23_self_attn_o_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(447979520))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448766016))), name = tensor("layers_23_self_attn_o_proj_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor obj_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = obj_dilations_0, groups = obj_groups_0, pad = obj_pad_0, pad_type = obj_pad_type_0, strides = obj_strides_0, weight = layers_23_self_attn_o_proj_weight_to_fp16_palettized, x = input_621_cast_fp16)[name = tensor("obj_cast_fp16")]; + tensor inputs_235_cast_fp16 = add(x = inputs_233_cast_fp16, y = obj_cast_fp16)[name = tensor("inputs_235_cast_fp16")]; + tensor out_235_axes_0 = const()[name = tensor("out_235_axes_0"), val = tensor([1])]; + tensor var_6799_to_fp16 = const()[name = tensor("op_6799_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_235_cast_fp16 = layer_norm(axes = out_235_axes_0, epsilon = var_6799_to_fp16, x = inputs_235_cast_fp16)[name = tensor("out_235_cast_fp16")]; + tensor input_623_gamma_0_to_fp16 = const()[name = tensor("input_623_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448766208)))]; + tensor input_623_beta_0_to_fp16 = const()[name = tensor("input_623_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448768320)))]; + tensor input_623_epsilon_0_to_fp16 = const()[name = tensor("input_623_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_623_cast_fp16 = batch_norm(beta = input_623_beta_0_to_fp16, epsilon = input_623_epsilon_0_to_fp16, gamma = input_623_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_235_cast_fp16)[name = tensor("input_623_cast_fp16")]; + tensor input_625_pad_type_0 = const()[name = tensor("input_625_pad_type_0"), val = tensor("valid")]; + tensor input_625_strides_0 = const()[name = tensor("input_625_strides_0"), val = tensor([1, 1])]; + tensor input_625_pad_0 = const()[name = tensor("input_625_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_625_dilations_0 = const()[name = tensor("input_625_dilations_0"), val = tensor([1, 1])]; + tensor input_625_groups_0 = const()[name = tensor("input_625_groups_0"), val = tensor(1)]; + tensor layers_23_conv_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448770432))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450343360))), name = tensor("layers_23_conv_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor([2048, 1024, 1, 1])]; + tensor input_625_cast_fp16 = conv(dilations = input_625_dilations_0, groups = input_625_groups_0, pad = input_625_pad_0, pad_type = input_625_pad_type_0, strides = input_625_strides_0, weight = layers_23_conv_pointwise_conv1_weight_to_fp16_palettized, x = input_623_cast_fp16)[name = tensor("input_625_cast_fp16")]; + tensor input_627_split_num_splits_0 = const()[name = tensor("input_627_split_num_splits_0"), val = tensor(2)]; + tensor input_627_split_axis_0 = const()[name = tensor("input_627_split_axis_0"), val = tensor(1)]; + tensor input_627_split_cast_fp16_0, tensor input_627_split_cast_fp16_1 = split(axis = input_627_split_axis_0, num_splits = input_627_split_num_splits_0, x = input_625_cast_fp16)[name = tensor("input_627_split_cast_fp16")]; + tensor input_627_split_1_sigmoid_cast_fp16 = sigmoid(x = input_627_split_cast_fp16_1)[name = tensor("input_627_split_1_sigmoid_cast_fp16")]; + tensor input_627_cast_fp16 = mul(x = input_627_split_cast_fp16_0, y = input_627_split_1_sigmoid_cast_fp16)[name = tensor("input_627_cast_fp16")]; + tensor input_629_pad_type_0 = const()[name = tensor("input_629_pad_type_0"), val = tensor("custom")]; + tensor input_629_pad_0 = const()[name = tensor("input_629_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_629_groups_0 = const()[name = tensor("input_629_groups_0"), val = tensor(1024)]; + tensor input_629_strides_0 = const()[name = tensor("input_629_strides_0"), val = tensor([1, 1])]; + tensor input_629_dilations_0 = const()[name = tensor("input_629_dilations_0"), val = tensor([1, 1])]; + tensor const_314_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450343552))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450350528))), name = tensor("const_314_to_fp16_palettized"), shape = tensor([1024, 1, 1, 9])]; + tensor const_315_to_fp16 = const()[name = tensor("const_315_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450350720)))]; + tensor input_631_cast_fp16 = conv(bias = const_315_to_fp16, dilations = input_629_dilations_0, groups = input_629_groups_0, pad = input_629_pad_0, pad_type = input_629_pad_type_0, strides = input_629_strides_0, weight = const_314_to_fp16_palettized, x = input_627_cast_fp16)[name = tensor("input_631_cast_fp16")]; + tensor input_633_cast_fp16 = silu(x = input_631_cast_fp16)[name = tensor("input_633_cast_fp16")]; + tensor x_143_pad_type_0 = const()[name = tensor("x_143_pad_type_0"), val = tensor("valid")]; + tensor x_143_strides_0 = const()[name = tensor("x_143_strides_0"), val = tensor([1, 1])]; + tensor x_143_pad_0 = const()[name = tensor("x_143_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_143_dilations_0 = const()[name = tensor("x_143_dilations_0"), val = tensor([1, 1])]; + tensor x_143_groups_0 = const()[name = tensor("x_143_groups_0"), val = tensor(1)]; + tensor layers_23_conv_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(450352832))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451139328))), name = tensor("layers_23_conv_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor([1024, 1024, 1, 1])]; + tensor x_143_cast_fp16 = conv(dilations = x_143_dilations_0, groups = x_143_groups_0, pad = x_143_pad_0, pad_type = x_143_pad_type_0, strides = x_143_strides_0, weight = layers_23_conv_pointwise_conv2_weight_to_fp16_palettized, x = input_633_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor inputs_237_cast_fp16 = add(x = inputs_235_cast_fp16, y = x_143_cast_fp16)[name = tensor("inputs_237_cast_fp16")]; + tensor out_237_axes_0 = const()[name = tensor("out_237_axes_0"), val = tensor([1])]; + tensor var_6847_to_fp16 = const()[name = tensor("op_6847_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_237_cast_fp16 = layer_norm(axes = out_237_axes_0, epsilon = var_6847_to_fp16, x = inputs_237_cast_fp16)[name = tensor("out_237_cast_fp16")]; + tensor input_635_gamma_0_to_fp16 = const()[name = tensor("input_635_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451139520)))]; + tensor input_635_beta_0_to_fp16 = const()[name = tensor("input_635_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451141632)))]; + tensor input_635_epsilon_0_to_fp16 = const()[name = tensor("input_635_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_635_cast_fp16 = batch_norm(beta = input_635_beta_0_to_fp16, epsilon = input_635_epsilon_0_to_fp16, gamma = input_635_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_237_cast_fp16)[name = tensor("input_635_cast_fp16")]; + tensor input_637_pad_type_0 = const()[name = tensor("input_637_pad_type_0"), val = tensor("valid")]; + tensor input_637_strides_0 = const()[name = tensor("input_637_strides_0"), val = tensor([1, 1])]; + tensor input_637_pad_0 = const()[name = tensor("input_637_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_637_dilations_0 = const()[name = tensor("input_637_dilations_0"), val = tensor([1, 1])]; + tensor input_637_groups_0 = const()[name = tensor("input_637_groups_0"), val = tensor(1)]; + tensor layers_23_feed_forward2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(451143744))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454289536))), name = tensor("layers_23_feed_forward2_fc1_weight_to_fp16_palettized"), shape = tensor([4096, 1024, 1, 1])]; + tensor input_637_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_637_dilations_0, groups = input_637_groups_0, pad = input_637_pad_0, pad_type = input_637_pad_type_0, strides = input_637_strides_0, weight = layers_23_feed_forward2_fc1_weight_to_fp16_palettized, x = input_635_cast_fp16)[name = tensor("input_637_cast_fp16")]; + tensor input_639_cast_fp16 = silu(x = input_637_cast_fp16)[name = tensor("input_639_cast_fp16")]; + tensor x_pad_type_0 = const()[name = tensor("x_pad_type_0"), val = tensor("valid")]; + tensor x_strides_0 = const()[name = tensor("x_strides_0"), val = tensor([1, 1])]; + tensor x_pad_0 = const()[name = tensor("x_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor x_dilations_0 = const()[name = tensor("x_dilations_0"), val = tensor([1, 1])]; + tensor x_groups_0 = const()[name = tensor("x_groups_0"), val = tensor(1)]; + tensor op_6875_weight_0_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454289728))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457435520))), name = tensor("op_6875_weight_0_to_fp16_palettized"), shape = tensor([1024, 4096, 1, 1])]; + tensor var_6875_cast_fp16 = conv(bias = input_17_mean_0_to_fp16, dilations = x_dilations_0, groups = x_groups_0, pad = x_pad_0, pad_type = x_pad_type_0, strides = x_strides_0, weight = op_6875_weight_0_to_fp16_palettized, x = input_639_cast_fp16)[name = tensor("op_6875_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_237_cast_fp16, y = var_6875_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_239_axes_0 = const()[name = tensor("out_239_axes_0"), val = tensor([1])]; + tensor var_6885_to_fp16 = const()[name = tensor("op_6885_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_239_cast_fp16 = layer_norm(axes = out_239_axes_0, epsilon = var_6885_to_fp16, x = inputs_cast_fp16)[name = tensor("out_239_cast_fp16")]; + tensor encoder_output_embeds_type_fp32_gamma_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457435712)))]; + tensor encoder_output_embeds_type_fp32_beta_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457437824)))]; + tensor encoder_output_embeds_type_fp32_epsilon_0_to_fp16 = const()[name = tensor("encoder_output_embeds_type_fp32_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor encoder_output_embeds = batch_norm(beta = encoder_output_embeds_type_fp32_beta_0_to_fp16, epsilon = encoder_output_embeds_type_fp32_epsilon_0_to_fp16, gamma = encoder_output_embeds_type_fp32_gamma_0_to_fp16, mean = input_17_mean_0_to_fp16, variance = input_17_variance_0_to_fp16, x = out_239_cast_fp16)[name = tensor("encoder_output_embeds_type_fp32_cast_fp16")]; + tensor var_6905_pad_type_0 = const()[name = tensor("op_6905_pad_type_0"), val = tensor("valid")]; + tensor var_6905_strides_0 = const()[name = tensor("op_6905_strides_0"), val = tensor([1, 1])]; + tensor var_6905_pad_0 = const()[name = tensor("op_6905_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6905_dilations_0 = const()[name = tensor("op_6905_dilations_0"), val = tensor([1, 1])]; + tensor var_6905_groups_0 = const()[name = tensor("op_6905_groups_0"), val = tensor(1)]; + tensor joint_projection_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457439936))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457931520))), name = tensor("joint_projection_weight_to_fp16_palettized"), shape = tensor([640, 1024, 1, 1])]; + tensor joint_projection_bias_to_fp16 = const()[name = tensor("joint_projection_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(457931712)))]; + tensor joint_projected_encoder_output_embeds = conv(bias = joint_projection_bias_to_fp16, dilations = var_6905_dilations_0, groups = var_6905_groups_0, pad = var_6905_pad_0, pad_type = var_6905_pad_type_0, strides = var_6905_strides_0, weight = joint_projection_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("op_6905_cast_fp16")]; + } -> (encoder_output_embeds, joint_projected_encoder_output_embeds); +} \ No newline at end of file