diff --git "a/nvidia_parakeet-v2/AudioEncoder.mlmodelc/model.mil" "b/nvidia_parakeet-v2/AudioEncoder.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/nvidia_parakeet-v2/AudioEncoder.mlmodelc/model.mil" @@ -0,0 +1,4393 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.2"}})] +{ + func main(tensor melspectrogram_features) { + 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 = const()[name = tensor("pre_encode_conv_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + 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(4736)))]; + 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, 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 = const()[name = tensor("pre_encode_conv_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5312)))]; + 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(9984)))]; + 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, 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 = const()[name = tensor("pre_encode_conv_3_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10560)))]; + 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(141696)))]; + 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, 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 = const()[name = tensor("pre_encode_conv_5_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142272)))]; + 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(146944)))]; + 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, 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 = const()[name = tensor("pre_encode_conv_6_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147520)))]; + 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(278656)))]; + 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, 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 var_119_cast_fp16 = reshape(shape = var_118, x = var_115_cast_fp16)[name = tensor("op_119_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 var_123_to_fp16 = const()[name = tensor("op_123_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(279232)))]; + 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(8667904)))]; + 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 = var_123_to_fp16, x = var_119_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; + tensor var_133 = const()[name = tensor("op_133"), val = tensor(3)]; + tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; + tensor var_164_to_fp16 = const()[name = tensor("op_164_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_164_to_fp16, x = inputs_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; + tensor input_15_mean_0_to_fp16 = const()[name = tensor("input_15_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8670016)))]; + tensor input_15_variance_0_to_fp16 = const()[name = tensor("input_15_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8672128)))]; + tensor input_15_gamma_0_to_fp16 = const()[name = tensor("input_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8674240)))]; + tensor input_15_beta_0_to_fp16 = const()[name = tensor("input_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8676352)))]; + tensor input_15_epsilon_0_to_fp16 = const()[name = tensor("input_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = batch_norm(beta = input_15_beta_0_to_fp16, epsilon = input_15_epsilon_0_to_fp16, gamma = input_15_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_pad_type_0 = const()[name = tensor("input_17_pad_type_0"), val = tensor("valid")]; + tensor input_17_strides_0 = const()[name = tensor("input_17_strides_0"), val = tensor([1, 1])]; + tensor input_17_pad_0 = const()[name = tensor("input_17_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_17_dilations_0 = const()[name = tensor("input_17_dilations_0"), val = tensor([1, 1])]; + tensor input_17_groups_0 = const()[name = tensor("input_17_groups_0"), val = tensor(1)]; + tensor layers_0_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_0_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8678464)))]; + 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(17067136)))]; + tensor input_17_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = layers_0_feed_forward1_fc1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_cast_fp16 = silu(x = input_17_cast_fp16)[name = tensor("input_19_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 var_192_weight_0_to_fp16 = const()[name = tensor("op_192_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17075392)))]; + tensor var_192_cast_fp16 = conv(bias = input_15_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 = var_192_weight_0_to_fp16, x = input_19_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = var_192_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_202_to_fp16 = const()[name = tensor("op_202_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_202_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(25464064)))]; + 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(25466176)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25468288)))]; + tensor query_1_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27565504)))]; + 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, 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 = const()[name = tensor("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29662720)))]; + tensor value_1_cast_fp16 = conv(bias = input_15_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, x = obj_1_cast_fp16)[name = tensor("value_1_cast_fp16")]; + tensor var_240_to_fp16 = const()[name = tensor("op_240_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31759936)))]; + tensor query_3_cast_fp16 = add(x = query_1_cast_fp16, y = var_240_to_fp16)[name = tensor("query_3_cast_fp16")]; + tensor var_243_to_fp16 = const()[name = tensor("op_243_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762048)))]; + tensor q_with_bias_v_1_cast_fp16 = add(x = query_1_cast_fp16, y = var_243_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 pos_enc_to_fp16 = const()[name = tensor("pos_enc_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31764160)))]; + tensor layers_0_self_attn_linear_pos_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32532224)))]; + tensor p_1_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_1_cast_fp16")]; + tensor var_254 = const()[name = tensor("op_254"), val = tensor([1, 8, 128, 188])]; + tensor var_255_cast_fp16 = reshape(shape = var_254, x = q_with_bias_v_1_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor var_256 = const()[name = tensor("op_256"), val = tensor([1, 8, 128, -1])]; + tensor var_257_cast_fp16 = reshape(shape = var_256, x = p_1_cast_fp16)[name = tensor("op_257_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_255_cast_fp16, y = var_257_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_266 = const()[name = tensor("op_266"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_5_cast_fp16 = reshape(shape = var_266, x = matrix_bd_3_cast_fp16)[name = tensor("matrix_bd_5_cast_fp16")]; + tensor var_270_begin_0 = const()[name = tensor("op_270_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_270_end_0 = const()[name = tensor("op_270_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_270_end_mask_0 = const()[name = tensor("op_270_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_270_cast_fp16 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = matrix_bd_5_cast_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271 = const()[name = tensor("op_271"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_7_cast_fp16 = reshape(shape = var_271, x = var_270_cast_fp16)[name = tensor("matrix_bd_7_cast_fp16")]; + tensor var_276_begin_0 = const()[name = tensor("op_276_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_276_end_0 = const()[name = tensor("op_276_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_276_end_mask_0 = const()[name = tensor("op_276_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_276_cast_fp16 = slice_by_index(begin = var_276_begin_0, end = var_276_end_0, end_mask = var_276_end_mask_0, x = matrix_bd_7_cast_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_to_fp16 = const()[name = tensor("op_277_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_1_cast_fp16 = mul(x = var_276_cast_fp16, y = var_277_to_fp16)[name = tensor("qk_mask_1_cast_fp16")]; + tensor var_281 = const()[name = tensor("op_281"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_1_cast_fp16 = reshape(shape = var_281, x = query_3_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; + tensor var_283_to_fp16 = const()[name = tensor("op_283_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_284_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_283_to_fp16)[name = tensor("op_284_cast_fp16")]; + tensor var_287 = const()[name = tensor("op_287"), val = tensor([1, 8, 128, 188])]; + tensor var_288_cast_fp16 = reshape(shape = var_287, x = key_1_cast_fp16)[name = tensor("op_288_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_284_cast_fp16, y = var_288_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_292_cast_fp16 = softmax(axis = var_133, x = mh_w_3_cast_fp16)[name = tensor("op_292_cast_fp16")]; + tensor var_293 = const()[name = tensor("op_293"), val = tensor([1, 8, 128, 188])]; + tensor var_294_cast_fp16 = reshape(shape = var_293, x = value_1_cast_fp16)[name = tensor("op_294_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_294_cast_fp16, y = var_292_cast_fp16)[name = tensor("attn_1_cast_fp16")]; + tensor var_297 = const()[name = tensor("op_297"), val = tensor([1, 1024, 1, 188])]; + tensor input_21_cast_fp16 = reshape(shape = var_297, x = attn_1_cast_fp16)[name = tensor("input_21_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 = const()[name = tensor("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34629440)))]; + tensor obj_5_cast_fp16 = conv(bias = input_15_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, x = input_21_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_315_to_fp16 = const()[name = tensor("op_315_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_315_to_fp16, x = inputs_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; + tensor input_23_gamma_0_to_fp16 = const()[name = tensor("input_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36726656)))]; + tensor input_23_beta_0_to_fp16 = const()[name = tensor("input_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36728768)))]; + tensor input_23_epsilon_0_to_fp16 = const()[name = tensor("input_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = batch_norm(beta = input_23_beta_0_to_fp16, epsilon = input_23_epsilon_0_to_fp16, gamma = input_23_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1, 1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1, 1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor layers_0_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_0_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36730880)))]; + tensor input_25_cast_fp16 = conv(dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = layers_0_conv_pointwise_conv1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_split_num_splits_0 = const()[name = tensor("input_27_split_num_splits_0"), val = tensor(2)]; + tensor input_27_split_axis_0 = const()[name = tensor("input_27_split_axis_0"), val = tensor(1)]; + tensor input_27_split_cast_fp16_0, tensor input_27_split_cast_fp16_1 = split(axis = input_27_split_axis_0, num_splits = input_27_split_num_splits_0, x = input_25_cast_fp16)[name = tensor("input_27_split_cast_fp16")]; + tensor input_27_split_1_sigmoid_cast_fp16 = sigmoid(x = input_27_split_cast_fp16_1)[name = tensor("input_27_split_1_sigmoid_cast_fp16")]; + tensor input_27_cast_fp16 = mul(x = input_27_split_cast_fp16_0, y = input_27_split_1_sigmoid_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("custom")]; + tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1024)]; + tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; + tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; + tensor const_268_to_fp16 = const()[name = tensor("const_268_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40925248)))]; + tensor const_269_to_fp16 = const()[name = tensor("const_269_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40943744)))]; + tensor input_31_cast_fp16 = conv(bias = const_269_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = const_268_to_fp16, x = input_27_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_cast_fp16 = silu(x = input_31_cast_fp16)[name = tensor("input_33_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 = const()[name = tensor("layers_0_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40945856)))]; + 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, x = input_33_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_363_to_fp16 = const()[name = tensor("op_363_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_363_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; + tensor input_35_gamma_0_to_fp16 = const()[name = tensor("input_35_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43043072)))]; + tensor input_35_beta_0_to_fp16 = const()[name = tensor("input_35_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43045184)))]; + tensor input_35_epsilon_0_to_fp16 = const()[name = tensor("input_35_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = batch_norm(beta = input_35_beta_0_to_fp16, epsilon = input_35_epsilon_0_to_fp16, gamma = input_35_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor input_37_pad_type_0 = const()[name = tensor("input_37_pad_type_0"), val = tensor("valid")]; + tensor input_37_strides_0 = const()[name = tensor("input_37_strides_0"), val = tensor([1, 1])]; + tensor input_37_pad_0 = const()[name = tensor("input_37_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_37_dilations_0 = const()[name = tensor("input_37_dilations_0"), val = tensor([1, 1])]; + tensor input_37_groups_0 = const()[name = tensor("input_37_groups_0"), val = tensor(1)]; + tensor layers_0_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_0_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43047296)))]; + tensor input_37_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_37_dilations_0, groups = input_37_groups_0, pad = input_37_pad_0, pad_type = input_37_pad_type_0, strides = input_37_strides_0, weight = layers_0_feed_forward2_fc1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_cast_fp16 = silu(x = input_37_cast_fp16)[name = tensor("input_39_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 var_391_weight_0_to_fp16 = const()[name = tensor("op_391_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51435968)))]; + tensor var_391_cast_fp16 = conv(bias = input_15_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 = var_391_weight_0_to_fp16, x = input_39_cast_fp16)[name = tensor("op_391_cast_fp16")]; + tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = var_391_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_401_to_fp16 = const()[name = tensor("op_401_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_401_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(59824640)))]; + 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(59826752)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor(3)]; + tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; + tensor var_446_to_fp16 = const()[name = tensor("op_446_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_446_to_fp16, x = inputs_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; + tensor input_41_gamma_0_to_fp16 = const()[name = tensor("input_41_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59828864)))]; + tensor input_41_beta_0_to_fp16 = const()[name = tensor("input_41_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59830976)))]; + tensor input_41_epsilon_0_to_fp16 = const()[name = tensor("input_41_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = batch_norm(beta = input_41_beta_0_to_fp16, epsilon = input_41_epsilon_0_to_fp16, gamma = input_41_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1, 1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1, 1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor layers_1_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_1_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59833088)))]; + tensor input_43_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = layers_1_feed_forward1_fc1_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = input_43_cast_fp16)[name = tensor("input_45_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 var_474_weight_0_to_fp16 = const()[name = tensor("op_474_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68221760)))]; + tensor var_474_cast_fp16 = conv(bias = input_15_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 = var_474_weight_0_to_fp16, x = input_45_cast_fp16)[name = tensor("op_474_cast_fp16")]; + tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = var_474_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_484_to_fp16 = const()[name = tensor("op_484_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_484_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(76610432)))]; + 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(76612544)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76614656)))]; + tensor query_5_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78711872)))]; + 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, 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 = const()[name = tensor("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(80809088)))]; + tensor value_3_cast_fp16 = conv(bias = input_15_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, x = obj_7_cast_fp16)[name = tensor("value_3_cast_fp16")]; + tensor var_522_to_fp16 = const()[name = tensor("op_522_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82906304)))]; + tensor query_7_cast_fp16 = add(x = query_5_cast_fp16, y = var_522_to_fp16)[name = tensor("query_7_cast_fp16")]; + tensor var_525_to_fp16 = const()[name = tensor("op_525_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82908416)))]; + tensor q_with_bias_v_3_cast_fp16 = add(x = query_5_cast_fp16, y = var_525_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 = const()[name = tensor("layers_1_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82910528)))]; + tensor p_3_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_3_cast_fp16")]; + tensor var_536 = const()[name = tensor("op_536"), val = tensor([1, 8, 128, 188])]; + tensor var_537_cast_fp16 = reshape(shape = var_536, x = q_with_bias_v_3_cast_fp16)[name = tensor("op_537_cast_fp16")]; + tensor var_538 = const()[name = tensor("op_538"), val = tensor([1, 8, 128, -1])]; + tensor var_539_cast_fp16 = reshape(shape = var_538, x = p_3_cast_fp16)[name = tensor("op_539_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_537_cast_fp16, y = var_539_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_548 = const()[name = tensor("op_548"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_13_cast_fp16 = reshape(shape = var_548, x = matrix_bd_11_cast_fp16)[name = tensor("matrix_bd_13_cast_fp16")]; + tensor var_552_begin_0 = const()[name = tensor("op_552_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_552_end_0 = const()[name = tensor("op_552_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_552_end_mask_0 = const()[name = tensor("op_552_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, x = matrix_bd_13_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_15_cast_fp16 = reshape(shape = var_553, x = var_552_cast_fp16)[name = tensor("matrix_bd_15_cast_fp16")]; + tensor var_558_begin_0 = const()[name = tensor("op_558_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_558_end_0 = const()[name = tensor("op_558_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_558_end_mask_0 = const()[name = tensor("op_558_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_558_cast_fp16 = slice_by_index(begin = var_558_begin_0, end = var_558_end_0, end_mask = var_558_end_mask_0, x = matrix_bd_15_cast_fp16)[name = tensor("op_558_cast_fp16")]; + tensor var_559_to_fp16 = const()[name = tensor("op_559_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_3_cast_fp16 = mul(x = var_558_cast_fp16, y = var_559_to_fp16)[name = tensor("qk_mask_3_cast_fp16")]; + tensor var_563 = const()[name = tensor("op_563"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_3_cast_fp16 = reshape(shape = var_563, x = query_7_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; + tensor var_565_to_fp16 = const()[name = tensor("op_565_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_566_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_565_to_fp16)[name = tensor("op_566_cast_fp16")]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 8, 128, 188])]; + tensor var_570_cast_fp16 = reshape(shape = var_569, x = key_3_cast_fp16)[name = tensor("op_570_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_566_cast_fp16, y = var_570_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_574_cast_fp16 = softmax(axis = var_415, x = mh_w_7_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 8, 128, 188])]; + tensor var_576_cast_fp16 = reshape(shape = var_575, x = value_3_cast_fp16)[name = tensor("op_576_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_576_cast_fp16, y = var_574_cast_fp16)[name = tensor("attn_3_cast_fp16")]; + tensor var_579 = const()[name = tensor("op_579"), val = tensor([1, 1024, 1, 188])]; + tensor input_47_cast_fp16 = reshape(shape = var_579, x = attn_3_cast_fp16)[name = tensor("input_47_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 = const()[name = tensor("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85007744)))]; + tensor obj_9_cast_fp16 = conv(bias = input_15_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, x = input_47_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_597_to_fp16 = const()[name = tensor("op_597_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_597_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; + tensor input_49_gamma_0_to_fp16 = const()[name = tensor("input_49_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87104960)))]; + tensor input_49_beta_0_to_fp16 = const()[name = tensor("input_49_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87107072)))]; + tensor input_49_epsilon_0_to_fp16 = const()[name = tensor("input_49_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = batch_norm(beta = input_49_beta_0_to_fp16, epsilon = input_49_epsilon_0_to_fp16, gamma = input_49_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_pad_type_0 = const()[name = tensor("input_51_pad_type_0"), val = tensor("valid")]; + tensor input_51_strides_0 = const()[name = tensor("input_51_strides_0"), val = tensor([1, 1])]; + tensor input_51_pad_0 = const()[name = tensor("input_51_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_51_dilations_0 = const()[name = tensor("input_51_dilations_0"), val = tensor([1, 1])]; + tensor input_51_groups_0 = const()[name = tensor("input_51_groups_0"), val = tensor(1)]; + tensor layers_1_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_1_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(87109184)))]; + tensor input_51_cast_fp16 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = layers_1_conv_pointwise_conv1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_split_num_splits_0 = const()[name = tensor("input_53_split_num_splits_0"), val = tensor(2)]; + tensor input_53_split_axis_0 = const()[name = tensor("input_53_split_axis_0"), val = tensor(1)]; + tensor input_53_split_cast_fp16_0, tensor input_53_split_cast_fp16_1 = split(axis = input_53_split_axis_0, num_splits = input_53_split_num_splits_0, x = input_51_cast_fp16)[name = tensor("input_53_split_cast_fp16")]; + tensor input_53_split_1_sigmoid_cast_fp16 = sigmoid(x = input_53_split_cast_fp16_1)[name = tensor("input_53_split_1_sigmoid_cast_fp16")]; + tensor input_53_cast_fp16 = mul(x = input_53_split_cast_fp16_0, y = input_53_split_1_sigmoid_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor input_55_pad_type_0 = const()[name = tensor("input_55_pad_type_0"), val = tensor("custom")]; + tensor input_55_pad_0 = const()[name = tensor("input_55_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_55_groups_0 = const()[name = tensor("input_55_groups_0"), val = tensor(1024)]; + tensor input_55_strides_0 = const()[name = tensor("input_55_strides_0"), val = tensor([1, 1])]; + tensor input_55_dilations_0 = const()[name = tensor("input_55_dilations_0"), val = tensor([1, 1])]; + tensor const_270_to_fp16 = const()[name = tensor("const_270_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91303552)))]; + tensor const_271_to_fp16 = const()[name = tensor("const_271_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91322048)))]; + tensor input_57_cast_fp16 = conv(bias = const_271_to_fp16, dilations = input_55_dilations_0, groups = input_55_groups_0, pad = input_55_pad_0, pad_type = input_55_pad_type_0, strides = input_55_strides_0, weight = const_270_to_fp16, x = input_53_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_cast_fp16 = silu(x = input_57_cast_fp16)[name = tensor("input_59_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 = const()[name = tensor("layers_1_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91324160)))]; + 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, x = input_59_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_645_to_fp16 = const()[name = tensor("op_645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_645_to_fp16, x = inputs_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; + tensor input_61_gamma_0_to_fp16 = const()[name = tensor("input_61_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93421376)))]; + tensor input_61_beta_0_to_fp16 = const()[name = tensor("input_61_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93423488)))]; + tensor input_61_epsilon_0_to_fp16 = const()[name = tensor("input_61_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_61_cast_fp16 = batch_norm(beta = input_61_beta_0_to_fp16, epsilon = input_61_epsilon_0_to_fp16, gamma = input_61_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_pad_type_0 = const()[name = tensor("input_63_pad_type_0"), val = tensor("valid")]; + tensor input_63_strides_0 = const()[name = tensor("input_63_strides_0"), val = tensor([1, 1])]; + tensor input_63_pad_0 = const()[name = tensor("input_63_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_63_dilations_0 = const()[name = tensor("input_63_dilations_0"), val = tensor([1, 1])]; + tensor input_63_groups_0 = const()[name = tensor("input_63_groups_0"), val = tensor(1)]; + tensor layers_1_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_1_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93425600)))]; + tensor input_63_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_63_dilations_0, groups = input_63_groups_0, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = input_63_strides_0, weight = layers_1_feed_forward2_fc1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_cast_fp16 = silu(x = input_63_cast_fp16)[name = tensor("input_65_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 var_673_weight_0_to_fp16 = const()[name = tensor("op_673_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101814272)))]; + tensor var_673_cast_fp16 = conv(bias = input_15_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 = var_673_weight_0_to_fp16, x = input_65_cast_fp16)[name = tensor("op_673_cast_fp16")]; + tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = var_673_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_683_to_fp16 = const()[name = tensor("op_683_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_683_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(110202944)))]; + 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(110205056)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor(3)]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; + tensor var_728_to_fp16 = const()[name = tensor("op_728_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_728_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; + tensor input_67_gamma_0_to_fp16 = const()[name = tensor("input_67_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110207168)))]; + tensor input_67_beta_0_to_fp16 = const()[name = tensor("input_67_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110209280)))]; + tensor input_67_epsilon_0_to_fp16 = const()[name = tensor("input_67_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_67_cast_fp16 = batch_norm(beta = input_67_beta_0_to_fp16, epsilon = input_67_epsilon_0_to_fp16, gamma = input_67_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; + tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; + tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; + tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; + tensor layers_2_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_2_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110211392)))]; + tensor input_69_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = layers_2_feed_forward1_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_cast_fp16 = silu(x = input_69_cast_fp16)[name = tensor("input_71_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 var_756_weight_0_to_fp16 = const()[name = tensor("op_756_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118600064)))]; + tensor var_756_cast_fp16 = conv(bias = input_15_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 = var_756_weight_0_to_fp16, x = input_71_cast_fp16)[name = tensor("op_756_cast_fp16")]; + tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = var_756_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_766_to_fp16 = const()[name = tensor("op_766_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_766_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(126988736)))]; + 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(126990848)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126992960)))]; + tensor query_9_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129090176)))]; + 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, 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 = const()[name = tensor("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131187392)))]; + tensor value_5_cast_fp16 = conv(bias = input_15_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, x = obj_11_cast_fp16)[name = tensor("value_5_cast_fp16")]; + tensor var_804_to_fp16 = const()[name = tensor("op_804_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133284608)))]; + tensor query_11_cast_fp16 = add(x = query_9_cast_fp16, y = var_804_to_fp16)[name = tensor("query_11_cast_fp16")]; + tensor var_807_to_fp16 = const()[name = tensor("op_807_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133286720)))]; + tensor q_with_bias_v_5_cast_fp16 = add(x = query_9_cast_fp16, y = var_807_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 = const()[name = tensor("layers_2_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133288832)))]; + tensor p_5_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_5_cast_fp16")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 8, 128, 188])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = q_with_bias_v_5_cast_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_820 = const()[name = tensor("op_820"), val = tensor([1, 8, 128, -1])]; + tensor var_821_cast_fp16 = reshape(shape = var_820, x = p_5_cast_fp16)[name = tensor("op_821_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_819_cast_fp16, y = var_821_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_830 = const()[name = tensor("op_830"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_21_cast_fp16 = reshape(shape = var_830, x = matrix_bd_19_cast_fp16)[name = tensor("matrix_bd_21_cast_fp16")]; + tensor var_834_begin_0 = const()[name = tensor("op_834_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_834_end_0 = const()[name = tensor("op_834_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_834_end_mask_0 = const()[name = tensor("op_834_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_834_cast_fp16 = slice_by_index(begin = var_834_begin_0, end = var_834_end_0, end_mask = var_834_end_mask_0, x = matrix_bd_21_cast_fp16)[name = tensor("op_834_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_23_cast_fp16 = reshape(shape = var_835, x = var_834_cast_fp16)[name = tensor("matrix_bd_23_cast_fp16")]; + tensor var_840_begin_0 = const()[name = tensor("op_840_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_840_end_0 = const()[name = tensor("op_840_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_840_end_mask_0 = const()[name = tensor("op_840_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_840_cast_fp16 = slice_by_index(begin = var_840_begin_0, end = var_840_end_0, end_mask = var_840_end_mask_0, x = matrix_bd_23_cast_fp16)[name = tensor("op_840_cast_fp16")]; + tensor var_841_to_fp16 = const()[name = tensor("op_841_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_5_cast_fp16 = mul(x = var_840_cast_fp16, y = var_841_to_fp16)[name = tensor("qk_mask_5_cast_fp16")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_5_cast_fp16 = reshape(shape = var_845, x = query_11_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; + tensor var_847_to_fp16 = const()[name = tensor("op_847_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_848_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_847_to_fp16)[name = tensor("op_848_cast_fp16")]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 8, 128, 188])]; + tensor var_852_cast_fp16 = reshape(shape = var_851, x = key_5_cast_fp16)[name = tensor("op_852_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_848_cast_fp16, y = var_852_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_856_cast_fp16 = softmax(axis = var_697, x = mh_w_11_cast_fp16)[name = tensor("op_856_cast_fp16")]; + tensor var_857 = const()[name = tensor("op_857"), val = tensor([1, 8, 128, 188])]; + tensor var_858_cast_fp16 = reshape(shape = var_857, x = value_5_cast_fp16)[name = tensor("op_858_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_858_cast_fp16, y = var_856_cast_fp16)[name = tensor("attn_5_cast_fp16")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 1024, 1, 188])]; + tensor input_73_cast_fp16 = reshape(shape = var_861, x = attn_5_cast_fp16)[name = tensor("input_73_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 = const()[name = tensor("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135386048)))]; + tensor obj_13_cast_fp16 = conv(bias = input_15_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, x = input_73_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_879_to_fp16 = const()[name = tensor("op_879_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_25_cast_fp16 = layer_norm(axes = out_25_axes_0, epsilon = var_879_to_fp16, x = inputs_25_cast_fp16)[name = tensor("out_25_cast_fp16")]; + tensor input_75_gamma_0_to_fp16 = const()[name = tensor("input_75_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137483264)))]; + tensor input_75_beta_0_to_fp16 = const()[name = tensor("input_75_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137485376)))]; + tensor input_75_epsilon_0_to_fp16 = const()[name = tensor("input_75_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = batch_norm(beta = input_75_beta_0_to_fp16, epsilon = input_75_epsilon_0_to_fp16, gamma = input_75_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_25_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_pad_type_0 = const()[name = tensor("input_77_pad_type_0"), val = tensor("valid")]; + tensor input_77_strides_0 = const()[name = tensor("input_77_strides_0"), val = tensor([1, 1])]; + tensor input_77_pad_0 = const()[name = tensor("input_77_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_77_dilations_0 = const()[name = tensor("input_77_dilations_0"), val = tensor([1, 1])]; + tensor input_77_groups_0 = const()[name = tensor("input_77_groups_0"), val = tensor(1)]; + tensor layers_2_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_2_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137487488)))]; + tensor input_77_cast_fp16 = conv(dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = layers_2_conv_pointwise_conv1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_split_num_splits_0 = const()[name = tensor("input_79_split_num_splits_0"), val = tensor(2)]; + tensor input_79_split_axis_0 = const()[name = tensor("input_79_split_axis_0"), val = tensor(1)]; + tensor input_79_split_cast_fp16_0, tensor input_79_split_cast_fp16_1 = split(axis = input_79_split_axis_0, num_splits = input_79_split_num_splits_0, x = input_77_cast_fp16)[name = tensor("input_79_split_cast_fp16")]; + tensor input_79_split_1_sigmoid_cast_fp16 = sigmoid(x = input_79_split_cast_fp16_1)[name = tensor("input_79_split_1_sigmoid_cast_fp16")]; + tensor input_79_cast_fp16 = mul(x = input_79_split_cast_fp16_0, y = input_79_split_1_sigmoid_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_pad_type_0 = const()[name = tensor("input_81_pad_type_0"), val = tensor("custom")]; + tensor input_81_pad_0 = const()[name = tensor("input_81_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_81_groups_0 = const()[name = tensor("input_81_groups_0"), val = tensor(1024)]; + tensor input_81_strides_0 = const()[name = tensor("input_81_strides_0"), val = tensor([1, 1])]; + tensor input_81_dilations_0 = const()[name = tensor("input_81_dilations_0"), val = tensor([1, 1])]; + tensor const_272_to_fp16 = const()[name = tensor("const_272_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141681856)))]; + tensor const_273_to_fp16 = const()[name = tensor("const_273_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141700352)))]; + tensor input_83_cast_fp16 = conv(bias = const_273_to_fp16, dilations = input_81_dilations_0, groups = input_81_groups_0, pad = input_81_pad_0, pad_type = input_81_pad_type_0, strides = input_81_strides_0, weight = const_272_to_fp16, x = input_79_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor input_85_cast_fp16 = silu(x = input_83_cast_fp16)[name = tensor("input_85_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 = const()[name = tensor("layers_2_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141702464)))]; + 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, x = input_85_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_927_to_fp16 = const()[name = tensor("op_927_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_27_cast_fp16 = layer_norm(axes = out_27_axes_0, epsilon = var_927_to_fp16, x = inputs_27_cast_fp16)[name = tensor("out_27_cast_fp16")]; + tensor input_87_gamma_0_to_fp16 = const()[name = tensor("input_87_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143799680)))]; + tensor input_87_beta_0_to_fp16 = const()[name = tensor("input_87_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143801792)))]; + tensor input_87_epsilon_0_to_fp16 = const()[name = tensor("input_87_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = batch_norm(beta = input_87_beta_0_to_fp16, epsilon = input_87_epsilon_0_to_fp16, gamma = input_87_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_27_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; + tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; + tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; + tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; + tensor layers_2_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_2_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143803904)))]; + tensor input_89_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = layers_2_feed_forward2_fc1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor input_91_cast_fp16 = silu(x = input_89_cast_fp16)[name = tensor("input_91_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 var_955_weight_0_to_fp16 = const()[name = tensor("op_955_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152192576)))]; + tensor var_955_cast_fp16 = conv(bias = input_15_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 = var_955_weight_0_to_fp16, x = input_91_cast_fp16)[name = tensor("op_955_cast_fp16")]; + tensor inputs_29_cast_fp16 = add(x = inputs_27_cast_fp16, y = var_955_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_965_to_fp16 = const()[name = tensor("op_965_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_29_cast_fp16 = layer_norm(axes = out_29_axes_0, epsilon = var_965_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(160581248)))]; + 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(160583360)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_29_cast_fp16)[name = tensor("inputs_31_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor(3)]; + tensor out_31_axes_0 = const()[name = tensor("out_31_axes_0"), val = tensor([1])]; + tensor var_1010_to_fp16 = const()[name = tensor("op_1010_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_31_cast_fp16 = layer_norm(axes = out_31_axes_0, epsilon = var_1010_to_fp16, x = inputs_31_cast_fp16)[name = tensor("out_31_cast_fp16")]; + tensor input_93_gamma_0_to_fp16 = const()[name = tensor("input_93_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160585472)))]; + tensor input_93_beta_0_to_fp16 = const()[name = tensor("input_93_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160587584)))]; + tensor input_93_epsilon_0_to_fp16 = const()[name = tensor("input_93_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = batch_norm(beta = input_93_beta_0_to_fp16, epsilon = input_93_epsilon_0_to_fp16, gamma = input_93_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_31_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_pad_type_0 = const()[name = tensor("input_95_pad_type_0"), val = tensor("valid")]; + tensor input_95_strides_0 = const()[name = tensor("input_95_strides_0"), val = tensor([1, 1])]; + tensor input_95_pad_0 = const()[name = tensor("input_95_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_95_dilations_0 = const()[name = tensor("input_95_dilations_0"), val = tensor([1, 1])]; + tensor input_95_groups_0 = const()[name = tensor("input_95_groups_0"), val = tensor(1)]; + tensor layers_3_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_3_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160589696)))]; + tensor input_95_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = layers_3_feed_forward1_fc1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = input_95_cast_fp16)[name = tensor("input_97_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 var_1038_weight_0_to_fp16 = const()[name = tensor("op_1038_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168978368)))]; + tensor var_1038_cast_fp16 = conv(bias = input_15_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 = var_1038_weight_0_to_fp16, x = input_97_cast_fp16)[name = tensor("op_1038_cast_fp16")]; + tensor inputs_33_cast_fp16 = add(x = inputs_31_cast_fp16, y = var_1038_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_1048_to_fp16 = const()[name = tensor("op_1048_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_33_cast_fp16 = layer_norm(axes = out_33_axes_0, epsilon = var_1048_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(177367040)))]; + 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(177369152)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177371264)))]; + tensor query_13_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179468480)))]; + 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, 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 = const()[name = tensor("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181565696)))]; + tensor value_7_cast_fp16 = conv(bias = input_15_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, x = obj_15_cast_fp16)[name = tensor("value_7_cast_fp16")]; + tensor var_1086_to_fp16 = const()[name = tensor("op_1086_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183662912)))]; + tensor query_15_cast_fp16 = add(x = query_13_cast_fp16, y = var_1086_to_fp16)[name = tensor("query_15_cast_fp16")]; + tensor var_1089_to_fp16 = const()[name = tensor("op_1089_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183665024)))]; + tensor q_with_bias_v_7_cast_fp16 = add(x = query_13_cast_fp16, y = var_1089_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 = const()[name = tensor("layers_3_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183667136)))]; + tensor p_7_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_7_cast_fp16")]; + tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1, 8, 128, 188])]; + tensor var_1101_cast_fp16 = reshape(shape = var_1100, x = q_with_bias_v_7_cast_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102 = const()[name = tensor("op_1102"), val = tensor([1, 8, 128, -1])]; + tensor var_1103_cast_fp16 = reshape(shape = var_1102, x = p_7_cast_fp16)[name = tensor("op_1103_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_1101_cast_fp16, y = var_1103_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_1112 = const()[name = tensor("op_1112"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_29_cast_fp16 = reshape(shape = var_1112, x = matrix_bd_27_cast_fp16)[name = tensor("matrix_bd_29_cast_fp16")]; + tensor var_1116_begin_0 = const()[name = tensor("op_1116_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1116_end_0 = const()[name = tensor("op_1116_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1116_end_mask_0 = const()[name = tensor("op_1116_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1116_cast_fp16 = slice_by_index(begin = var_1116_begin_0, end = var_1116_end_0, end_mask = var_1116_end_mask_0, x = matrix_bd_29_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_31_cast_fp16 = reshape(shape = var_1117, x = var_1116_cast_fp16)[name = tensor("matrix_bd_31_cast_fp16")]; + tensor var_1122_begin_0 = const()[name = tensor("op_1122_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1122_end_0 = const()[name = tensor("op_1122_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1122_end_mask_0 = const()[name = tensor("op_1122_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1122_cast_fp16 = slice_by_index(begin = var_1122_begin_0, end = var_1122_end_0, end_mask = var_1122_end_mask_0, x = matrix_bd_31_cast_fp16)[name = tensor("op_1122_cast_fp16")]; + tensor var_1123_to_fp16 = const()[name = tensor("op_1123_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_7_cast_fp16 = mul(x = var_1122_cast_fp16, y = var_1123_to_fp16)[name = tensor("qk_mask_7_cast_fp16")]; + tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_7_cast_fp16 = reshape(shape = var_1127, x = query_15_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; + tensor var_1129_to_fp16 = const()[name = tensor("op_1129_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1130_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_1129_to_fp16)[name = tensor("op_1130_cast_fp16")]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 8, 128, 188])]; + tensor var_1134_cast_fp16 = reshape(shape = var_1133, x = key_7_cast_fp16)[name = tensor("op_1134_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_1130_cast_fp16, y = var_1134_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_1138_cast_fp16 = softmax(axis = var_979, x = mh_w_15_cast_fp16)[name = tensor("op_1138_cast_fp16")]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([1, 8, 128, 188])]; + tensor var_1140_cast_fp16 = reshape(shape = var_1139, x = value_7_cast_fp16)[name = tensor("op_1140_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_1140_cast_fp16, y = var_1138_cast_fp16)[name = tensor("attn_7_cast_fp16")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1024, 1, 188])]; + tensor input_99_cast_fp16 = reshape(shape = var_1143, x = attn_7_cast_fp16)[name = tensor("input_99_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 = const()[name = tensor("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185764352)))]; + tensor obj_17_cast_fp16 = conv(bias = input_15_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, x = input_99_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_1161_to_fp16 = const()[name = tensor("op_1161_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_35_cast_fp16 = layer_norm(axes = out_35_axes_0, epsilon = var_1161_to_fp16, x = inputs_35_cast_fp16)[name = tensor("out_35_cast_fp16")]; + tensor input_101_gamma_0_to_fp16 = const()[name = tensor("input_101_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187861568)))]; + tensor input_101_beta_0_to_fp16 = const()[name = tensor("input_101_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187863680)))]; + tensor input_101_epsilon_0_to_fp16 = const()[name = tensor("input_101_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = batch_norm(beta = input_101_beta_0_to_fp16, epsilon = input_101_epsilon_0_to_fp16, gamma = input_101_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_35_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor input_103_pad_type_0 = const()[name = tensor("input_103_pad_type_0"), val = tensor("valid")]; + tensor input_103_strides_0 = const()[name = tensor("input_103_strides_0"), val = tensor([1, 1])]; + tensor input_103_pad_0 = const()[name = tensor("input_103_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_103_dilations_0 = const()[name = tensor("input_103_dilations_0"), val = tensor([1, 1])]; + tensor input_103_groups_0 = const()[name = tensor("input_103_groups_0"), val = tensor(1)]; + tensor layers_3_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_3_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187865792)))]; + tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = layers_3_conv_pointwise_conv1_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_split_num_splits_0 = const()[name = tensor("input_105_split_num_splits_0"), val = tensor(2)]; + tensor input_105_split_axis_0 = const()[name = tensor("input_105_split_axis_0"), val = tensor(1)]; + tensor input_105_split_cast_fp16_0, tensor input_105_split_cast_fp16_1 = split(axis = input_105_split_axis_0, num_splits = input_105_split_num_splits_0, x = input_103_cast_fp16)[name = tensor("input_105_split_cast_fp16")]; + tensor input_105_split_1_sigmoid_cast_fp16 = sigmoid(x = input_105_split_cast_fp16_1)[name = tensor("input_105_split_1_sigmoid_cast_fp16")]; + tensor input_105_cast_fp16 = mul(x = input_105_split_cast_fp16_0, y = input_105_split_1_sigmoid_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_pad_type_0 = const()[name = tensor("input_107_pad_type_0"), val = tensor("custom")]; + tensor input_107_pad_0 = const()[name = tensor("input_107_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_107_groups_0 = const()[name = tensor("input_107_groups_0"), val = tensor(1024)]; + tensor input_107_strides_0 = const()[name = tensor("input_107_strides_0"), val = tensor([1, 1])]; + tensor input_107_dilations_0 = const()[name = tensor("input_107_dilations_0"), val = tensor([1, 1])]; + tensor const_274_to_fp16 = const()[name = tensor("const_274_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192060160)))]; + tensor const_275_to_fp16 = const()[name = tensor("const_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192078656)))]; + tensor input_109_cast_fp16 = conv(bias = const_275_to_fp16, dilations = input_107_dilations_0, groups = input_107_groups_0, pad = input_107_pad_0, pad_type = input_107_pad_type_0, strides = input_107_strides_0, weight = const_274_to_fp16, x = input_105_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_cast_fp16 = silu(x = input_109_cast_fp16)[name = tensor("input_111_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 = const()[name = tensor("layers_3_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192080768)))]; + 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, x = input_111_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_1209_to_fp16 = const()[name = tensor("op_1209_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_37_cast_fp16 = layer_norm(axes = out_37_axes_0, epsilon = var_1209_to_fp16, x = inputs_37_cast_fp16)[name = tensor("out_37_cast_fp16")]; + tensor input_113_gamma_0_to_fp16 = const()[name = tensor("input_113_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194177984)))]; + tensor input_113_beta_0_to_fp16 = const()[name = tensor("input_113_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194180096)))]; + tensor input_113_epsilon_0_to_fp16 = const()[name = tensor("input_113_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = batch_norm(beta = input_113_beta_0_to_fp16, epsilon = input_113_epsilon_0_to_fp16, gamma = input_113_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_37_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor input_115_pad_type_0 = const()[name = tensor("input_115_pad_type_0"), val = tensor("valid")]; + tensor input_115_strides_0 = const()[name = tensor("input_115_strides_0"), val = tensor([1, 1])]; + tensor input_115_pad_0 = const()[name = tensor("input_115_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_115_dilations_0 = const()[name = tensor("input_115_dilations_0"), val = tensor([1, 1])]; + tensor input_115_groups_0 = const()[name = tensor("input_115_groups_0"), val = tensor(1)]; + tensor layers_3_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_3_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194182208)))]; + tensor input_115_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_115_dilations_0, groups = input_115_groups_0, pad = input_115_pad_0, pad_type = input_115_pad_type_0, strides = input_115_strides_0, weight = layers_3_feed_forward2_fc1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_cast_fp16 = silu(x = input_115_cast_fp16)[name = tensor("input_117_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 var_1237_weight_0_to_fp16 = const()[name = tensor("op_1237_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202570880)))]; + tensor var_1237_cast_fp16 = conv(bias = input_15_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 = var_1237_weight_0_to_fp16, x = input_117_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor inputs_39_cast_fp16 = add(x = inputs_37_cast_fp16, y = var_1237_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_1247_to_fp16 = const()[name = tensor("op_1247_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_39_cast_fp16 = layer_norm(axes = out_39_axes_0, epsilon = var_1247_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(210959552)))]; + 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(210961664)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_39_cast_fp16)[name = tensor("inputs_41_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor(3)]; + tensor out_41_axes_0 = const()[name = tensor("out_41_axes_0"), val = tensor([1])]; + tensor var_1292_to_fp16 = const()[name = tensor("op_1292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_41_cast_fp16 = layer_norm(axes = out_41_axes_0, epsilon = var_1292_to_fp16, x = inputs_41_cast_fp16)[name = tensor("out_41_cast_fp16")]; + tensor input_119_gamma_0_to_fp16 = const()[name = tensor("input_119_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210963776)))]; + tensor input_119_beta_0_to_fp16 = const()[name = tensor("input_119_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210965888)))]; + tensor input_119_epsilon_0_to_fp16 = const()[name = tensor("input_119_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = batch_norm(beta = input_119_beta_0_to_fp16, epsilon = input_119_epsilon_0_to_fp16, gamma = input_119_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_41_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor input_121_pad_type_0 = const()[name = tensor("input_121_pad_type_0"), val = tensor("valid")]; + tensor input_121_strides_0 = const()[name = tensor("input_121_strides_0"), val = tensor([1, 1])]; + tensor input_121_pad_0 = const()[name = tensor("input_121_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_121_dilations_0 = const()[name = tensor("input_121_dilations_0"), val = tensor([1, 1])]; + tensor input_121_groups_0 = const()[name = tensor("input_121_groups_0"), val = tensor(1)]; + tensor layers_4_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_4_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210968000)))]; + tensor input_121_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_121_dilations_0, groups = input_121_groups_0, pad = input_121_pad_0, pad_type = input_121_pad_type_0, strides = input_121_strides_0, weight = layers_4_feed_forward1_fc1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_cast_fp16 = silu(x = input_121_cast_fp16)[name = tensor("input_123_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 var_1320_weight_0_to_fp16 = const()[name = tensor("op_1320_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219356672)))]; + tensor var_1320_cast_fp16 = conv(bias = input_15_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 = var_1320_weight_0_to_fp16, x = input_123_cast_fp16)[name = tensor("op_1320_cast_fp16")]; + tensor inputs_43_cast_fp16 = add(x = inputs_41_cast_fp16, y = var_1320_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_1330_to_fp16 = const()[name = tensor("op_1330_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_43_cast_fp16 = layer_norm(axes = out_43_axes_0, epsilon = var_1330_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(227745344)))]; + 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(227747456)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227749568)))]; + tensor query_17_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229846784)))]; + 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, 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 = const()[name = tensor("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231944000)))]; + tensor value_9_cast_fp16 = conv(bias = input_15_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, x = obj_19_cast_fp16)[name = tensor("value_9_cast_fp16")]; + tensor var_1368_to_fp16 = const()[name = tensor("op_1368_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234041216)))]; + tensor query_19_cast_fp16 = add(x = query_17_cast_fp16, y = var_1368_to_fp16)[name = tensor("query_19_cast_fp16")]; + tensor var_1371_to_fp16 = const()[name = tensor("op_1371_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234043328)))]; + tensor q_with_bias_v_9_cast_fp16 = add(x = query_17_cast_fp16, y = var_1371_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 = const()[name = tensor("layers_4_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(234045440)))]; + tensor p_9_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_9_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([1, 8, 128, 188])]; + tensor var_1383_cast_fp16 = reshape(shape = var_1382, x = q_with_bias_v_9_cast_fp16)[name = tensor("op_1383_cast_fp16")]; + tensor var_1384 = const()[name = tensor("op_1384"), val = tensor([1, 8, 128, -1])]; + tensor var_1385_cast_fp16 = reshape(shape = var_1384, x = p_9_cast_fp16)[name = tensor("op_1385_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_1383_cast_fp16, y = var_1385_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_1394 = const()[name = tensor("op_1394"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_37_cast_fp16 = reshape(shape = var_1394, x = matrix_bd_35_cast_fp16)[name = tensor("matrix_bd_37_cast_fp16")]; + tensor var_1398_begin_0 = const()[name = tensor("op_1398_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1398_end_0 = const()[name = tensor("op_1398_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1398_end_mask_0 = const()[name = tensor("op_1398_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1398_cast_fp16 = slice_by_index(begin = var_1398_begin_0, end = var_1398_end_0, end_mask = var_1398_end_mask_0, x = matrix_bd_37_cast_fp16)[name = tensor("op_1398_cast_fp16")]; + tensor var_1399 = const()[name = tensor("op_1399"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_39_cast_fp16 = reshape(shape = var_1399, x = var_1398_cast_fp16)[name = tensor("matrix_bd_39_cast_fp16")]; + tensor var_1404_begin_0 = const()[name = tensor("op_1404_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1404_end_0 = const()[name = tensor("op_1404_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1404_end_mask_0 = const()[name = tensor("op_1404_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1404_cast_fp16 = slice_by_index(begin = var_1404_begin_0, end = var_1404_end_0, end_mask = var_1404_end_mask_0, x = matrix_bd_39_cast_fp16)[name = tensor("op_1404_cast_fp16")]; + tensor var_1405_to_fp16 = const()[name = tensor("op_1405_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_9_cast_fp16 = mul(x = var_1404_cast_fp16, y = var_1405_to_fp16)[name = tensor("qk_mask_9_cast_fp16")]; + tensor var_1409 = const()[name = tensor("op_1409"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_9_cast_fp16 = reshape(shape = var_1409, x = query_19_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; + tensor var_1411_to_fp16 = const()[name = tensor("op_1411_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1412_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_1411_to_fp16)[name = tensor("op_1412_cast_fp16")]; + tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([1, 8, 128, 188])]; + tensor var_1416_cast_fp16 = reshape(shape = var_1415, x = key_9_cast_fp16)[name = tensor("op_1416_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_1412_cast_fp16, y = var_1416_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_1420_cast_fp16 = softmax(axis = var_1261, x = mh_w_19_cast_fp16)[name = tensor("op_1420_cast_fp16")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([1, 8, 128, 188])]; + tensor var_1422_cast_fp16 = reshape(shape = var_1421, x = value_9_cast_fp16)[name = tensor("op_1422_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_1422_cast_fp16, y = var_1420_cast_fp16)[name = tensor("attn_9_cast_fp16")]; + tensor var_1425 = const()[name = tensor("op_1425"), val = tensor([1, 1024, 1, 188])]; + tensor input_125_cast_fp16 = reshape(shape = var_1425, x = attn_9_cast_fp16)[name = tensor("input_125_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 = const()[name = tensor("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236142656)))]; + tensor obj_21_cast_fp16 = conv(bias = input_15_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, x = input_125_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_1443_to_fp16 = const()[name = tensor("op_1443_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_45_cast_fp16 = layer_norm(axes = out_45_axes_0, epsilon = var_1443_to_fp16, x = inputs_45_cast_fp16)[name = tensor("out_45_cast_fp16")]; + tensor input_127_gamma_0_to_fp16 = const()[name = tensor("input_127_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238239872)))]; + tensor input_127_beta_0_to_fp16 = const()[name = tensor("input_127_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238241984)))]; + tensor input_127_epsilon_0_to_fp16 = const()[name = tensor("input_127_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_127_cast_fp16 = batch_norm(beta = input_127_beta_0_to_fp16, epsilon = input_127_epsilon_0_to_fp16, gamma = input_127_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_45_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_pad_type_0 = const()[name = tensor("input_129_pad_type_0"), val = tensor("valid")]; + tensor input_129_strides_0 = const()[name = tensor("input_129_strides_0"), val = tensor([1, 1])]; + tensor input_129_pad_0 = const()[name = tensor("input_129_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_129_dilations_0 = const()[name = tensor("input_129_dilations_0"), val = tensor([1, 1])]; + tensor input_129_groups_0 = const()[name = tensor("input_129_groups_0"), val = tensor(1)]; + tensor layers_4_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_4_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(238244096)))]; + tensor input_129_cast_fp16 = conv(dilations = input_129_dilations_0, groups = input_129_groups_0, pad = input_129_pad_0, pad_type = input_129_pad_type_0, strides = input_129_strides_0, weight = layers_4_conv_pointwise_conv1_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_split_num_splits_0 = const()[name = tensor("input_131_split_num_splits_0"), val = tensor(2)]; + tensor input_131_split_axis_0 = const()[name = tensor("input_131_split_axis_0"), val = tensor(1)]; + tensor input_131_split_cast_fp16_0, tensor input_131_split_cast_fp16_1 = split(axis = input_131_split_axis_0, num_splits = input_131_split_num_splits_0, x = input_129_cast_fp16)[name = tensor("input_131_split_cast_fp16")]; + tensor input_131_split_1_sigmoid_cast_fp16 = sigmoid(x = input_131_split_cast_fp16_1)[name = tensor("input_131_split_1_sigmoid_cast_fp16")]; + tensor input_131_cast_fp16 = mul(x = input_131_split_cast_fp16_0, y = input_131_split_1_sigmoid_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor input_133_pad_type_0 = const()[name = tensor("input_133_pad_type_0"), val = tensor("custom")]; + tensor input_133_pad_0 = const()[name = tensor("input_133_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_133_groups_0 = const()[name = tensor("input_133_groups_0"), val = tensor(1024)]; + tensor input_133_strides_0 = const()[name = tensor("input_133_strides_0"), val = tensor([1, 1])]; + tensor input_133_dilations_0 = const()[name = tensor("input_133_dilations_0"), val = tensor([1, 1])]; + tensor const_276_to_fp16 = const()[name = tensor("const_276_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242438464)))]; + tensor const_277_to_fp16 = const()[name = tensor("const_277_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242456960)))]; + tensor input_135_cast_fp16 = conv(bias = const_277_to_fp16, dilations = input_133_dilations_0, groups = input_133_groups_0, pad = input_133_pad_0, pad_type = input_133_pad_type_0, strides = input_133_strides_0, weight = const_276_to_fp16, x = input_131_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_cast_fp16 = silu(x = input_135_cast_fp16)[name = tensor("input_137_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 = const()[name = tensor("layers_4_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242459072)))]; + 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, x = input_137_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_1491_to_fp16 = const()[name = tensor("op_1491_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_47_cast_fp16 = layer_norm(axes = out_47_axes_0, epsilon = var_1491_to_fp16, x = inputs_47_cast_fp16)[name = tensor("out_47_cast_fp16")]; + tensor input_139_gamma_0_to_fp16 = const()[name = tensor("input_139_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244556288)))]; + tensor input_139_beta_0_to_fp16 = const()[name = tensor("input_139_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244558400)))]; + tensor input_139_epsilon_0_to_fp16 = const()[name = tensor("input_139_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_139_cast_fp16 = batch_norm(beta = input_139_beta_0_to_fp16, epsilon = input_139_epsilon_0_to_fp16, gamma = input_139_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_47_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_pad_type_0 = const()[name = tensor("input_141_pad_type_0"), val = tensor("valid")]; + tensor input_141_strides_0 = const()[name = tensor("input_141_strides_0"), val = tensor([1, 1])]; + tensor input_141_pad_0 = const()[name = tensor("input_141_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_141_dilations_0 = const()[name = tensor("input_141_dilations_0"), val = tensor([1, 1])]; + tensor input_141_groups_0 = const()[name = tensor("input_141_groups_0"), val = tensor(1)]; + tensor layers_4_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_4_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244560512)))]; + tensor input_141_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_141_dilations_0, groups = input_141_groups_0, pad = input_141_pad_0, pad_type = input_141_pad_type_0, strides = input_141_strides_0, weight = layers_4_feed_forward2_fc1_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_cast_fp16 = silu(x = input_141_cast_fp16)[name = tensor("input_143_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 var_1519_weight_0_to_fp16 = const()[name = tensor("op_1519_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252949184)))]; + tensor var_1519_cast_fp16 = conv(bias = input_15_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 = var_1519_weight_0_to_fp16, x = input_143_cast_fp16)[name = tensor("op_1519_cast_fp16")]; + tensor inputs_49_cast_fp16 = add(x = inputs_47_cast_fp16, y = var_1519_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_1529_to_fp16 = const()[name = tensor("op_1529_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_49_cast_fp16 = layer_norm(axes = out_49_axes_0, epsilon = var_1529_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(261337856)))]; + 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(261339968)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_49_cast_fp16)[name = tensor("inputs_51_cast_fp16")]; + tensor var_1543 = const()[name = tensor("op_1543"), val = tensor(3)]; + tensor out_51_axes_0 = const()[name = tensor("out_51_axes_0"), val = tensor([1])]; + tensor var_1574_to_fp16 = const()[name = tensor("op_1574_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_51_cast_fp16 = layer_norm(axes = out_51_axes_0, epsilon = var_1574_to_fp16, x = inputs_51_cast_fp16)[name = tensor("out_51_cast_fp16")]; + tensor input_145_gamma_0_to_fp16 = const()[name = tensor("input_145_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261342080)))]; + tensor input_145_beta_0_to_fp16 = const()[name = tensor("input_145_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261344192)))]; + tensor input_145_epsilon_0_to_fp16 = const()[name = tensor("input_145_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_145_cast_fp16 = batch_norm(beta = input_145_beta_0_to_fp16, epsilon = input_145_epsilon_0_to_fp16, gamma = input_145_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_51_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_pad_type_0 = const()[name = tensor("input_147_pad_type_0"), val = tensor("valid")]; + tensor input_147_strides_0 = const()[name = tensor("input_147_strides_0"), val = tensor([1, 1])]; + tensor input_147_pad_0 = const()[name = tensor("input_147_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_147_dilations_0 = const()[name = tensor("input_147_dilations_0"), val = tensor([1, 1])]; + tensor input_147_groups_0 = const()[name = tensor("input_147_groups_0"), val = tensor(1)]; + tensor layers_5_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_5_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261346304)))]; + tensor input_147_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_147_dilations_0, groups = input_147_groups_0, pad = input_147_pad_0, pad_type = input_147_pad_type_0, strides = input_147_strides_0, weight = layers_5_feed_forward1_fc1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_cast_fp16 = silu(x = input_147_cast_fp16)[name = tensor("input_149_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 var_1602_weight_0_to_fp16 = const()[name = tensor("op_1602_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269734976)))]; + tensor var_1602_cast_fp16 = conv(bias = input_15_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 = var_1602_weight_0_to_fp16, x = input_149_cast_fp16)[name = tensor("op_1602_cast_fp16")]; + tensor inputs_53_cast_fp16 = add(x = inputs_51_cast_fp16, y = var_1602_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_1612_to_fp16 = const()[name = tensor("op_1612_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_53_cast_fp16 = layer_norm(axes = out_53_axes_0, epsilon = var_1612_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(278123648)))]; + 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(278125760)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278127872)))]; + tensor query_21_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280225088)))]; + 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, 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 = const()[name = tensor("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(282322304)))]; + tensor value_11_cast_fp16 = conv(bias = input_15_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, x = obj_23_cast_fp16)[name = tensor("value_11_cast_fp16")]; + tensor var_1650_to_fp16 = const()[name = tensor("op_1650_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284419520)))]; + tensor query_23_cast_fp16 = add(x = query_21_cast_fp16, y = var_1650_to_fp16)[name = tensor("query_23_cast_fp16")]; + tensor var_1653_to_fp16 = const()[name = tensor("op_1653_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284421632)))]; + tensor q_with_bias_v_11_cast_fp16 = add(x = query_21_cast_fp16, y = var_1653_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 = const()[name = tensor("layers_5_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284423744)))]; + tensor p_11_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_11_cast_fp16")]; + tensor var_1664 = const()[name = tensor("op_1664"), val = tensor([1, 8, 128, 188])]; + tensor var_1665_cast_fp16 = reshape(shape = var_1664, x = q_with_bias_v_11_cast_fp16)[name = tensor("op_1665_cast_fp16")]; + tensor var_1666 = const()[name = tensor("op_1666"), val = tensor([1, 8, 128, -1])]; + tensor var_1667_cast_fp16 = reshape(shape = var_1666, x = p_11_cast_fp16)[name = tensor("op_1667_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_1665_cast_fp16, y = var_1667_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_1676 = const()[name = tensor("op_1676"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_45_cast_fp16 = reshape(shape = var_1676, x = matrix_bd_43_cast_fp16)[name = tensor("matrix_bd_45_cast_fp16")]; + tensor var_1680_begin_0 = const()[name = tensor("op_1680_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1680_end_0 = const()[name = tensor("op_1680_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1680_end_mask_0 = const()[name = tensor("op_1680_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1680_cast_fp16 = slice_by_index(begin = var_1680_begin_0, end = var_1680_end_0, end_mask = var_1680_end_mask_0, x = matrix_bd_45_cast_fp16)[name = tensor("op_1680_cast_fp16")]; + tensor var_1681 = const()[name = tensor("op_1681"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_47_cast_fp16 = reshape(shape = var_1681, x = var_1680_cast_fp16)[name = tensor("matrix_bd_47_cast_fp16")]; + tensor var_1686_begin_0 = const()[name = tensor("op_1686_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1686_end_0 = const()[name = tensor("op_1686_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1686_end_mask_0 = const()[name = tensor("op_1686_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1686_cast_fp16 = slice_by_index(begin = var_1686_begin_0, end = var_1686_end_0, end_mask = var_1686_end_mask_0, x = matrix_bd_47_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_to_fp16 = const()[name = tensor("op_1687_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_11_cast_fp16 = mul(x = var_1686_cast_fp16, y = var_1687_to_fp16)[name = tensor("qk_mask_11_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_11_cast_fp16 = reshape(shape = var_1691, x = query_23_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; + tensor var_1693_to_fp16 = const()[name = tensor("op_1693_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1694_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1693_to_fp16)[name = tensor("op_1694_cast_fp16")]; + tensor var_1697 = const()[name = tensor("op_1697"), val = tensor([1, 8, 128, 188])]; + tensor var_1698_cast_fp16 = reshape(shape = var_1697, x = key_11_cast_fp16)[name = tensor("op_1698_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_1694_cast_fp16, y = var_1698_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_1702_cast_fp16 = softmax(axis = var_1543, x = mh_w_23_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor var_1703 = const()[name = tensor("op_1703"), val = tensor([1, 8, 128, 188])]; + tensor var_1704_cast_fp16 = reshape(shape = var_1703, x = value_11_cast_fp16)[name = tensor("op_1704_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_1704_cast_fp16, y = var_1702_cast_fp16)[name = tensor("attn_11_cast_fp16")]; + tensor var_1707 = const()[name = tensor("op_1707"), val = tensor([1, 1024, 1, 188])]; + tensor input_151_cast_fp16 = reshape(shape = var_1707, x = attn_11_cast_fp16)[name = tensor("input_151_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 = const()[name = tensor("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286520960)))]; + tensor obj_25_cast_fp16 = conv(bias = input_15_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, x = input_151_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_1725_to_fp16 = const()[name = tensor("op_1725_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_55_cast_fp16 = layer_norm(axes = out_55_axes_0, epsilon = var_1725_to_fp16, x = inputs_55_cast_fp16)[name = tensor("out_55_cast_fp16")]; + tensor input_153_gamma_0_to_fp16 = const()[name = tensor("input_153_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288618176)))]; + tensor input_153_beta_0_to_fp16 = const()[name = tensor("input_153_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288620288)))]; + tensor input_153_epsilon_0_to_fp16 = const()[name = tensor("input_153_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = batch_norm(beta = input_153_beta_0_to_fp16, epsilon = input_153_epsilon_0_to_fp16, gamma = input_153_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_55_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_pad_type_0 = const()[name = tensor("input_155_pad_type_0"), val = tensor("valid")]; + tensor input_155_strides_0 = const()[name = tensor("input_155_strides_0"), val = tensor([1, 1])]; + tensor input_155_pad_0 = const()[name = tensor("input_155_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_155_dilations_0 = const()[name = tensor("input_155_dilations_0"), val = tensor([1, 1])]; + tensor input_155_groups_0 = const()[name = tensor("input_155_groups_0"), val = tensor(1)]; + tensor layers_5_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_5_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(288622400)))]; + tensor input_155_cast_fp16 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = layers_5_conv_pointwise_conv1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor input_157_split_num_splits_0 = const()[name = tensor("input_157_split_num_splits_0"), val = tensor(2)]; + tensor input_157_split_axis_0 = const()[name = tensor("input_157_split_axis_0"), val = tensor(1)]; + tensor input_157_split_cast_fp16_0, tensor input_157_split_cast_fp16_1 = split(axis = input_157_split_axis_0, num_splits = input_157_split_num_splits_0, x = input_155_cast_fp16)[name = tensor("input_157_split_cast_fp16")]; + tensor input_157_split_1_sigmoid_cast_fp16 = sigmoid(x = input_157_split_cast_fp16_1)[name = tensor("input_157_split_1_sigmoid_cast_fp16")]; + tensor input_157_cast_fp16 = mul(x = input_157_split_cast_fp16_0, y = input_157_split_1_sigmoid_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_pad_type_0 = const()[name = tensor("input_159_pad_type_0"), val = tensor("custom")]; + tensor input_159_pad_0 = const()[name = tensor("input_159_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_159_groups_0 = const()[name = tensor("input_159_groups_0"), val = tensor(1024)]; + tensor input_159_strides_0 = const()[name = tensor("input_159_strides_0"), val = tensor([1, 1])]; + tensor input_159_dilations_0 = const()[name = tensor("input_159_dilations_0"), val = tensor([1, 1])]; + tensor const_278_to_fp16 = const()[name = tensor("const_278_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292816768)))]; + tensor const_279_to_fp16 = const()[name = tensor("const_279_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292835264)))]; + tensor input_161_cast_fp16 = conv(bias = const_279_to_fp16, dilations = input_159_dilations_0, groups = input_159_groups_0, pad = input_159_pad_0, pad_type = input_159_pad_type_0, strides = input_159_strides_0, weight = const_278_to_fp16, x = input_157_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor input_163_cast_fp16 = silu(x = input_161_cast_fp16)[name = tensor("input_163_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 = const()[name = tensor("layers_5_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292837376)))]; + 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, x = input_163_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_1773_to_fp16 = const()[name = tensor("op_1773_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_57_cast_fp16 = layer_norm(axes = out_57_axes_0, epsilon = var_1773_to_fp16, x = inputs_57_cast_fp16)[name = tensor("out_57_cast_fp16")]; + tensor input_165_gamma_0_to_fp16 = const()[name = tensor("input_165_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294934592)))]; + tensor input_165_beta_0_to_fp16 = const()[name = tensor("input_165_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294936704)))]; + tensor input_165_epsilon_0_to_fp16 = const()[name = tensor("input_165_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = batch_norm(beta = input_165_beta_0_to_fp16, epsilon = input_165_epsilon_0_to_fp16, gamma = input_165_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_57_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_pad_type_0 = const()[name = tensor("input_167_pad_type_0"), val = tensor("valid")]; + tensor input_167_strides_0 = const()[name = tensor("input_167_strides_0"), val = tensor([1, 1])]; + tensor input_167_pad_0 = const()[name = tensor("input_167_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_167_dilations_0 = const()[name = tensor("input_167_dilations_0"), val = tensor([1, 1])]; + tensor input_167_groups_0 = const()[name = tensor("input_167_groups_0"), val = tensor(1)]; + tensor layers_5_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_5_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294938816)))]; + tensor input_167_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_167_dilations_0, groups = input_167_groups_0, pad = input_167_pad_0, pad_type = input_167_pad_type_0, strides = input_167_strides_0, weight = layers_5_feed_forward2_fc1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_cast_fp16 = silu(x = input_167_cast_fp16)[name = tensor("input_169_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 var_1801_weight_0_to_fp16 = const()[name = tensor("op_1801_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303327488)))]; + tensor var_1801_cast_fp16 = conv(bias = input_15_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 = var_1801_weight_0_to_fp16, x = input_169_cast_fp16)[name = tensor("op_1801_cast_fp16")]; + tensor inputs_59_cast_fp16 = add(x = inputs_57_cast_fp16, y = var_1801_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_1811_to_fp16 = const()[name = tensor("op_1811_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_59_cast_fp16 = layer_norm(axes = out_59_axes_0, epsilon = var_1811_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(311716160)))]; + 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(311718272)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_59_cast_fp16)[name = tensor("inputs_61_cast_fp16")]; + tensor var_1825 = const()[name = tensor("op_1825"), val = tensor(3)]; + tensor out_61_axes_0 = const()[name = tensor("out_61_axes_0"), val = tensor([1])]; + tensor var_1856_to_fp16 = const()[name = tensor("op_1856_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_61_cast_fp16 = layer_norm(axes = out_61_axes_0, epsilon = var_1856_to_fp16, x = inputs_61_cast_fp16)[name = tensor("out_61_cast_fp16")]; + tensor input_171_gamma_0_to_fp16 = const()[name = tensor("input_171_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311720384)))]; + tensor input_171_beta_0_to_fp16 = const()[name = tensor("input_171_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311722496)))]; + tensor input_171_epsilon_0_to_fp16 = const()[name = tensor("input_171_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = batch_norm(beta = input_171_beta_0_to_fp16, epsilon = input_171_epsilon_0_to_fp16, gamma = input_171_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_61_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_pad_type_0 = const()[name = tensor("input_173_pad_type_0"), val = tensor("valid")]; + tensor input_173_strides_0 = const()[name = tensor("input_173_strides_0"), val = tensor([1, 1])]; + tensor input_173_pad_0 = const()[name = tensor("input_173_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_173_dilations_0 = const()[name = tensor("input_173_dilations_0"), val = tensor([1, 1])]; + tensor input_173_groups_0 = const()[name = tensor("input_173_groups_0"), val = tensor(1)]; + tensor layers_6_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_6_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311724608)))]; + tensor input_173_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_173_dilations_0, groups = input_173_groups_0, pad = input_173_pad_0, pad_type = input_173_pad_type_0, strides = input_173_strides_0, weight = layers_6_feed_forward1_fc1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor input_175_cast_fp16 = silu(x = input_173_cast_fp16)[name = tensor("input_175_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 var_1884_weight_0_to_fp16 = const()[name = tensor("op_1884_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320113280)))]; + tensor var_1884_cast_fp16 = conv(bias = input_15_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 = var_1884_weight_0_to_fp16, x = input_175_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor inputs_63_cast_fp16 = add(x = inputs_61_cast_fp16, y = var_1884_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_1894_to_fp16 = const()[name = tensor("op_1894_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_63_cast_fp16 = layer_norm(axes = out_63_axes_0, epsilon = var_1894_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(328501952)))]; + 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(328504064)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328506176)))]; + tensor query_25_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(330603392)))]; + 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, 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 = const()[name = tensor("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(332700608)))]; + tensor value_13_cast_fp16 = conv(bias = input_15_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, x = obj_27_cast_fp16)[name = tensor("value_13_cast_fp16")]; + tensor var_1932_to_fp16 = const()[name = tensor("op_1932_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334797824)))]; + tensor query_27_cast_fp16 = add(x = query_25_cast_fp16, y = var_1932_to_fp16)[name = tensor("query_27_cast_fp16")]; + tensor var_1935_to_fp16 = const()[name = tensor("op_1935_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334799936)))]; + tensor q_with_bias_v_13_cast_fp16 = add(x = query_25_cast_fp16, y = var_1935_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 = const()[name = tensor("layers_6_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334802048)))]; + tensor p_13_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_13_cast_fp16")]; + tensor var_1946 = const()[name = tensor("op_1946"), val = tensor([1, 8, 128, 188])]; + tensor var_1947_cast_fp16 = reshape(shape = var_1946, x = q_with_bias_v_13_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor var_1948 = const()[name = tensor("op_1948"), val = tensor([1, 8, 128, -1])]; + tensor var_1949_cast_fp16 = reshape(shape = var_1948, x = p_13_cast_fp16)[name = tensor("op_1949_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_1947_cast_fp16, y = var_1949_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_1958 = const()[name = tensor("op_1958"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_53_cast_fp16 = reshape(shape = var_1958, x = matrix_bd_51_cast_fp16)[name = tensor("matrix_bd_53_cast_fp16")]; + tensor var_1962_begin_0 = const()[name = tensor("op_1962_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_1962_end_0 = const()[name = tensor("op_1962_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_1962_end_mask_0 = const()[name = tensor("op_1962_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_1962_cast_fp16 = slice_by_index(begin = var_1962_begin_0, end = var_1962_end_0, end_mask = var_1962_end_mask_0, x = matrix_bd_53_cast_fp16)[name = tensor("op_1962_cast_fp16")]; + tensor var_1963 = const()[name = tensor("op_1963"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_55_cast_fp16 = reshape(shape = var_1963, x = var_1962_cast_fp16)[name = tensor("matrix_bd_55_cast_fp16")]; + tensor var_1968_begin_0 = const()[name = tensor("op_1968_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_1968_end_0 = const()[name = tensor("op_1968_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_1968_end_mask_0 = const()[name = tensor("op_1968_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_1968_cast_fp16 = slice_by_index(begin = var_1968_begin_0, end = var_1968_end_0, end_mask = var_1968_end_mask_0, x = matrix_bd_55_cast_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor var_1969_to_fp16 = const()[name = tensor("op_1969_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_13_cast_fp16 = mul(x = var_1968_cast_fp16, y = var_1969_to_fp16)[name = tensor("qk_mask_13_cast_fp16")]; + tensor var_1973 = const()[name = tensor("op_1973"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_13_cast_fp16 = reshape(shape = var_1973, x = query_27_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; + tensor var_1975_to_fp16 = const()[name = tensor("op_1975_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_1976_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1975_to_fp16)[name = tensor("op_1976_cast_fp16")]; + tensor var_1979 = const()[name = tensor("op_1979"), val = tensor([1, 8, 128, 188])]; + tensor var_1980_cast_fp16 = reshape(shape = var_1979, x = key_13_cast_fp16)[name = tensor("op_1980_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_1976_cast_fp16, y = var_1980_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_1984_cast_fp16 = softmax(axis = var_1825, x = mh_w_27_cast_fp16)[name = tensor("op_1984_cast_fp16")]; + tensor var_1985 = const()[name = tensor("op_1985"), val = tensor([1, 8, 128, 188])]; + tensor var_1986_cast_fp16 = reshape(shape = var_1985, x = value_13_cast_fp16)[name = tensor("op_1986_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_1986_cast_fp16, y = var_1984_cast_fp16)[name = tensor("attn_13_cast_fp16")]; + tensor var_1989 = const()[name = tensor("op_1989"), val = tensor([1, 1024, 1, 188])]; + tensor input_177_cast_fp16 = reshape(shape = var_1989, x = attn_13_cast_fp16)[name = tensor("input_177_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 = const()[name = tensor("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336899264)))]; + tensor obj_29_cast_fp16 = conv(bias = input_15_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, x = input_177_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_2007_to_fp16 = const()[name = tensor("op_2007_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_65_cast_fp16 = layer_norm(axes = out_65_axes_0, epsilon = var_2007_to_fp16, x = inputs_65_cast_fp16)[name = tensor("out_65_cast_fp16")]; + tensor input_179_gamma_0_to_fp16 = const()[name = tensor("input_179_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338996480)))]; + tensor input_179_beta_0_to_fp16 = const()[name = tensor("input_179_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(338998592)))]; + tensor input_179_epsilon_0_to_fp16 = const()[name = tensor("input_179_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = batch_norm(beta = input_179_beta_0_to_fp16, epsilon = input_179_epsilon_0_to_fp16, gamma = input_179_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_65_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor input_181_pad_type_0 = const()[name = tensor("input_181_pad_type_0"), val = tensor("valid")]; + tensor input_181_strides_0 = const()[name = tensor("input_181_strides_0"), val = tensor([1, 1])]; + tensor input_181_pad_0 = const()[name = tensor("input_181_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_181_dilations_0 = const()[name = tensor("input_181_dilations_0"), val = tensor([1, 1])]; + tensor input_181_groups_0 = const()[name = tensor("input_181_groups_0"), val = tensor(1)]; + tensor layers_6_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_6_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(339000704)))]; + tensor input_181_cast_fp16 = conv(dilations = input_181_dilations_0, groups = input_181_groups_0, pad = input_181_pad_0, pad_type = input_181_pad_type_0, strides = input_181_strides_0, weight = layers_6_conv_pointwise_conv1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_split_num_splits_0 = const()[name = tensor("input_183_split_num_splits_0"), val = tensor(2)]; + tensor input_183_split_axis_0 = const()[name = tensor("input_183_split_axis_0"), val = tensor(1)]; + tensor input_183_split_cast_fp16_0, tensor input_183_split_cast_fp16_1 = split(axis = input_183_split_axis_0, num_splits = input_183_split_num_splits_0, x = input_181_cast_fp16)[name = tensor("input_183_split_cast_fp16")]; + tensor input_183_split_1_sigmoid_cast_fp16 = sigmoid(x = input_183_split_cast_fp16_1)[name = tensor("input_183_split_1_sigmoid_cast_fp16")]; + tensor input_183_cast_fp16 = mul(x = input_183_split_cast_fp16_0, y = input_183_split_1_sigmoid_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_pad_type_0 = const()[name = tensor("input_185_pad_type_0"), val = tensor("custom")]; + tensor input_185_pad_0 = const()[name = tensor("input_185_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_185_groups_0 = const()[name = tensor("input_185_groups_0"), val = tensor(1024)]; + tensor input_185_strides_0 = const()[name = tensor("input_185_strides_0"), val = tensor([1, 1])]; + tensor input_185_dilations_0 = const()[name = tensor("input_185_dilations_0"), val = tensor([1, 1])]; + tensor const_280_to_fp16 = const()[name = tensor("const_280_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343195072)))]; + tensor const_281_to_fp16 = const()[name = tensor("const_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343213568)))]; + tensor input_187_cast_fp16 = conv(bias = const_281_to_fp16, dilations = input_185_dilations_0, groups = input_185_groups_0, pad = input_185_pad_0, pad_type = input_185_pad_type_0, strides = input_185_strides_0, weight = const_280_to_fp16, x = input_183_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = input_187_cast_fp16)[name = tensor("input_189_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 = const()[name = tensor("layers_6_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(343215680)))]; + 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, x = input_189_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_2055_to_fp16 = const()[name = tensor("op_2055_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_67_cast_fp16 = layer_norm(axes = out_67_axes_0, epsilon = var_2055_to_fp16, x = inputs_67_cast_fp16)[name = tensor("out_67_cast_fp16")]; + tensor input_191_gamma_0_to_fp16 = const()[name = tensor("input_191_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345312896)))]; + tensor input_191_beta_0_to_fp16 = const()[name = tensor("input_191_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345315008)))]; + tensor input_191_epsilon_0_to_fp16 = const()[name = tensor("input_191_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = batch_norm(beta = input_191_beta_0_to_fp16, epsilon = input_191_epsilon_0_to_fp16, gamma = input_191_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_67_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor input_193_pad_type_0 = const()[name = tensor("input_193_pad_type_0"), val = tensor("valid")]; + tensor input_193_strides_0 = const()[name = tensor("input_193_strides_0"), val = tensor([1, 1])]; + tensor input_193_pad_0 = const()[name = tensor("input_193_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_193_dilations_0 = const()[name = tensor("input_193_dilations_0"), val = tensor([1, 1])]; + tensor input_193_groups_0 = const()[name = tensor("input_193_groups_0"), val = tensor(1)]; + tensor layers_6_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_6_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345317120)))]; + tensor input_193_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_193_dilations_0, groups = input_193_groups_0, pad = input_193_pad_0, pad_type = input_193_pad_type_0, strides = input_193_strides_0, weight = layers_6_feed_forward2_fc1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_cast_fp16 = silu(x = input_193_cast_fp16)[name = tensor("input_195_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 var_2083_weight_0_to_fp16 = const()[name = tensor("op_2083_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353705792)))]; + tensor var_2083_cast_fp16 = conv(bias = input_15_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 = var_2083_weight_0_to_fp16, x = input_195_cast_fp16)[name = tensor("op_2083_cast_fp16")]; + tensor inputs_69_cast_fp16 = add(x = inputs_67_cast_fp16, y = var_2083_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_2093_to_fp16 = const()[name = tensor("op_2093_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_69_cast_fp16 = layer_norm(axes = out_69_axes_0, epsilon = var_2093_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(362094464)))]; + 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(362096576)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_69_cast_fp16)[name = tensor("inputs_71_cast_fp16")]; + tensor var_2107 = const()[name = tensor("op_2107"), val = tensor(3)]; + tensor out_71_axes_0 = const()[name = tensor("out_71_axes_0"), val = tensor([1])]; + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_71_cast_fp16 = layer_norm(axes = out_71_axes_0, epsilon = var_2138_to_fp16, x = inputs_71_cast_fp16)[name = tensor("out_71_cast_fp16")]; + tensor input_197_gamma_0_to_fp16 = const()[name = tensor("input_197_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362098688)))]; + tensor input_197_beta_0_to_fp16 = const()[name = tensor("input_197_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362100800)))]; + tensor input_197_epsilon_0_to_fp16 = const()[name = tensor("input_197_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = batch_norm(beta = input_197_beta_0_to_fp16, epsilon = input_197_epsilon_0_to_fp16, gamma = input_197_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_71_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_pad_type_0 = const()[name = tensor("input_199_pad_type_0"), val = tensor("valid")]; + tensor input_199_strides_0 = const()[name = tensor("input_199_strides_0"), val = tensor([1, 1])]; + tensor input_199_pad_0 = const()[name = tensor("input_199_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_199_dilations_0 = const()[name = tensor("input_199_dilations_0"), val = tensor([1, 1])]; + tensor input_199_groups_0 = const()[name = tensor("input_199_groups_0"), val = tensor(1)]; + tensor layers_7_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_7_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(362102912)))]; + tensor input_199_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_199_dilations_0, groups = input_199_groups_0, pad = input_199_pad_0, pad_type = input_199_pad_type_0, strides = input_199_strides_0, weight = layers_7_feed_forward1_fc1_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_cast_fp16 = silu(x = input_199_cast_fp16)[name = tensor("input_201_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 var_2166_weight_0_to_fp16 = const()[name = tensor("op_2166_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370491584)))]; + tensor var_2166_cast_fp16 = conv(bias = input_15_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 = var_2166_weight_0_to_fp16, x = input_201_cast_fp16)[name = tensor("op_2166_cast_fp16")]; + tensor inputs_73_cast_fp16 = add(x = inputs_71_cast_fp16, y = var_2166_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_2176_to_fp16 = const()[name = tensor("op_2176_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_73_cast_fp16 = layer_norm(axes = out_73_axes_0, epsilon = var_2176_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(378880256)))]; + 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(378882368)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378884480)))]; + tensor query_29_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(380981696)))]; + 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, 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 = const()[name = tensor("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(383078912)))]; + tensor value_15_cast_fp16 = conv(bias = input_15_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, x = obj_31_cast_fp16)[name = tensor("value_15_cast_fp16")]; + tensor var_2214_to_fp16 = const()[name = tensor("op_2214_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385176128)))]; + tensor query_31_cast_fp16 = add(x = query_29_cast_fp16, y = var_2214_to_fp16)[name = tensor("query_31_cast_fp16")]; + tensor var_2217_to_fp16 = const()[name = tensor("op_2217_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385178240)))]; + tensor q_with_bias_v_15_cast_fp16 = add(x = query_29_cast_fp16, y = var_2217_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 = const()[name = tensor("layers_7_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385180352)))]; + tensor p_15_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_15_cast_fp16")]; + tensor var_2228 = const()[name = tensor("op_2228"), val = tensor([1, 8, 128, 188])]; + tensor var_2229_cast_fp16 = reshape(shape = var_2228, x = q_with_bias_v_15_cast_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor var_2230 = const()[name = tensor("op_2230"), val = tensor([1, 8, 128, -1])]; + tensor var_2231_cast_fp16 = reshape(shape = var_2230, x = p_15_cast_fp16)[name = tensor("op_2231_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_2229_cast_fp16, y = var_2231_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_2240 = const()[name = tensor("op_2240"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_61_cast_fp16 = reshape(shape = var_2240, x = matrix_bd_59_cast_fp16)[name = tensor("matrix_bd_61_cast_fp16")]; + tensor var_2244_begin_0 = const()[name = tensor("op_2244_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2244_end_0 = const()[name = tensor("op_2244_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_2244_end_mask_0 = const()[name = tensor("op_2244_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2244_cast_fp16 = slice_by_index(begin = var_2244_begin_0, end = var_2244_end_0, end_mask = var_2244_end_mask_0, x = matrix_bd_61_cast_fp16)[name = tensor("op_2244_cast_fp16")]; + tensor var_2245 = const()[name = tensor("op_2245"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_63_cast_fp16 = reshape(shape = var_2245, x = var_2244_cast_fp16)[name = tensor("matrix_bd_63_cast_fp16")]; + tensor var_2250_begin_0 = const()[name = tensor("op_2250_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2250_end_0 = const()[name = tensor("op_2250_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_2250_end_mask_0 = const()[name = tensor("op_2250_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_2250_cast_fp16 = slice_by_index(begin = var_2250_begin_0, end = var_2250_end_0, end_mask = var_2250_end_mask_0, x = matrix_bd_63_cast_fp16)[name = tensor("op_2250_cast_fp16")]; + tensor var_2251_to_fp16 = const()[name = tensor("op_2251_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_15_cast_fp16 = mul(x = var_2250_cast_fp16, y = var_2251_to_fp16)[name = tensor("qk_mask_15_cast_fp16")]; + tensor var_2255 = const()[name = tensor("op_2255"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_15_cast_fp16 = reshape(shape = var_2255, x = query_31_cast_fp16)[name = tensor("mh_q_15_cast_fp16")]; + tensor var_2257_to_fp16 = const()[name = tensor("op_2257_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_2258_cast_fp16 = mul(x = mh_q_15_cast_fp16, y = var_2257_to_fp16)[name = tensor("op_2258_cast_fp16")]; + tensor var_2261 = const()[name = tensor("op_2261"), val = tensor([1, 8, 128, 188])]; + tensor var_2262_cast_fp16 = reshape(shape = var_2261, x = key_15_cast_fp16)[name = tensor("op_2262_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_2258_cast_fp16, y = var_2262_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_2266_cast_fp16 = softmax(axis = var_2107, x = mh_w_31_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor var_2267 = const()[name = tensor("op_2267"), val = tensor([1, 8, 128, 188])]; + tensor var_2268_cast_fp16 = reshape(shape = var_2267, x = value_15_cast_fp16)[name = tensor("op_2268_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_2268_cast_fp16, y = var_2266_cast_fp16)[name = tensor("attn_15_cast_fp16")]; + tensor var_2271 = const()[name = tensor("op_2271"), val = tensor([1, 1024, 1, 188])]; + tensor input_203_cast_fp16 = reshape(shape = var_2271, x = attn_15_cast_fp16)[name = tensor("input_203_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 = const()[name = tensor("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387277568)))]; + tensor obj_33_cast_fp16 = conv(bias = input_15_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, x = input_203_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_2289_to_fp16 = const()[name = tensor("op_2289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_75_cast_fp16 = layer_norm(axes = out_75_axes_0, epsilon = var_2289_to_fp16, x = inputs_75_cast_fp16)[name = tensor("out_75_cast_fp16")]; + tensor input_205_gamma_0_to_fp16 = const()[name = tensor("input_205_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389374784)))]; + tensor input_205_beta_0_to_fp16 = const()[name = tensor("input_205_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389376896)))]; + tensor input_205_epsilon_0_to_fp16 = const()[name = tensor("input_205_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_205_cast_fp16 = batch_norm(beta = input_205_beta_0_to_fp16, epsilon = input_205_epsilon_0_to_fp16, gamma = input_205_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_75_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_pad_type_0 = const()[name = tensor("input_207_pad_type_0"), val = tensor("valid")]; + tensor input_207_strides_0 = const()[name = tensor("input_207_strides_0"), val = tensor([1, 1])]; + tensor input_207_pad_0 = const()[name = tensor("input_207_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_207_dilations_0 = const()[name = tensor("input_207_dilations_0"), val = tensor([1, 1])]; + tensor input_207_groups_0 = const()[name = tensor("input_207_groups_0"), val = tensor(1)]; + tensor layers_7_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_7_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(389379008)))]; + tensor input_207_cast_fp16 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = layers_7_conv_pointwise_conv1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_split_num_splits_0 = const()[name = tensor("input_209_split_num_splits_0"), val = tensor(2)]; + tensor input_209_split_axis_0 = const()[name = tensor("input_209_split_axis_0"), val = tensor(1)]; + tensor input_209_split_cast_fp16_0, tensor input_209_split_cast_fp16_1 = split(axis = input_209_split_axis_0, num_splits = input_209_split_num_splits_0, x = input_207_cast_fp16)[name = tensor("input_209_split_cast_fp16")]; + tensor input_209_split_1_sigmoid_cast_fp16 = sigmoid(x = input_209_split_cast_fp16_1)[name = tensor("input_209_split_1_sigmoid_cast_fp16")]; + tensor input_209_cast_fp16 = mul(x = input_209_split_cast_fp16_0, y = input_209_split_1_sigmoid_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor input_211_pad_type_0 = const()[name = tensor("input_211_pad_type_0"), val = tensor("custom")]; + tensor input_211_pad_0 = const()[name = tensor("input_211_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_211_groups_0 = const()[name = tensor("input_211_groups_0"), val = tensor(1024)]; + tensor input_211_strides_0 = const()[name = tensor("input_211_strides_0"), val = tensor([1, 1])]; + tensor input_211_dilations_0 = const()[name = tensor("input_211_dilations_0"), val = tensor([1, 1])]; + tensor const_282_to_fp16 = const()[name = tensor("const_282_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393573376)))]; + tensor const_283_to_fp16 = const()[name = tensor("const_283_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393591872)))]; + tensor input_213_cast_fp16 = conv(bias = const_283_to_fp16, dilations = input_211_dilations_0, groups = input_211_groups_0, pad = input_211_pad_0, pad_type = input_211_pad_type_0, strides = input_211_strides_0, weight = const_282_to_fp16, x = input_209_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_cast_fp16 = silu(x = input_213_cast_fp16)[name = tensor("input_215_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 = const()[name = tensor("layers_7_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(393593984)))]; + 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, x = input_215_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_2337_to_fp16 = const()[name = tensor("op_2337_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_77_cast_fp16 = layer_norm(axes = out_77_axes_0, epsilon = var_2337_to_fp16, x = inputs_77_cast_fp16)[name = tensor("out_77_cast_fp16")]; + tensor input_217_gamma_0_to_fp16 = const()[name = tensor("input_217_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395691200)))]; + tensor input_217_beta_0_to_fp16 = const()[name = tensor("input_217_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395693312)))]; + tensor input_217_epsilon_0_to_fp16 = const()[name = tensor("input_217_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_217_cast_fp16 = batch_norm(beta = input_217_beta_0_to_fp16, epsilon = input_217_epsilon_0_to_fp16, gamma = input_217_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_77_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_pad_type_0 = const()[name = tensor("input_219_pad_type_0"), val = tensor("valid")]; + tensor input_219_strides_0 = const()[name = tensor("input_219_strides_0"), val = tensor([1, 1])]; + tensor input_219_pad_0 = const()[name = tensor("input_219_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_219_dilations_0 = const()[name = tensor("input_219_dilations_0"), val = tensor([1, 1])]; + tensor input_219_groups_0 = const()[name = tensor("input_219_groups_0"), val = tensor(1)]; + tensor layers_7_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_7_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395695424)))]; + tensor input_219_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_219_dilations_0, groups = input_219_groups_0, pad = input_219_pad_0, pad_type = input_219_pad_type_0, strides = input_219_strides_0, weight = layers_7_feed_forward2_fc1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_cast_fp16 = silu(x = input_219_cast_fp16)[name = tensor("input_221_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 var_2365_weight_0_to_fp16 = const()[name = tensor("op_2365_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(404084096)))]; + tensor var_2365_cast_fp16 = conv(bias = input_15_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 = var_2365_weight_0_to_fp16, x = input_221_cast_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor inputs_79_cast_fp16 = add(x = inputs_77_cast_fp16, y = var_2365_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_2375_to_fp16 = const()[name = tensor("op_2375_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_79_cast_fp16 = layer_norm(axes = out_79_axes_0, epsilon = var_2375_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(412472768)))]; + 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(412474880)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_79_cast_fp16)[name = tensor("inputs_81_cast_fp16")]; + tensor var_2389 = const()[name = tensor("op_2389"), val = tensor(3)]; + tensor out_81_axes_0 = const()[name = tensor("out_81_axes_0"), val = tensor([1])]; + tensor var_2420_to_fp16 = const()[name = tensor("op_2420_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_81_cast_fp16 = layer_norm(axes = out_81_axes_0, epsilon = var_2420_to_fp16, x = inputs_81_cast_fp16)[name = tensor("out_81_cast_fp16")]; + tensor input_223_gamma_0_to_fp16 = const()[name = tensor("input_223_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412476992)))]; + tensor input_223_beta_0_to_fp16 = const()[name = tensor("input_223_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412479104)))]; + tensor input_223_epsilon_0_to_fp16 = const()[name = tensor("input_223_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_223_cast_fp16 = batch_norm(beta = input_223_beta_0_to_fp16, epsilon = input_223_epsilon_0_to_fp16, gamma = input_223_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_81_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_pad_type_0 = const()[name = tensor("input_225_pad_type_0"), val = tensor("valid")]; + tensor input_225_strides_0 = const()[name = tensor("input_225_strides_0"), val = tensor([1, 1])]; + tensor input_225_pad_0 = const()[name = tensor("input_225_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_225_dilations_0 = const()[name = tensor("input_225_dilations_0"), val = tensor([1, 1])]; + tensor input_225_groups_0 = const()[name = tensor("input_225_groups_0"), val = tensor(1)]; + tensor layers_8_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_8_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412481216)))]; + tensor input_225_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_225_dilations_0, groups = input_225_groups_0, pad = input_225_pad_0, pad_type = input_225_pad_type_0, strides = input_225_strides_0, weight = layers_8_feed_forward1_fc1_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_cast_fp16 = silu(x = input_225_cast_fp16)[name = tensor("input_227_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 var_2448_weight_0_to_fp16 = const()[name = tensor("op_2448_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420869888)))]; + tensor var_2448_cast_fp16 = conv(bias = input_15_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 = var_2448_weight_0_to_fp16, x = input_227_cast_fp16)[name = tensor("op_2448_cast_fp16")]; + tensor inputs_83_cast_fp16 = add(x = inputs_81_cast_fp16, y = var_2448_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_2458_to_fp16 = const()[name = tensor("op_2458_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_83_cast_fp16 = layer_norm(axes = out_83_axes_0, epsilon = var_2458_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(429258560)))]; + 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(429260672)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429262784)))]; + tensor query_33_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(431360000)))]; + 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, 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 = const()[name = tensor("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(433457216)))]; + tensor value_17_cast_fp16 = conv(bias = input_15_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, x = obj_35_cast_fp16)[name = tensor("value_17_cast_fp16")]; + tensor var_2496_to_fp16 = const()[name = tensor("op_2496_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435554432)))]; + tensor query_35_cast_fp16 = add(x = query_33_cast_fp16, y = var_2496_to_fp16)[name = tensor("query_35_cast_fp16")]; + tensor var_2499_to_fp16 = const()[name = tensor("op_2499_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435556544)))]; + tensor q_with_bias_v_17_cast_fp16 = add(x = query_33_cast_fp16, y = var_2499_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 = const()[name = tensor("layers_8_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435558656)))]; + tensor p_17_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_17_cast_fp16")]; + tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([1, 8, 128, 188])]; + tensor var_2511_cast_fp16 = reshape(shape = var_2510, x = q_with_bias_v_17_cast_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512 = const()[name = tensor("op_2512"), val = tensor([1, 8, 128, -1])]; + tensor var_2513_cast_fp16 = reshape(shape = var_2512, x = p_17_cast_fp16)[name = tensor("op_2513_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_2511_cast_fp16, y = var_2513_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_2522 = const()[name = tensor("op_2522"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_69_cast_fp16 = reshape(shape = var_2522, x = matrix_bd_67_cast_fp16)[name = tensor("matrix_bd_69_cast_fp16")]; + tensor var_2526_begin_0 = const()[name = tensor("op_2526_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2526_end_0 = const()[name = tensor("op_2526_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_2526_end_mask_0 = const()[name = tensor("op_2526_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2526_cast_fp16 = slice_by_index(begin = var_2526_begin_0, end = var_2526_end_0, end_mask = var_2526_end_mask_0, x = matrix_bd_69_cast_fp16)[name = tensor("op_2526_cast_fp16")]; + tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_71_cast_fp16 = reshape(shape = var_2527, x = var_2526_cast_fp16)[name = tensor("matrix_bd_71_cast_fp16")]; + tensor var_2532_begin_0 = const()[name = tensor("op_2532_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2532_end_0 = const()[name = tensor("op_2532_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_2532_end_mask_0 = const()[name = tensor("op_2532_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_2532_cast_fp16 = slice_by_index(begin = var_2532_begin_0, end = var_2532_end_0, end_mask = var_2532_end_mask_0, x = matrix_bd_71_cast_fp16)[name = tensor("op_2532_cast_fp16")]; + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_17_cast_fp16 = mul(x = var_2532_cast_fp16, y = var_2533_to_fp16)[name = tensor("qk_mask_17_cast_fp16")]; + tensor var_2537 = const()[name = tensor("op_2537"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_17_cast_fp16 = reshape(shape = var_2537, x = query_35_cast_fp16)[name = tensor("mh_q_17_cast_fp16")]; + tensor var_2539_to_fp16 = const()[name = tensor("op_2539_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_2540_cast_fp16 = mul(x = mh_q_17_cast_fp16, y = var_2539_to_fp16)[name = tensor("op_2540_cast_fp16")]; + tensor var_2543 = const()[name = tensor("op_2543"), val = tensor([1, 8, 128, 188])]; + tensor var_2544_cast_fp16 = reshape(shape = var_2543, x = key_17_cast_fp16)[name = tensor("op_2544_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_2540_cast_fp16, y = var_2544_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_2548_cast_fp16 = softmax(axis = var_2389, x = mh_w_35_cast_fp16)[name = tensor("op_2548_cast_fp16")]; + tensor var_2549 = const()[name = tensor("op_2549"), val = tensor([1, 8, 128, 188])]; + tensor var_2550_cast_fp16 = reshape(shape = var_2549, x = value_17_cast_fp16)[name = tensor("op_2550_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_2550_cast_fp16, y = var_2548_cast_fp16)[name = tensor("attn_17_cast_fp16")]; + tensor var_2553 = const()[name = tensor("op_2553"), val = tensor([1, 1024, 1, 188])]; + tensor input_229_cast_fp16 = reshape(shape = var_2553, x = attn_17_cast_fp16)[name = tensor("input_229_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 = const()[name = tensor("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437655872)))]; + tensor obj_37_cast_fp16 = conv(bias = input_15_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, x = input_229_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_2571_to_fp16 = const()[name = tensor("op_2571_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_85_cast_fp16 = layer_norm(axes = out_85_axes_0, epsilon = var_2571_to_fp16, x = inputs_85_cast_fp16)[name = tensor("out_85_cast_fp16")]; + tensor input_231_gamma_0_to_fp16 = const()[name = tensor("input_231_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439753088)))]; + tensor input_231_beta_0_to_fp16 = const()[name = tensor("input_231_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439755200)))]; + tensor input_231_epsilon_0_to_fp16 = const()[name = tensor("input_231_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = batch_norm(beta = input_231_beta_0_to_fp16, epsilon = input_231_epsilon_0_to_fp16, gamma = input_231_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_85_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_pad_type_0 = const()[name = tensor("input_233_pad_type_0"), val = tensor("valid")]; + tensor input_233_strides_0 = const()[name = tensor("input_233_strides_0"), val = tensor([1, 1])]; + tensor input_233_pad_0 = const()[name = tensor("input_233_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_233_dilations_0 = const()[name = tensor("input_233_dilations_0"), val = tensor([1, 1])]; + tensor input_233_groups_0 = const()[name = tensor("input_233_groups_0"), val = tensor(1)]; + tensor layers_8_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_8_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(439757312)))]; + tensor input_233_cast_fp16 = conv(dilations = input_233_dilations_0, groups = input_233_groups_0, pad = input_233_pad_0, pad_type = input_233_pad_type_0, strides = input_233_strides_0, weight = layers_8_conv_pointwise_conv1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor input_235_split_num_splits_0 = const()[name = tensor("input_235_split_num_splits_0"), val = tensor(2)]; + tensor input_235_split_axis_0 = const()[name = tensor("input_235_split_axis_0"), val = tensor(1)]; + tensor input_235_split_cast_fp16_0, tensor input_235_split_cast_fp16_1 = split(axis = input_235_split_axis_0, num_splits = input_235_split_num_splits_0, x = input_233_cast_fp16)[name = tensor("input_235_split_cast_fp16")]; + tensor input_235_split_1_sigmoid_cast_fp16 = sigmoid(x = input_235_split_cast_fp16_1)[name = tensor("input_235_split_1_sigmoid_cast_fp16")]; + tensor input_235_cast_fp16 = mul(x = input_235_split_cast_fp16_0, y = input_235_split_1_sigmoid_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_pad_type_0 = const()[name = tensor("input_237_pad_type_0"), val = tensor("custom")]; + tensor input_237_pad_0 = const()[name = tensor("input_237_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_237_groups_0 = const()[name = tensor("input_237_groups_0"), val = tensor(1024)]; + tensor input_237_strides_0 = const()[name = tensor("input_237_strides_0"), val = tensor([1, 1])]; + tensor input_237_dilations_0 = const()[name = tensor("input_237_dilations_0"), val = tensor([1, 1])]; + tensor const_284_to_fp16 = const()[name = tensor("const_284_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443951680)))]; + tensor const_285_to_fp16 = const()[name = tensor("const_285_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443970176)))]; + tensor input_239_cast_fp16 = conv(bias = const_285_to_fp16, dilations = input_237_dilations_0, groups = input_237_groups_0, pad = input_237_pad_0, pad_type = input_237_pad_type_0, strides = input_237_strides_0, weight = const_284_to_fp16, x = input_235_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = input_239_cast_fp16)[name = tensor("input_241_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 = const()[name = tensor("layers_8_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(443972288)))]; + 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, x = input_241_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_2619_to_fp16 = const()[name = tensor("op_2619_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_87_cast_fp16 = layer_norm(axes = out_87_axes_0, epsilon = var_2619_to_fp16, x = inputs_87_cast_fp16)[name = tensor("out_87_cast_fp16")]; + tensor input_243_gamma_0_to_fp16 = const()[name = tensor("input_243_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446069504)))]; + tensor input_243_beta_0_to_fp16 = const()[name = tensor("input_243_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446071616)))]; + tensor input_243_epsilon_0_to_fp16 = const()[name = tensor("input_243_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = batch_norm(beta = input_243_beta_0_to_fp16, epsilon = input_243_epsilon_0_to_fp16, gamma = input_243_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_87_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_pad_type_0 = const()[name = tensor("input_245_pad_type_0"), val = tensor("valid")]; + tensor input_245_strides_0 = const()[name = tensor("input_245_strides_0"), val = tensor([1, 1])]; + tensor input_245_pad_0 = const()[name = tensor("input_245_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_245_dilations_0 = const()[name = tensor("input_245_dilations_0"), val = tensor([1, 1])]; + tensor input_245_groups_0 = const()[name = tensor("input_245_groups_0"), val = tensor(1)]; + tensor layers_8_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_8_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(446073728)))]; + tensor input_245_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_245_dilations_0, groups = input_245_groups_0, pad = input_245_pad_0, pad_type = input_245_pad_type_0, strides = input_245_strides_0, weight = layers_8_feed_forward2_fc1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor input_247_cast_fp16 = silu(x = input_245_cast_fp16)[name = tensor("input_247_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 var_2647_weight_0_to_fp16 = const()[name = tensor("op_2647_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454462400)))]; + tensor var_2647_cast_fp16 = conv(bias = input_15_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 = var_2647_weight_0_to_fp16, x = input_247_cast_fp16)[name = tensor("op_2647_cast_fp16")]; + tensor inputs_89_cast_fp16 = add(x = inputs_87_cast_fp16, y = var_2647_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_2657_to_fp16 = const()[name = tensor("op_2657_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_89_cast_fp16 = layer_norm(axes = out_89_axes_0, epsilon = var_2657_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(462851072)))]; + 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(462853184)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_89_cast_fp16)[name = tensor("inputs_91_cast_fp16")]; + tensor var_2671 = const()[name = tensor("op_2671"), val = tensor(3)]; + tensor out_91_axes_0 = const()[name = tensor("out_91_axes_0"), val = tensor([1])]; + tensor var_2702_to_fp16 = const()[name = tensor("op_2702_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_91_cast_fp16 = layer_norm(axes = out_91_axes_0, epsilon = var_2702_to_fp16, x = inputs_91_cast_fp16)[name = tensor("out_91_cast_fp16")]; + tensor input_249_gamma_0_to_fp16 = const()[name = tensor("input_249_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462855296)))]; + tensor input_249_beta_0_to_fp16 = const()[name = tensor("input_249_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462857408)))]; + tensor input_249_epsilon_0_to_fp16 = const()[name = tensor("input_249_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = batch_norm(beta = input_249_beta_0_to_fp16, epsilon = input_249_epsilon_0_to_fp16, gamma = input_249_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_91_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_pad_type_0 = const()[name = tensor("input_251_pad_type_0"), val = tensor("valid")]; + tensor input_251_strides_0 = const()[name = tensor("input_251_strides_0"), val = tensor([1, 1])]; + tensor input_251_pad_0 = const()[name = tensor("input_251_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_251_dilations_0 = const()[name = tensor("input_251_dilations_0"), val = tensor([1, 1])]; + tensor input_251_groups_0 = const()[name = tensor("input_251_groups_0"), val = tensor(1)]; + tensor layers_9_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_9_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462859520)))]; + tensor input_251_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_251_dilations_0, groups = input_251_groups_0, pad = input_251_pad_0, pad_type = input_251_pad_type_0, strides = input_251_strides_0, weight = layers_9_feed_forward1_fc1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor input_253_cast_fp16 = silu(x = input_251_cast_fp16)[name = tensor("input_253_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 var_2730_weight_0_to_fp16 = const()[name = tensor("op_2730_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471248192)))]; + tensor var_2730_cast_fp16 = conv(bias = input_15_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 = var_2730_weight_0_to_fp16, x = input_253_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor inputs_93_cast_fp16 = add(x = inputs_91_cast_fp16, y = var_2730_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_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_93_cast_fp16 = layer_norm(axes = out_93_axes_0, epsilon = var_2740_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(479636864)))]; + 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(479638976)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479641088)))]; + tensor query_37_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(481738304)))]; + 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, 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 = const()[name = tensor("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(483835520)))]; + tensor value_19_cast_fp16 = conv(bias = input_15_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, x = obj_39_cast_fp16)[name = tensor("value_19_cast_fp16")]; + tensor var_2778_to_fp16 = const()[name = tensor("op_2778_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485932736)))]; + tensor query_39_cast_fp16 = add(x = query_37_cast_fp16, y = var_2778_to_fp16)[name = tensor("query_39_cast_fp16")]; + tensor var_2781_to_fp16 = const()[name = tensor("op_2781_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485934848)))]; + tensor q_with_bias_v_19_cast_fp16 = add(x = query_37_cast_fp16, y = var_2781_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 = const()[name = tensor("layers_9_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485936960)))]; + tensor p_19_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_19_cast_fp16")]; + tensor var_2792 = const()[name = tensor("op_2792"), val = tensor([1, 8, 128, 188])]; + tensor var_2793_cast_fp16 = reshape(shape = var_2792, x = q_with_bias_v_19_cast_fp16)[name = tensor("op_2793_cast_fp16")]; + tensor var_2794 = const()[name = tensor("op_2794"), val = tensor([1, 8, 128, -1])]; + tensor var_2795_cast_fp16 = reshape(shape = var_2794, x = p_19_cast_fp16)[name = tensor("op_2795_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_2793_cast_fp16, y = var_2795_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_2804 = const()[name = tensor("op_2804"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_77_cast_fp16 = reshape(shape = var_2804, x = matrix_bd_75_cast_fp16)[name = tensor("matrix_bd_77_cast_fp16")]; + tensor var_2808_begin_0 = const()[name = tensor("op_2808_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_2808_end_0 = const()[name = tensor("op_2808_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_2808_end_mask_0 = const()[name = tensor("op_2808_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_2808_cast_fp16 = slice_by_index(begin = var_2808_begin_0, end = var_2808_end_0, end_mask = var_2808_end_mask_0, x = matrix_bd_77_cast_fp16)[name = tensor("op_2808_cast_fp16")]; + tensor var_2809 = const()[name = tensor("op_2809"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_79_cast_fp16 = reshape(shape = var_2809, x = var_2808_cast_fp16)[name = tensor("matrix_bd_79_cast_fp16")]; + tensor var_2814_begin_0 = const()[name = tensor("op_2814_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_2814_end_0 = const()[name = tensor("op_2814_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_2814_end_mask_0 = const()[name = tensor("op_2814_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_2814_cast_fp16 = slice_by_index(begin = var_2814_begin_0, end = var_2814_end_0, end_mask = var_2814_end_mask_0, x = matrix_bd_79_cast_fp16)[name = tensor("op_2814_cast_fp16")]; + tensor var_2815_to_fp16 = const()[name = tensor("op_2815_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_19_cast_fp16 = mul(x = var_2814_cast_fp16, y = var_2815_to_fp16)[name = tensor("qk_mask_19_cast_fp16")]; + tensor var_2819 = const()[name = tensor("op_2819"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_19_cast_fp16 = reshape(shape = var_2819, x = query_39_cast_fp16)[name = tensor("mh_q_19_cast_fp16")]; + tensor var_2821_to_fp16 = const()[name = tensor("op_2821_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_2822_cast_fp16 = mul(x = mh_q_19_cast_fp16, y = var_2821_to_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor var_2825 = const()[name = tensor("op_2825"), val = tensor([1, 8, 128, 188])]; + tensor var_2826_cast_fp16 = reshape(shape = var_2825, x = key_19_cast_fp16)[name = tensor("op_2826_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_2822_cast_fp16, y = var_2826_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_2830_cast_fp16 = softmax(axis = var_2671, x = mh_w_39_cast_fp16)[name = tensor("op_2830_cast_fp16")]; + tensor var_2831 = const()[name = tensor("op_2831"), val = tensor([1, 8, 128, 188])]; + tensor var_2832_cast_fp16 = reshape(shape = var_2831, x = value_19_cast_fp16)[name = tensor("op_2832_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_2832_cast_fp16, y = var_2830_cast_fp16)[name = tensor("attn_19_cast_fp16")]; + tensor var_2835 = const()[name = tensor("op_2835"), val = tensor([1, 1024, 1, 188])]; + tensor input_255_cast_fp16 = reshape(shape = var_2835, x = attn_19_cast_fp16)[name = tensor("input_255_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 = const()[name = tensor("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(488034176)))]; + tensor obj_41_cast_fp16 = conv(bias = input_15_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, x = input_255_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_2853_to_fp16 = const()[name = tensor("op_2853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_95_cast_fp16 = layer_norm(axes = out_95_axes_0, epsilon = var_2853_to_fp16, x = inputs_95_cast_fp16)[name = tensor("out_95_cast_fp16")]; + tensor input_257_gamma_0_to_fp16 = const()[name = tensor("input_257_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490131392)))]; + tensor input_257_beta_0_to_fp16 = const()[name = tensor("input_257_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490133504)))]; + tensor input_257_epsilon_0_to_fp16 = const()[name = tensor("input_257_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = batch_norm(beta = input_257_beta_0_to_fp16, epsilon = input_257_epsilon_0_to_fp16, gamma = input_257_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_95_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor input_259_pad_type_0 = const()[name = tensor("input_259_pad_type_0"), val = tensor("valid")]; + tensor input_259_strides_0 = const()[name = tensor("input_259_strides_0"), val = tensor([1, 1])]; + tensor input_259_pad_0 = const()[name = tensor("input_259_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_259_dilations_0 = const()[name = tensor("input_259_dilations_0"), val = tensor([1, 1])]; + tensor input_259_groups_0 = const()[name = tensor("input_259_groups_0"), val = tensor(1)]; + tensor layers_9_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_9_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(490135616)))]; + tensor input_259_cast_fp16 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = layers_9_conv_pointwise_conv1_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_split_num_splits_0 = const()[name = tensor("input_261_split_num_splits_0"), val = tensor(2)]; + tensor input_261_split_axis_0 = const()[name = tensor("input_261_split_axis_0"), val = tensor(1)]; + tensor input_261_split_cast_fp16_0, tensor input_261_split_cast_fp16_1 = split(axis = input_261_split_axis_0, num_splits = input_261_split_num_splits_0, x = input_259_cast_fp16)[name = tensor("input_261_split_cast_fp16")]; + tensor input_261_split_1_sigmoid_cast_fp16 = sigmoid(x = input_261_split_cast_fp16_1)[name = tensor("input_261_split_1_sigmoid_cast_fp16")]; + tensor input_261_cast_fp16 = mul(x = input_261_split_cast_fp16_0, y = input_261_split_1_sigmoid_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_pad_type_0 = const()[name = tensor("input_263_pad_type_0"), val = tensor("custom")]; + tensor input_263_pad_0 = const()[name = tensor("input_263_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_263_groups_0 = const()[name = tensor("input_263_groups_0"), val = tensor(1024)]; + tensor input_263_strides_0 = const()[name = tensor("input_263_strides_0"), val = tensor([1, 1])]; + tensor input_263_dilations_0 = const()[name = tensor("input_263_dilations_0"), val = tensor([1, 1])]; + tensor const_286_to_fp16 = const()[name = tensor("const_286_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494329984)))]; + tensor const_287_to_fp16 = const()[name = tensor("const_287_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494348480)))]; + tensor input_265_cast_fp16 = conv(bias = const_287_to_fp16, dilations = input_263_dilations_0, groups = input_263_groups_0, pad = input_263_pad_0, pad_type = input_263_pad_type_0, strides = input_263_strides_0, weight = const_286_to_fp16, x = input_261_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_cast_fp16 = silu(x = input_265_cast_fp16)[name = tensor("input_267_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 = const()[name = tensor("layers_9_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(494350592)))]; + 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, x = input_267_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_2901_to_fp16 = const()[name = tensor("op_2901_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_97_cast_fp16 = layer_norm(axes = out_97_axes_0, epsilon = var_2901_to_fp16, x = inputs_97_cast_fp16)[name = tensor("out_97_cast_fp16")]; + tensor input_269_gamma_0_to_fp16 = const()[name = tensor("input_269_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496447808)))]; + tensor input_269_beta_0_to_fp16 = const()[name = tensor("input_269_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496449920)))]; + tensor input_269_epsilon_0_to_fp16 = const()[name = tensor("input_269_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = batch_norm(beta = input_269_beta_0_to_fp16, epsilon = input_269_epsilon_0_to_fp16, gamma = input_269_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_97_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor input_271_pad_type_0 = const()[name = tensor("input_271_pad_type_0"), val = tensor("valid")]; + tensor input_271_strides_0 = const()[name = tensor("input_271_strides_0"), val = tensor([1, 1])]; + tensor input_271_pad_0 = const()[name = tensor("input_271_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_271_dilations_0 = const()[name = tensor("input_271_dilations_0"), val = tensor([1, 1])]; + tensor input_271_groups_0 = const()[name = tensor("input_271_groups_0"), val = tensor(1)]; + tensor layers_9_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_9_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496452032)))]; + tensor input_271_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_271_dilations_0, groups = input_271_groups_0, pad = input_271_pad_0, pad_type = input_271_pad_type_0, strides = input_271_strides_0, weight = layers_9_feed_forward2_fc1_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_cast_fp16 = silu(x = input_271_cast_fp16)[name = tensor("input_273_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 var_2929_weight_0_to_fp16 = const()[name = tensor("op_2929_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504840704)))]; + tensor var_2929_cast_fp16 = conv(bias = input_15_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 = var_2929_weight_0_to_fp16, x = input_273_cast_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor inputs_99_cast_fp16 = add(x = inputs_97_cast_fp16, y = var_2929_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_2939_to_fp16 = const()[name = tensor("op_2939_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_99_cast_fp16 = layer_norm(axes = out_99_axes_0, epsilon = var_2939_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(513229376)))]; + 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(513231488)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_99_cast_fp16)[name = tensor("inputs_101_cast_fp16")]; + tensor var_2953 = const()[name = tensor("op_2953"), val = tensor(3)]; + tensor out_101_axes_0 = const()[name = tensor("out_101_axes_0"), val = tensor([1])]; + tensor var_2984_to_fp16 = const()[name = tensor("op_2984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_101_cast_fp16 = layer_norm(axes = out_101_axes_0, epsilon = var_2984_to_fp16, x = inputs_101_cast_fp16)[name = tensor("out_101_cast_fp16")]; + tensor input_275_gamma_0_to_fp16 = const()[name = tensor("input_275_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513233600)))]; + tensor input_275_beta_0_to_fp16 = const()[name = tensor("input_275_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513235712)))]; + tensor input_275_epsilon_0_to_fp16 = const()[name = tensor("input_275_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = batch_norm(beta = input_275_beta_0_to_fp16, epsilon = input_275_epsilon_0_to_fp16, gamma = input_275_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_101_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor input_277_pad_type_0 = const()[name = tensor("input_277_pad_type_0"), val = tensor("valid")]; + tensor input_277_strides_0 = const()[name = tensor("input_277_strides_0"), val = tensor([1, 1])]; + tensor input_277_pad_0 = const()[name = tensor("input_277_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_277_dilations_0 = const()[name = tensor("input_277_dilations_0"), val = tensor([1, 1])]; + tensor input_277_groups_0 = const()[name = tensor("input_277_groups_0"), val = tensor(1)]; + tensor layers_10_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_10_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513237824)))]; + tensor input_277_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_277_dilations_0, groups = input_277_groups_0, pad = input_277_pad_0, pad_type = input_277_pad_type_0, strides = input_277_strides_0, weight = layers_10_feed_forward1_fc1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_cast_fp16 = silu(x = input_277_cast_fp16)[name = tensor("input_279_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 var_3012_weight_0_to_fp16 = const()[name = tensor("op_3012_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521626496)))]; + tensor var_3012_cast_fp16 = conv(bias = input_15_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 = var_3012_weight_0_to_fp16, x = input_279_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor inputs_103_cast_fp16 = add(x = inputs_101_cast_fp16, y = var_3012_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_3022_to_fp16 = const()[name = tensor("op_3022_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_103_cast_fp16 = layer_norm(axes = out_103_axes_0, epsilon = var_3022_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(530015168)))]; + 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(530017280)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(530019392)))]; + tensor query_41_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(532116608)))]; + 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, 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 = const()[name = tensor("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(534213824)))]; + tensor value_21_cast_fp16 = conv(bias = input_15_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, x = obj_43_cast_fp16)[name = tensor("value_21_cast_fp16")]; + tensor var_3060_to_fp16 = const()[name = tensor("op_3060_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536311040)))]; + tensor query_43_cast_fp16 = add(x = query_41_cast_fp16, y = var_3060_to_fp16)[name = tensor("query_43_cast_fp16")]; + tensor var_3063_to_fp16 = const()[name = tensor("op_3063_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536313152)))]; + tensor q_with_bias_v_21_cast_fp16 = add(x = query_41_cast_fp16, y = var_3063_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 = const()[name = tensor("layers_10_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536315264)))]; + tensor p_21_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_21_cast_fp16")]; + tensor var_3074 = const()[name = tensor("op_3074"), val = tensor([1, 8, 128, 188])]; + tensor var_3075_cast_fp16 = reshape(shape = var_3074, x = q_with_bias_v_21_cast_fp16)[name = tensor("op_3075_cast_fp16")]; + tensor var_3076 = const()[name = tensor("op_3076"), val = tensor([1, 8, 128, -1])]; + tensor var_3077_cast_fp16 = reshape(shape = var_3076, x = p_21_cast_fp16)[name = tensor("op_3077_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_3075_cast_fp16, y = var_3077_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_3086 = const()[name = tensor("op_3086"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_85_cast_fp16 = reshape(shape = var_3086, x = matrix_bd_83_cast_fp16)[name = tensor("matrix_bd_85_cast_fp16")]; + tensor var_3090_begin_0 = const()[name = tensor("op_3090_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3090_end_0 = const()[name = tensor("op_3090_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3090_end_mask_0 = const()[name = tensor("op_3090_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3090_cast_fp16 = slice_by_index(begin = var_3090_begin_0, end = var_3090_end_0, end_mask = var_3090_end_mask_0, x = matrix_bd_85_cast_fp16)[name = tensor("op_3090_cast_fp16")]; + tensor var_3091 = const()[name = tensor("op_3091"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_87_cast_fp16 = reshape(shape = var_3091, x = var_3090_cast_fp16)[name = tensor("matrix_bd_87_cast_fp16")]; + tensor var_3096_begin_0 = const()[name = tensor("op_3096_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3096_end_0 = const()[name = tensor("op_3096_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3096_end_mask_0 = const()[name = tensor("op_3096_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3096_cast_fp16 = slice_by_index(begin = var_3096_begin_0, end = var_3096_end_0, end_mask = var_3096_end_mask_0, x = matrix_bd_87_cast_fp16)[name = tensor("op_3096_cast_fp16")]; + tensor var_3097_to_fp16 = const()[name = tensor("op_3097_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_21_cast_fp16 = mul(x = var_3096_cast_fp16, y = var_3097_to_fp16)[name = tensor("qk_mask_21_cast_fp16")]; + tensor var_3101 = const()[name = tensor("op_3101"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_21_cast_fp16 = reshape(shape = var_3101, x = query_43_cast_fp16)[name = tensor("mh_q_21_cast_fp16")]; + tensor var_3103_to_fp16 = const()[name = tensor("op_3103_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3104_cast_fp16 = mul(x = mh_q_21_cast_fp16, y = var_3103_to_fp16)[name = tensor("op_3104_cast_fp16")]; + tensor var_3107 = const()[name = tensor("op_3107"), val = tensor([1, 8, 128, 188])]; + tensor var_3108_cast_fp16 = reshape(shape = var_3107, x = key_21_cast_fp16)[name = tensor("op_3108_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_3104_cast_fp16, y = var_3108_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_3112_cast_fp16 = softmax(axis = var_2953, x = mh_w_43_cast_fp16)[name = tensor("op_3112_cast_fp16")]; + tensor var_3113 = const()[name = tensor("op_3113"), val = tensor([1, 8, 128, 188])]; + tensor var_3114_cast_fp16 = reshape(shape = var_3113, x = value_21_cast_fp16)[name = tensor("op_3114_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_3114_cast_fp16, y = var_3112_cast_fp16)[name = tensor("attn_21_cast_fp16")]; + tensor var_3117 = const()[name = tensor("op_3117"), val = tensor([1, 1024, 1, 188])]; + tensor input_281_cast_fp16 = reshape(shape = var_3117, x = attn_21_cast_fp16)[name = tensor("input_281_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 = const()[name = tensor("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538412480)))]; + tensor obj_45_cast_fp16 = conv(bias = input_15_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, x = input_281_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_3135_to_fp16 = const()[name = tensor("op_3135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_105_cast_fp16 = layer_norm(axes = out_105_axes_0, epsilon = var_3135_to_fp16, x = inputs_105_cast_fp16)[name = tensor("out_105_cast_fp16")]; + tensor input_283_gamma_0_to_fp16 = const()[name = tensor("input_283_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540509696)))]; + tensor input_283_beta_0_to_fp16 = const()[name = tensor("input_283_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540511808)))]; + tensor input_283_epsilon_0_to_fp16 = const()[name = tensor("input_283_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_283_cast_fp16 = batch_norm(beta = input_283_beta_0_to_fp16, epsilon = input_283_epsilon_0_to_fp16, gamma = input_283_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_105_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor input_285_pad_type_0 = const()[name = tensor("input_285_pad_type_0"), val = tensor("valid")]; + tensor input_285_strides_0 = const()[name = tensor("input_285_strides_0"), val = tensor([1, 1])]; + tensor input_285_pad_0 = const()[name = tensor("input_285_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_285_dilations_0 = const()[name = tensor("input_285_dilations_0"), val = tensor([1, 1])]; + tensor input_285_groups_0 = const()[name = tensor("input_285_groups_0"), val = tensor(1)]; + tensor layers_10_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_10_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(540513920)))]; + tensor input_285_cast_fp16 = conv(dilations = input_285_dilations_0, groups = input_285_groups_0, pad = input_285_pad_0, pad_type = input_285_pad_type_0, strides = input_285_strides_0, weight = layers_10_conv_pointwise_conv1_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_split_num_splits_0 = const()[name = tensor("input_287_split_num_splits_0"), val = tensor(2)]; + tensor input_287_split_axis_0 = const()[name = tensor("input_287_split_axis_0"), val = tensor(1)]; + tensor input_287_split_cast_fp16_0, tensor input_287_split_cast_fp16_1 = split(axis = input_287_split_axis_0, num_splits = input_287_split_num_splits_0, x = input_285_cast_fp16)[name = tensor("input_287_split_cast_fp16")]; + tensor input_287_split_1_sigmoid_cast_fp16 = sigmoid(x = input_287_split_cast_fp16_1)[name = tensor("input_287_split_1_sigmoid_cast_fp16")]; + tensor input_287_cast_fp16 = mul(x = input_287_split_cast_fp16_0, y = input_287_split_1_sigmoid_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor input_289_pad_type_0 = const()[name = tensor("input_289_pad_type_0"), val = tensor("custom")]; + tensor input_289_pad_0 = const()[name = tensor("input_289_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_289_groups_0 = const()[name = tensor("input_289_groups_0"), val = tensor(1024)]; + tensor input_289_strides_0 = const()[name = tensor("input_289_strides_0"), val = tensor([1, 1])]; + tensor input_289_dilations_0 = const()[name = tensor("input_289_dilations_0"), val = tensor([1, 1])]; + tensor const_288_to_fp16 = const()[name = tensor("const_288_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(544708288)))]; + tensor const_289_to_fp16 = const()[name = tensor("const_289_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(544726784)))]; + tensor input_291_cast_fp16 = conv(bias = const_289_to_fp16, dilations = input_289_dilations_0, groups = input_289_groups_0, pad = input_289_pad_0, pad_type = input_289_pad_type_0, strides = input_289_strides_0, weight = const_288_to_fp16, x = input_287_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = input_291_cast_fp16)[name = tensor("input_293_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 = const()[name = tensor("layers_10_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(544728896)))]; + 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, x = input_293_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_3183_to_fp16 = const()[name = tensor("op_3183_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_107_cast_fp16 = layer_norm(axes = out_107_axes_0, epsilon = var_3183_to_fp16, x = inputs_107_cast_fp16)[name = tensor("out_107_cast_fp16")]; + tensor input_295_gamma_0_to_fp16 = const()[name = tensor("input_295_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546826112)))]; + tensor input_295_beta_0_to_fp16 = const()[name = tensor("input_295_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546828224)))]; + tensor input_295_epsilon_0_to_fp16 = const()[name = tensor("input_295_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_295_cast_fp16 = batch_norm(beta = input_295_beta_0_to_fp16, epsilon = input_295_epsilon_0_to_fp16, gamma = input_295_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_107_cast_fp16)[name = tensor("input_295_cast_fp16")]; + tensor input_297_pad_type_0 = const()[name = tensor("input_297_pad_type_0"), val = tensor("valid")]; + tensor input_297_strides_0 = const()[name = tensor("input_297_strides_0"), val = tensor([1, 1])]; + tensor input_297_pad_0 = const()[name = tensor("input_297_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_297_dilations_0 = const()[name = tensor("input_297_dilations_0"), val = tensor([1, 1])]; + tensor input_297_groups_0 = const()[name = tensor("input_297_groups_0"), val = tensor(1)]; + tensor layers_10_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_10_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546830336)))]; + tensor input_297_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_297_dilations_0, groups = input_297_groups_0, pad = input_297_pad_0, pad_type = input_297_pad_type_0, strides = input_297_strides_0, weight = layers_10_feed_forward2_fc1_weight_to_fp16, x = input_295_cast_fp16)[name = tensor("input_297_cast_fp16")]; + tensor input_299_cast_fp16 = silu(x = input_297_cast_fp16)[name = tensor("input_299_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 var_3211_weight_0_to_fp16 = const()[name = tensor("op_3211_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(555219008)))]; + tensor var_3211_cast_fp16 = conv(bias = input_15_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 = var_3211_weight_0_to_fp16, x = input_299_cast_fp16)[name = tensor("op_3211_cast_fp16")]; + tensor inputs_109_cast_fp16 = add(x = inputs_107_cast_fp16, y = var_3211_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_3221_to_fp16 = const()[name = tensor("op_3221_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_109_cast_fp16 = layer_norm(axes = out_109_axes_0, epsilon = var_3221_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(563607680)))]; + 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(563609792)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_109_cast_fp16)[name = tensor("inputs_111_cast_fp16")]; + tensor var_3235 = const()[name = tensor("op_3235"), val = tensor(3)]; + tensor out_111_axes_0 = const()[name = tensor("out_111_axes_0"), val = tensor([1])]; + tensor var_3266_to_fp16 = const()[name = tensor("op_3266_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_111_cast_fp16 = layer_norm(axes = out_111_axes_0, epsilon = var_3266_to_fp16, x = inputs_111_cast_fp16)[name = tensor("out_111_cast_fp16")]; + tensor input_301_gamma_0_to_fp16 = const()[name = tensor("input_301_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563611904)))]; + tensor input_301_beta_0_to_fp16 = const()[name = tensor("input_301_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563614016)))]; + tensor input_301_epsilon_0_to_fp16 = const()[name = tensor("input_301_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_301_cast_fp16 = batch_norm(beta = input_301_beta_0_to_fp16, epsilon = input_301_epsilon_0_to_fp16, gamma = input_301_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_111_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor input_303_pad_type_0 = const()[name = tensor("input_303_pad_type_0"), val = tensor("valid")]; + tensor input_303_strides_0 = const()[name = tensor("input_303_strides_0"), val = tensor([1, 1])]; + tensor input_303_pad_0 = const()[name = tensor("input_303_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_303_dilations_0 = const()[name = tensor("input_303_dilations_0"), val = tensor([1, 1])]; + tensor input_303_groups_0 = const()[name = tensor("input_303_groups_0"), val = tensor(1)]; + tensor layers_11_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_11_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563616128)))]; + tensor input_303_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_303_dilations_0, groups = input_303_groups_0, pad = input_303_pad_0, pad_type = input_303_pad_type_0, strides = input_303_strides_0, weight = layers_11_feed_forward1_fc1_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("input_303_cast_fp16")]; + tensor input_305_cast_fp16 = silu(x = input_303_cast_fp16)[name = tensor("input_305_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 var_3294_weight_0_to_fp16 = const()[name = tensor("op_3294_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(572004800)))]; + tensor var_3294_cast_fp16 = conv(bias = input_15_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 = var_3294_weight_0_to_fp16, x = input_305_cast_fp16)[name = tensor("op_3294_cast_fp16")]; + tensor inputs_113_cast_fp16 = add(x = inputs_111_cast_fp16, y = var_3294_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_3304_to_fp16 = const()[name = tensor("op_3304_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_113_cast_fp16 = layer_norm(axes = out_113_axes_0, epsilon = var_3304_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(580393472)))]; + 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(580395584)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580397696)))]; + tensor query_45_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(582494912)))]; + 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, 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 = const()[name = tensor("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(584592128)))]; + tensor value_23_cast_fp16 = conv(bias = input_15_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, x = obj_47_cast_fp16)[name = tensor("value_23_cast_fp16")]; + tensor var_3342_to_fp16 = const()[name = tensor("op_3342_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586689344)))]; + tensor query_47_cast_fp16 = add(x = query_45_cast_fp16, y = var_3342_to_fp16)[name = tensor("query_47_cast_fp16")]; + tensor var_3345_to_fp16 = const()[name = tensor("op_3345_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586691456)))]; + tensor q_with_bias_v_23_cast_fp16 = add(x = query_45_cast_fp16, y = var_3345_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 = const()[name = tensor("layers_11_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586693568)))]; + tensor p_23_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_23_cast_fp16")]; + tensor var_3356 = const()[name = tensor("op_3356"), val = tensor([1, 8, 128, 188])]; + tensor var_3357_cast_fp16 = reshape(shape = var_3356, x = q_with_bias_v_23_cast_fp16)[name = tensor("op_3357_cast_fp16")]; + tensor var_3358 = const()[name = tensor("op_3358"), val = tensor([1, 8, 128, -1])]; + tensor var_3359_cast_fp16 = reshape(shape = var_3358, x = p_23_cast_fp16)[name = tensor("op_3359_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_3357_cast_fp16, y = var_3359_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_3368 = const()[name = tensor("op_3368"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_93_cast_fp16 = reshape(shape = var_3368, x = matrix_bd_91_cast_fp16)[name = tensor("matrix_bd_93_cast_fp16")]; + tensor var_3372_begin_0 = const()[name = tensor("op_3372_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3372_end_0 = const()[name = tensor("op_3372_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3372_end_mask_0 = const()[name = tensor("op_3372_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3372_cast_fp16 = slice_by_index(begin = var_3372_begin_0, end = var_3372_end_0, end_mask = var_3372_end_mask_0, x = matrix_bd_93_cast_fp16)[name = tensor("op_3372_cast_fp16")]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_95_cast_fp16 = reshape(shape = var_3373, x = var_3372_cast_fp16)[name = tensor("matrix_bd_95_cast_fp16")]; + tensor var_3378_begin_0 = const()[name = tensor("op_3378_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3378_end_0 = const()[name = tensor("op_3378_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3378_end_mask_0 = const()[name = tensor("op_3378_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3378_cast_fp16 = slice_by_index(begin = var_3378_begin_0, end = var_3378_end_0, end_mask = var_3378_end_mask_0, x = matrix_bd_95_cast_fp16)[name = tensor("op_3378_cast_fp16")]; + tensor var_3379_to_fp16 = const()[name = tensor("op_3379_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_23_cast_fp16 = mul(x = var_3378_cast_fp16, y = var_3379_to_fp16)[name = tensor("qk_mask_23_cast_fp16")]; + tensor var_3383 = const()[name = tensor("op_3383"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_23_cast_fp16 = reshape(shape = var_3383, x = query_47_cast_fp16)[name = tensor("mh_q_23_cast_fp16")]; + tensor var_3385_to_fp16 = const()[name = tensor("op_3385_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3386_cast_fp16 = mul(x = mh_q_23_cast_fp16, y = var_3385_to_fp16)[name = tensor("op_3386_cast_fp16")]; + tensor var_3389 = const()[name = tensor("op_3389"), val = tensor([1, 8, 128, 188])]; + tensor var_3390_cast_fp16 = reshape(shape = var_3389, x = key_23_cast_fp16)[name = tensor("op_3390_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_3386_cast_fp16, y = var_3390_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_3394_cast_fp16 = softmax(axis = var_3235, x = mh_w_47_cast_fp16)[name = tensor("op_3394_cast_fp16")]; + tensor var_3395 = const()[name = tensor("op_3395"), val = tensor([1, 8, 128, 188])]; + tensor var_3396_cast_fp16 = reshape(shape = var_3395, x = value_23_cast_fp16)[name = tensor("op_3396_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_3396_cast_fp16, y = var_3394_cast_fp16)[name = tensor("attn_23_cast_fp16")]; + tensor var_3399 = const()[name = tensor("op_3399"), val = tensor([1, 1024, 1, 188])]; + tensor input_307_cast_fp16 = reshape(shape = var_3399, x = attn_23_cast_fp16)[name = tensor("input_307_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 = const()[name = tensor("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588790784)))]; + tensor obj_49_cast_fp16 = conv(bias = input_15_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, x = input_307_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_3417_to_fp16 = const()[name = tensor("op_3417_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_115_cast_fp16 = layer_norm(axes = out_115_axes_0, epsilon = var_3417_to_fp16, x = inputs_115_cast_fp16)[name = tensor("out_115_cast_fp16")]; + tensor input_309_gamma_0_to_fp16 = const()[name = tensor("input_309_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(590888000)))]; + tensor input_309_beta_0_to_fp16 = const()[name = tensor("input_309_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(590890112)))]; + tensor input_309_epsilon_0_to_fp16 = const()[name = tensor("input_309_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_309_cast_fp16 = batch_norm(beta = input_309_beta_0_to_fp16, epsilon = input_309_epsilon_0_to_fp16, gamma = input_309_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_115_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor input_311_pad_type_0 = const()[name = tensor("input_311_pad_type_0"), val = tensor("valid")]; + tensor input_311_strides_0 = const()[name = tensor("input_311_strides_0"), val = tensor([1, 1])]; + tensor input_311_pad_0 = const()[name = tensor("input_311_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_311_dilations_0 = const()[name = tensor("input_311_dilations_0"), val = tensor([1, 1])]; + tensor input_311_groups_0 = const()[name = tensor("input_311_groups_0"), val = tensor(1)]; + tensor layers_11_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_11_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(590892224)))]; + tensor input_311_cast_fp16 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = layers_11_conv_pointwise_conv1_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("input_311_cast_fp16")]; + tensor input_313_split_num_splits_0 = const()[name = tensor("input_313_split_num_splits_0"), val = tensor(2)]; + tensor input_313_split_axis_0 = const()[name = tensor("input_313_split_axis_0"), val = tensor(1)]; + tensor input_313_split_cast_fp16_0, tensor input_313_split_cast_fp16_1 = split(axis = input_313_split_axis_0, num_splits = input_313_split_num_splits_0, x = input_311_cast_fp16)[name = tensor("input_313_split_cast_fp16")]; + tensor input_313_split_1_sigmoid_cast_fp16 = sigmoid(x = input_313_split_cast_fp16_1)[name = tensor("input_313_split_1_sigmoid_cast_fp16")]; + tensor input_313_cast_fp16 = mul(x = input_313_split_cast_fp16_0, y = input_313_split_1_sigmoid_cast_fp16)[name = tensor("input_313_cast_fp16")]; + tensor input_315_pad_type_0 = const()[name = tensor("input_315_pad_type_0"), val = tensor("custom")]; + tensor input_315_pad_0 = const()[name = tensor("input_315_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_315_groups_0 = const()[name = tensor("input_315_groups_0"), val = tensor(1024)]; + tensor input_315_strides_0 = const()[name = tensor("input_315_strides_0"), val = tensor([1, 1])]; + tensor input_315_dilations_0 = const()[name = tensor("input_315_dilations_0"), val = tensor([1, 1])]; + tensor const_290_to_fp16 = const()[name = tensor("const_290_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595086592)))]; + tensor const_291_to_fp16 = const()[name = tensor("const_291_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595105088)))]; + tensor input_317_cast_fp16 = conv(bias = const_291_to_fp16, dilations = input_315_dilations_0, groups = input_315_groups_0, pad = input_315_pad_0, pad_type = input_315_pad_type_0, strides = input_315_strides_0, weight = const_290_to_fp16, x = input_313_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor input_319_cast_fp16 = silu(x = input_317_cast_fp16)[name = tensor("input_319_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 = const()[name = tensor("layers_11_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(595107200)))]; + 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, x = input_319_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_3465_to_fp16 = const()[name = tensor("op_3465_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_117_cast_fp16 = layer_norm(axes = out_117_axes_0, epsilon = var_3465_to_fp16, x = inputs_117_cast_fp16)[name = tensor("out_117_cast_fp16")]; + tensor input_321_gamma_0_to_fp16 = const()[name = tensor("input_321_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597204416)))]; + tensor input_321_beta_0_to_fp16 = const()[name = tensor("input_321_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597206528)))]; + tensor input_321_epsilon_0_to_fp16 = const()[name = tensor("input_321_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_321_cast_fp16 = batch_norm(beta = input_321_beta_0_to_fp16, epsilon = input_321_epsilon_0_to_fp16, gamma = input_321_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_117_cast_fp16)[name = tensor("input_321_cast_fp16")]; + tensor input_323_pad_type_0 = const()[name = tensor("input_323_pad_type_0"), val = tensor("valid")]; + tensor input_323_strides_0 = const()[name = tensor("input_323_strides_0"), val = tensor([1, 1])]; + tensor input_323_pad_0 = const()[name = tensor("input_323_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_323_dilations_0 = const()[name = tensor("input_323_dilations_0"), val = tensor([1, 1])]; + tensor input_323_groups_0 = const()[name = tensor("input_323_groups_0"), val = tensor(1)]; + tensor layers_11_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_11_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(597208640)))]; + tensor input_323_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_323_dilations_0, groups = input_323_groups_0, pad = input_323_pad_0, pad_type = input_323_pad_type_0, strides = input_323_strides_0, weight = layers_11_feed_forward2_fc1_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("input_323_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = input_323_cast_fp16)[name = tensor("input_325_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 var_3493_weight_0_to_fp16 = const()[name = tensor("op_3493_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605597312)))]; + tensor var_3493_cast_fp16 = conv(bias = input_15_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 = var_3493_weight_0_to_fp16, x = input_325_cast_fp16)[name = tensor("op_3493_cast_fp16")]; + tensor inputs_119_cast_fp16 = add(x = inputs_117_cast_fp16, y = var_3493_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_3503_to_fp16 = const()[name = tensor("op_3503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_119_cast_fp16 = layer_norm(axes = out_119_axes_0, epsilon = var_3503_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(613985984)))]; + 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(613988096)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_119_cast_fp16)[name = tensor("inputs_121_cast_fp16")]; + tensor var_3517 = const()[name = tensor("op_3517"), val = tensor(3)]; + tensor out_121_axes_0 = const()[name = tensor("out_121_axes_0"), val = tensor([1])]; + tensor var_3548_to_fp16 = const()[name = tensor("op_3548_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_121_cast_fp16 = layer_norm(axes = out_121_axes_0, epsilon = var_3548_to_fp16, x = inputs_121_cast_fp16)[name = tensor("out_121_cast_fp16")]; + tensor input_327_gamma_0_to_fp16 = const()[name = tensor("input_327_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(613990208)))]; + tensor input_327_beta_0_to_fp16 = const()[name = tensor("input_327_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(613992320)))]; + tensor input_327_epsilon_0_to_fp16 = const()[name = tensor("input_327_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_327_cast_fp16 = batch_norm(beta = input_327_beta_0_to_fp16, epsilon = input_327_epsilon_0_to_fp16, gamma = input_327_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_121_cast_fp16)[name = tensor("input_327_cast_fp16")]; + tensor input_329_pad_type_0 = const()[name = tensor("input_329_pad_type_0"), val = tensor("valid")]; + tensor input_329_strides_0 = const()[name = tensor("input_329_strides_0"), val = tensor([1, 1])]; + tensor input_329_pad_0 = const()[name = tensor("input_329_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_329_dilations_0 = const()[name = tensor("input_329_dilations_0"), val = tensor([1, 1])]; + tensor input_329_groups_0 = const()[name = tensor("input_329_groups_0"), val = tensor(1)]; + tensor layers_12_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_12_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(613994432)))]; + tensor input_329_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_329_dilations_0, groups = input_329_groups_0, pad = input_329_pad_0, pad_type = input_329_pad_type_0, strides = input_329_strides_0, weight = layers_12_feed_forward1_fc1_weight_to_fp16, x = input_327_cast_fp16)[name = tensor("input_329_cast_fp16")]; + tensor input_331_cast_fp16 = silu(x = input_329_cast_fp16)[name = tensor("input_331_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 var_3576_weight_0_to_fp16 = const()[name = tensor("op_3576_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(622383104)))]; + tensor var_3576_cast_fp16 = conv(bias = input_15_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 = var_3576_weight_0_to_fp16, x = input_331_cast_fp16)[name = tensor("op_3576_cast_fp16")]; + tensor inputs_123_cast_fp16 = add(x = inputs_121_cast_fp16, y = var_3576_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_3586_to_fp16 = const()[name = tensor("op_3586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_123_cast_fp16 = layer_norm(axes = out_123_axes_0, epsilon = var_3586_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(630771776)))]; + 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(630773888)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(630776000)))]; + tensor query_49_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(632873216)))]; + 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, 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 = const()[name = tensor("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(634970432)))]; + tensor value_25_cast_fp16 = conv(bias = input_15_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, x = obj_51_cast_fp16)[name = tensor("value_25_cast_fp16")]; + tensor var_3624_to_fp16 = const()[name = tensor("op_3624_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(637067648)))]; + tensor query_51_cast_fp16 = add(x = query_49_cast_fp16, y = var_3624_to_fp16)[name = tensor("query_51_cast_fp16")]; + tensor var_3627_to_fp16 = const()[name = tensor("op_3627_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(637069760)))]; + tensor q_with_bias_v_25_cast_fp16 = add(x = query_49_cast_fp16, y = var_3627_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 = const()[name = tensor("layers_12_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(637071872)))]; + tensor p_25_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_25_cast_fp16")]; + tensor var_3638 = const()[name = tensor("op_3638"), val = tensor([1, 8, 128, 188])]; + tensor var_3639_cast_fp16 = reshape(shape = var_3638, x = q_with_bias_v_25_cast_fp16)[name = tensor("op_3639_cast_fp16")]; + tensor var_3640 = const()[name = tensor("op_3640"), val = tensor([1, 8, 128, -1])]; + tensor var_3641_cast_fp16 = reshape(shape = var_3640, x = p_25_cast_fp16)[name = tensor("op_3641_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_3639_cast_fp16, y = var_3641_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_3650 = const()[name = tensor("op_3650"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_101_cast_fp16 = reshape(shape = var_3650, x = matrix_bd_99_cast_fp16)[name = tensor("matrix_bd_101_cast_fp16")]; + tensor var_3654_begin_0 = const()[name = tensor("op_3654_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3654_end_0 = const()[name = tensor("op_3654_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3654_end_mask_0 = const()[name = tensor("op_3654_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3654_cast_fp16 = slice_by_index(begin = var_3654_begin_0, end = var_3654_end_0, end_mask = var_3654_end_mask_0, x = matrix_bd_101_cast_fp16)[name = tensor("op_3654_cast_fp16")]; + tensor var_3655 = const()[name = tensor("op_3655"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_103_cast_fp16 = reshape(shape = var_3655, x = var_3654_cast_fp16)[name = tensor("matrix_bd_103_cast_fp16")]; + tensor var_3660_begin_0 = const()[name = tensor("op_3660_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3660_end_0 = const()[name = tensor("op_3660_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3660_end_mask_0 = const()[name = tensor("op_3660_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3660_cast_fp16 = slice_by_index(begin = var_3660_begin_0, end = var_3660_end_0, end_mask = var_3660_end_mask_0, x = matrix_bd_103_cast_fp16)[name = tensor("op_3660_cast_fp16")]; + tensor var_3661_to_fp16 = const()[name = tensor("op_3661_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_25_cast_fp16 = mul(x = var_3660_cast_fp16, y = var_3661_to_fp16)[name = tensor("qk_mask_25_cast_fp16")]; + tensor var_3665 = const()[name = tensor("op_3665"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_25_cast_fp16 = reshape(shape = var_3665, x = query_51_cast_fp16)[name = tensor("mh_q_25_cast_fp16")]; + tensor var_3667_to_fp16 = const()[name = tensor("op_3667_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3668_cast_fp16 = mul(x = mh_q_25_cast_fp16, y = var_3667_to_fp16)[name = tensor("op_3668_cast_fp16")]; + tensor var_3671 = const()[name = tensor("op_3671"), val = tensor([1, 8, 128, 188])]; + tensor var_3672_cast_fp16 = reshape(shape = var_3671, x = key_25_cast_fp16)[name = tensor("op_3672_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_3668_cast_fp16, y = var_3672_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_3676_cast_fp16 = softmax(axis = var_3517, x = mh_w_51_cast_fp16)[name = tensor("op_3676_cast_fp16")]; + tensor var_3677 = const()[name = tensor("op_3677"), val = tensor([1, 8, 128, 188])]; + tensor var_3678_cast_fp16 = reshape(shape = var_3677, x = value_25_cast_fp16)[name = tensor("op_3678_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_3678_cast_fp16, y = var_3676_cast_fp16)[name = tensor("attn_25_cast_fp16")]; + tensor var_3681 = const()[name = tensor("op_3681"), val = tensor([1, 1024, 1, 188])]; + tensor input_333_cast_fp16 = reshape(shape = var_3681, x = attn_25_cast_fp16)[name = tensor("input_333_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 = const()[name = tensor("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(639169088)))]; + tensor obj_53_cast_fp16 = conv(bias = input_15_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, x = input_333_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_3699_to_fp16 = const()[name = tensor("op_3699_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_125_cast_fp16 = layer_norm(axes = out_125_axes_0, epsilon = var_3699_to_fp16, x = inputs_125_cast_fp16)[name = tensor("out_125_cast_fp16")]; + tensor input_335_gamma_0_to_fp16 = const()[name = tensor("input_335_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(641266304)))]; + tensor input_335_beta_0_to_fp16 = const()[name = tensor("input_335_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(641268416)))]; + tensor input_335_epsilon_0_to_fp16 = const()[name = tensor("input_335_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_335_cast_fp16 = batch_norm(beta = input_335_beta_0_to_fp16, epsilon = input_335_epsilon_0_to_fp16, gamma = input_335_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_125_cast_fp16)[name = tensor("input_335_cast_fp16")]; + tensor input_337_pad_type_0 = const()[name = tensor("input_337_pad_type_0"), val = tensor("valid")]; + tensor input_337_strides_0 = const()[name = tensor("input_337_strides_0"), val = tensor([1, 1])]; + tensor input_337_pad_0 = const()[name = tensor("input_337_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_337_dilations_0 = const()[name = tensor("input_337_dilations_0"), val = tensor([1, 1])]; + tensor input_337_groups_0 = const()[name = tensor("input_337_groups_0"), val = tensor(1)]; + tensor layers_12_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_12_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(641270528)))]; + tensor input_337_cast_fp16 = conv(dilations = input_337_dilations_0, groups = input_337_groups_0, pad = input_337_pad_0, pad_type = input_337_pad_type_0, strides = input_337_strides_0, weight = layers_12_conv_pointwise_conv1_weight_to_fp16, x = input_335_cast_fp16)[name = tensor("input_337_cast_fp16")]; + tensor input_339_split_num_splits_0 = const()[name = tensor("input_339_split_num_splits_0"), val = tensor(2)]; + tensor input_339_split_axis_0 = const()[name = tensor("input_339_split_axis_0"), val = tensor(1)]; + tensor input_339_split_cast_fp16_0, tensor input_339_split_cast_fp16_1 = split(axis = input_339_split_axis_0, num_splits = input_339_split_num_splits_0, x = input_337_cast_fp16)[name = tensor("input_339_split_cast_fp16")]; + tensor input_339_split_1_sigmoid_cast_fp16 = sigmoid(x = input_339_split_cast_fp16_1)[name = tensor("input_339_split_1_sigmoid_cast_fp16")]; + tensor input_339_cast_fp16 = mul(x = input_339_split_cast_fp16_0, y = input_339_split_1_sigmoid_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor input_341_pad_type_0 = const()[name = tensor("input_341_pad_type_0"), val = tensor("custom")]; + tensor input_341_pad_0 = const()[name = tensor("input_341_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_341_groups_0 = const()[name = tensor("input_341_groups_0"), val = tensor(1024)]; + tensor input_341_strides_0 = const()[name = tensor("input_341_strides_0"), val = tensor([1, 1])]; + tensor input_341_dilations_0 = const()[name = tensor("input_341_dilations_0"), val = tensor([1, 1])]; + tensor const_292_to_fp16 = const()[name = tensor("const_292_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(645464896)))]; + tensor const_293_to_fp16 = const()[name = tensor("const_293_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(645483392)))]; + tensor input_343_cast_fp16 = conv(bias = const_293_to_fp16, dilations = input_341_dilations_0, groups = input_341_groups_0, pad = input_341_pad_0, pad_type = input_341_pad_type_0, strides = input_341_strides_0, weight = const_292_to_fp16, x = input_339_cast_fp16)[name = tensor("input_343_cast_fp16")]; + tensor input_345_cast_fp16 = silu(x = input_343_cast_fp16)[name = tensor("input_345_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 = const()[name = tensor("layers_12_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(645485504)))]; + 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, x = input_345_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_3747_to_fp16 = const()[name = tensor("op_3747_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_127_cast_fp16 = layer_norm(axes = out_127_axes_0, epsilon = var_3747_to_fp16, x = inputs_127_cast_fp16)[name = tensor("out_127_cast_fp16")]; + tensor input_347_gamma_0_to_fp16 = const()[name = tensor("input_347_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647582720)))]; + tensor input_347_beta_0_to_fp16 = const()[name = tensor("input_347_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647584832)))]; + tensor input_347_epsilon_0_to_fp16 = const()[name = tensor("input_347_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_347_cast_fp16 = batch_norm(beta = input_347_beta_0_to_fp16, epsilon = input_347_epsilon_0_to_fp16, gamma = input_347_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_127_cast_fp16)[name = tensor("input_347_cast_fp16")]; + tensor input_349_pad_type_0 = const()[name = tensor("input_349_pad_type_0"), val = tensor("valid")]; + tensor input_349_strides_0 = const()[name = tensor("input_349_strides_0"), val = tensor([1, 1])]; + tensor input_349_pad_0 = const()[name = tensor("input_349_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_349_dilations_0 = const()[name = tensor("input_349_dilations_0"), val = tensor([1, 1])]; + tensor input_349_groups_0 = const()[name = tensor("input_349_groups_0"), val = tensor(1)]; + tensor layers_12_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_12_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(647586944)))]; + tensor input_349_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_349_dilations_0, groups = input_349_groups_0, pad = input_349_pad_0, pad_type = input_349_pad_type_0, strides = input_349_strides_0, weight = layers_12_feed_forward2_fc1_weight_to_fp16, x = input_347_cast_fp16)[name = tensor("input_349_cast_fp16")]; + tensor input_351_cast_fp16 = silu(x = input_349_cast_fp16)[name = tensor("input_351_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 var_3775_weight_0_to_fp16 = const()[name = tensor("op_3775_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(655975616)))]; + tensor var_3775_cast_fp16 = conv(bias = input_15_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 = var_3775_weight_0_to_fp16, x = input_351_cast_fp16)[name = tensor("op_3775_cast_fp16")]; + tensor inputs_129_cast_fp16 = add(x = inputs_127_cast_fp16, y = var_3775_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_3785_to_fp16 = const()[name = tensor("op_3785_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_129_cast_fp16 = layer_norm(axes = out_129_axes_0, epsilon = var_3785_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(664364288)))]; + 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(664366400)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_129_cast_fp16)[name = tensor("inputs_131_cast_fp16")]; + tensor var_3799 = const()[name = tensor("op_3799"), val = tensor(3)]; + tensor out_131_axes_0 = const()[name = tensor("out_131_axes_0"), val = tensor([1])]; + tensor var_3830_to_fp16 = const()[name = tensor("op_3830_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_131_cast_fp16 = layer_norm(axes = out_131_axes_0, epsilon = var_3830_to_fp16, x = inputs_131_cast_fp16)[name = tensor("out_131_cast_fp16")]; + tensor input_353_gamma_0_to_fp16 = const()[name = tensor("input_353_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(664368512)))]; + tensor input_353_beta_0_to_fp16 = const()[name = tensor("input_353_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(664370624)))]; + tensor input_353_epsilon_0_to_fp16 = const()[name = tensor("input_353_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_353_cast_fp16 = batch_norm(beta = input_353_beta_0_to_fp16, epsilon = input_353_epsilon_0_to_fp16, gamma = input_353_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_131_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_pad_type_0 = const()[name = tensor("input_355_pad_type_0"), val = tensor("valid")]; + tensor input_355_strides_0 = const()[name = tensor("input_355_strides_0"), val = tensor([1, 1])]; + tensor input_355_pad_0 = const()[name = tensor("input_355_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_355_dilations_0 = const()[name = tensor("input_355_dilations_0"), val = tensor([1, 1])]; + tensor input_355_groups_0 = const()[name = tensor("input_355_groups_0"), val = tensor(1)]; + tensor layers_13_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_13_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(664372736)))]; + tensor input_355_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_355_dilations_0, groups = input_355_groups_0, pad = input_355_pad_0, pad_type = input_355_pad_type_0, strides = input_355_strides_0, weight = layers_13_feed_forward1_fc1_weight_to_fp16, x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor input_357_cast_fp16 = silu(x = input_355_cast_fp16)[name = tensor("input_357_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 var_3858_weight_0_to_fp16 = const()[name = tensor("op_3858_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(672761408)))]; + tensor var_3858_cast_fp16 = conv(bias = input_15_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 = var_3858_weight_0_to_fp16, x = input_357_cast_fp16)[name = tensor("op_3858_cast_fp16")]; + tensor inputs_133_cast_fp16 = add(x = inputs_131_cast_fp16, y = var_3858_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_3868_to_fp16 = const()[name = tensor("op_3868_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_133_cast_fp16 = layer_norm(axes = out_133_axes_0, epsilon = var_3868_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(681150080)))]; + 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(681152192)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(681154304)))]; + tensor query_53_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(683251520)))]; + 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, 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 = const()[name = tensor("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(685348736)))]; + tensor value_27_cast_fp16 = conv(bias = input_15_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, x = obj_55_cast_fp16)[name = tensor("value_27_cast_fp16")]; + tensor var_3906_to_fp16 = const()[name = tensor("op_3906_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(687445952)))]; + tensor query_55_cast_fp16 = add(x = query_53_cast_fp16, y = var_3906_to_fp16)[name = tensor("query_55_cast_fp16")]; + tensor var_3909_to_fp16 = const()[name = tensor("op_3909_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(687448064)))]; + tensor q_with_bias_v_27_cast_fp16 = add(x = query_53_cast_fp16, y = var_3909_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 = const()[name = tensor("layers_13_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(687450176)))]; + tensor p_27_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_27_cast_fp16")]; + tensor var_3920 = const()[name = tensor("op_3920"), val = tensor([1, 8, 128, 188])]; + tensor var_3921_cast_fp16 = reshape(shape = var_3920, x = q_with_bias_v_27_cast_fp16)[name = tensor("op_3921_cast_fp16")]; + tensor var_3922 = const()[name = tensor("op_3922"), val = tensor([1, 8, 128, -1])]; + tensor var_3923_cast_fp16 = reshape(shape = var_3922, x = p_27_cast_fp16)[name = tensor("op_3923_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_3921_cast_fp16, y = var_3923_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_3932 = const()[name = tensor("op_3932"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_109_cast_fp16 = reshape(shape = var_3932, x = matrix_bd_107_cast_fp16)[name = tensor("matrix_bd_109_cast_fp16")]; + tensor var_3936_begin_0 = const()[name = tensor("op_3936_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_3936_end_0 = const()[name = tensor("op_3936_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_3936_end_mask_0 = const()[name = tensor("op_3936_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_3936_cast_fp16 = slice_by_index(begin = var_3936_begin_0, end = var_3936_end_0, end_mask = var_3936_end_mask_0, x = matrix_bd_109_cast_fp16)[name = tensor("op_3936_cast_fp16")]; + tensor var_3937 = const()[name = tensor("op_3937"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_111_cast_fp16 = reshape(shape = var_3937, x = var_3936_cast_fp16)[name = tensor("matrix_bd_111_cast_fp16")]; + tensor var_3942_begin_0 = const()[name = tensor("op_3942_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_3942_end_0 = const()[name = tensor("op_3942_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_3942_end_mask_0 = const()[name = tensor("op_3942_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_3942_cast_fp16 = slice_by_index(begin = var_3942_begin_0, end = var_3942_end_0, end_mask = var_3942_end_mask_0, x = matrix_bd_111_cast_fp16)[name = tensor("op_3942_cast_fp16")]; + tensor var_3943_to_fp16 = const()[name = tensor("op_3943_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_27_cast_fp16 = mul(x = var_3942_cast_fp16, y = var_3943_to_fp16)[name = tensor("qk_mask_27_cast_fp16")]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_27_cast_fp16 = reshape(shape = var_3947, x = query_55_cast_fp16)[name = tensor("mh_q_27_cast_fp16")]; + tensor var_3949_to_fp16 = const()[name = tensor("op_3949_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_3950_cast_fp16 = mul(x = mh_q_27_cast_fp16, y = var_3949_to_fp16)[name = tensor("op_3950_cast_fp16")]; + tensor var_3953 = const()[name = tensor("op_3953"), val = tensor([1, 8, 128, 188])]; + tensor var_3954_cast_fp16 = reshape(shape = var_3953, x = key_27_cast_fp16)[name = tensor("op_3954_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_3950_cast_fp16, y = var_3954_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_3958_cast_fp16 = softmax(axis = var_3799, x = mh_w_55_cast_fp16)[name = tensor("op_3958_cast_fp16")]; + tensor var_3959 = const()[name = tensor("op_3959"), val = tensor([1, 8, 128, 188])]; + tensor var_3960_cast_fp16 = reshape(shape = var_3959, x = value_27_cast_fp16)[name = tensor("op_3960_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_3960_cast_fp16, y = var_3958_cast_fp16)[name = tensor("attn_27_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1024, 1, 188])]; + tensor input_359_cast_fp16 = reshape(shape = var_3963, x = attn_27_cast_fp16)[name = tensor("input_359_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 = const()[name = tensor("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(689547392)))]; + tensor obj_57_cast_fp16 = conv(bias = input_15_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, x = input_359_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_3981_to_fp16 = const()[name = tensor("op_3981_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_135_cast_fp16 = layer_norm(axes = out_135_axes_0, epsilon = var_3981_to_fp16, x = inputs_135_cast_fp16)[name = tensor("out_135_cast_fp16")]; + tensor input_361_gamma_0_to_fp16 = const()[name = tensor("input_361_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(691644608)))]; + tensor input_361_beta_0_to_fp16 = const()[name = tensor("input_361_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(691646720)))]; + tensor input_361_epsilon_0_to_fp16 = const()[name = tensor("input_361_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_361_cast_fp16 = batch_norm(beta = input_361_beta_0_to_fp16, epsilon = input_361_epsilon_0_to_fp16, gamma = input_361_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_135_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor input_363_pad_type_0 = const()[name = tensor("input_363_pad_type_0"), val = tensor("valid")]; + tensor input_363_strides_0 = const()[name = tensor("input_363_strides_0"), val = tensor([1, 1])]; + tensor input_363_pad_0 = const()[name = tensor("input_363_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_363_dilations_0 = const()[name = tensor("input_363_dilations_0"), val = tensor([1, 1])]; + tensor input_363_groups_0 = const()[name = tensor("input_363_groups_0"), val = tensor(1)]; + tensor layers_13_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_13_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(691648832)))]; + tensor input_363_cast_fp16 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = layers_13_conv_pointwise_conv1_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("input_363_cast_fp16")]; + tensor input_365_split_num_splits_0 = const()[name = tensor("input_365_split_num_splits_0"), val = tensor(2)]; + tensor input_365_split_axis_0 = const()[name = tensor("input_365_split_axis_0"), val = tensor(1)]; + tensor input_365_split_cast_fp16_0, tensor input_365_split_cast_fp16_1 = split(axis = input_365_split_axis_0, num_splits = input_365_split_num_splits_0, x = input_363_cast_fp16)[name = tensor("input_365_split_cast_fp16")]; + tensor input_365_split_1_sigmoid_cast_fp16 = sigmoid(x = input_365_split_cast_fp16_1)[name = tensor("input_365_split_1_sigmoid_cast_fp16")]; + tensor input_365_cast_fp16 = mul(x = input_365_split_cast_fp16_0, y = input_365_split_1_sigmoid_cast_fp16)[name = tensor("input_365_cast_fp16")]; + tensor input_367_pad_type_0 = const()[name = tensor("input_367_pad_type_0"), val = tensor("custom")]; + tensor input_367_pad_0 = const()[name = tensor("input_367_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_367_groups_0 = const()[name = tensor("input_367_groups_0"), val = tensor(1024)]; + tensor input_367_strides_0 = const()[name = tensor("input_367_strides_0"), val = tensor([1, 1])]; + tensor input_367_dilations_0 = const()[name = tensor("input_367_dilations_0"), val = tensor([1, 1])]; + tensor const_294_to_fp16 = const()[name = tensor("const_294_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(695843200)))]; + tensor const_295_to_fp16 = const()[name = tensor("const_295_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(695861696)))]; + tensor input_369_cast_fp16 = conv(bias = const_295_to_fp16, dilations = input_367_dilations_0, groups = input_367_groups_0, pad = input_367_pad_0, pad_type = input_367_pad_type_0, strides = input_367_strides_0, weight = const_294_to_fp16, x = input_365_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor input_371_cast_fp16 = silu(x = input_369_cast_fp16)[name = tensor("input_371_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 = const()[name = tensor("layers_13_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(695863808)))]; + 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, x = input_371_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_4029_to_fp16 = const()[name = tensor("op_4029_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_137_cast_fp16 = layer_norm(axes = out_137_axes_0, epsilon = var_4029_to_fp16, x = inputs_137_cast_fp16)[name = tensor("out_137_cast_fp16")]; + tensor input_373_gamma_0_to_fp16 = const()[name = tensor("input_373_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(697961024)))]; + tensor input_373_beta_0_to_fp16 = const()[name = tensor("input_373_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(697963136)))]; + tensor input_373_epsilon_0_to_fp16 = const()[name = tensor("input_373_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_373_cast_fp16 = batch_norm(beta = input_373_beta_0_to_fp16, epsilon = input_373_epsilon_0_to_fp16, gamma = input_373_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_137_cast_fp16)[name = tensor("input_373_cast_fp16")]; + tensor input_375_pad_type_0 = const()[name = tensor("input_375_pad_type_0"), val = tensor("valid")]; + tensor input_375_strides_0 = const()[name = tensor("input_375_strides_0"), val = tensor([1, 1])]; + tensor input_375_pad_0 = const()[name = tensor("input_375_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_375_dilations_0 = const()[name = tensor("input_375_dilations_0"), val = tensor([1, 1])]; + tensor input_375_groups_0 = const()[name = tensor("input_375_groups_0"), val = tensor(1)]; + tensor layers_13_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_13_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(697965248)))]; + tensor input_375_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_375_dilations_0, groups = input_375_groups_0, pad = input_375_pad_0, pad_type = input_375_pad_type_0, strides = input_375_strides_0, weight = layers_13_feed_forward2_fc1_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("input_375_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = input_375_cast_fp16)[name = tensor("input_377_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 var_4057_weight_0_to_fp16 = const()[name = tensor("op_4057_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(706353920)))]; + tensor var_4057_cast_fp16 = conv(bias = input_15_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 = var_4057_weight_0_to_fp16, x = input_377_cast_fp16)[name = tensor("op_4057_cast_fp16")]; + tensor inputs_139_cast_fp16 = add(x = inputs_137_cast_fp16, y = var_4057_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_4067_to_fp16 = const()[name = tensor("op_4067_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_139_cast_fp16 = layer_norm(axes = out_139_axes_0, epsilon = var_4067_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(714742592)))]; + 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(714744704)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_139_cast_fp16)[name = tensor("inputs_141_cast_fp16")]; + tensor var_4081 = const()[name = tensor("op_4081"), val = tensor(3)]; + tensor out_141_axes_0 = const()[name = tensor("out_141_axes_0"), val = tensor([1])]; + tensor var_4112_to_fp16 = const()[name = tensor("op_4112_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_141_cast_fp16 = layer_norm(axes = out_141_axes_0, epsilon = var_4112_to_fp16, x = inputs_141_cast_fp16)[name = tensor("out_141_cast_fp16")]; + tensor input_379_gamma_0_to_fp16 = const()[name = tensor("input_379_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(714746816)))]; + tensor input_379_beta_0_to_fp16 = const()[name = tensor("input_379_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(714748928)))]; + tensor input_379_epsilon_0_to_fp16 = const()[name = tensor("input_379_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_379_cast_fp16 = batch_norm(beta = input_379_beta_0_to_fp16, epsilon = input_379_epsilon_0_to_fp16, gamma = input_379_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_141_cast_fp16)[name = tensor("input_379_cast_fp16")]; + tensor input_381_pad_type_0 = const()[name = tensor("input_381_pad_type_0"), val = tensor("valid")]; + tensor input_381_strides_0 = const()[name = tensor("input_381_strides_0"), val = tensor([1, 1])]; + tensor input_381_pad_0 = const()[name = tensor("input_381_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_381_dilations_0 = const()[name = tensor("input_381_dilations_0"), val = tensor([1, 1])]; + tensor input_381_groups_0 = const()[name = tensor("input_381_groups_0"), val = tensor(1)]; + tensor layers_14_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_14_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(714751040)))]; + tensor input_381_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_381_dilations_0, groups = input_381_groups_0, pad = input_381_pad_0, pad_type = input_381_pad_type_0, strides = input_381_strides_0, weight = layers_14_feed_forward1_fc1_weight_to_fp16, x = input_379_cast_fp16)[name = tensor("input_381_cast_fp16")]; + tensor input_383_cast_fp16 = silu(x = input_381_cast_fp16)[name = tensor("input_383_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 var_4140_weight_0_to_fp16 = const()[name = tensor("op_4140_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(723139712)))]; + tensor var_4140_cast_fp16 = conv(bias = input_15_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 = var_4140_weight_0_to_fp16, x = input_383_cast_fp16)[name = tensor("op_4140_cast_fp16")]; + tensor inputs_143_cast_fp16 = add(x = inputs_141_cast_fp16, y = var_4140_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_4150_to_fp16 = const()[name = tensor("op_4150_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_143_cast_fp16 = layer_norm(axes = out_143_axes_0, epsilon = var_4150_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(731528384)))]; + 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(731530496)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(731532608)))]; + tensor query_57_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(733629824)))]; + 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, 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 = const()[name = tensor("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(735727040)))]; + tensor value_29_cast_fp16 = conv(bias = input_15_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, x = obj_59_cast_fp16)[name = tensor("value_29_cast_fp16")]; + tensor var_4188_to_fp16 = const()[name = tensor("op_4188_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737824256)))]; + tensor query_59_cast_fp16 = add(x = query_57_cast_fp16, y = var_4188_to_fp16)[name = tensor("query_59_cast_fp16")]; + tensor var_4191_to_fp16 = const()[name = tensor("op_4191_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737826368)))]; + tensor q_with_bias_v_29_cast_fp16 = add(x = query_57_cast_fp16, y = var_4191_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 = const()[name = tensor("layers_14_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(737828480)))]; + tensor p_29_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_29_cast_fp16")]; + tensor var_4202 = const()[name = tensor("op_4202"), val = tensor([1, 8, 128, 188])]; + tensor var_4203_cast_fp16 = reshape(shape = var_4202, x = q_with_bias_v_29_cast_fp16)[name = tensor("op_4203_cast_fp16")]; + tensor var_4204 = const()[name = tensor("op_4204"), val = tensor([1, 8, 128, -1])]; + tensor var_4205_cast_fp16 = reshape(shape = var_4204, x = p_29_cast_fp16)[name = tensor("op_4205_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_4203_cast_fp16, y = var_4205_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_4214 = const()[name = tensor("op_4214"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_117_cast_fp16 = reshape(shape = var_4214, x = matrix_bd_115_cast_fp16)[name = tensor("matrix_bd_117_cast_fp16")]; + tensor var_4218_begin_0 = const()[name = tensor("op_4218_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4218_end_0 = const()[name = tensor("op_4218_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_4218_end_mask_0 = const()[name = tensor("op_4218_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4218_cast_fp16 = slice_by_index(begin = var_4218_begin_0, end = var_4218_end_0, end_mask = var_4218_end_mask_0, x = matrix_bd_117_cast_fp16)[name = tensor("op_4218_cast_fp16")]; + tensor var_4219 = const()[name = tensor("op_4219"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_119_cast_fp16 = reshape(shape = var_4219, x = var_4218_cast_fp16)[name = tensor("matrix_bd_119_cast_fp16")]; + tensor var_4224_begin_0 = const()[name = tensor("op_4224_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4224_end_0 = const()[name = tensor("op_4224_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_4224_end_mask_0 = const()[name = tensor("op_4224_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_4224_cast_fp16 = slice_by_index(begin = var_4224_begin_0, end = var_4224_end_0, end_mask = var_4224_end_mask_0, x = matrix_bd_119_cast_fp16)[name = tensor("op_4224_cast_fp16")]; + tensor var_4225_to_fp16 = const()[name = tensor("op_4225_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_29_cast_fp16 = mul(x = var_4224_cast_fp16, y = var_4225_to_fp16)[name = tensor("qk_mask_29_cast_fp16")]; + tensor var_4229 = const()[name = tensor("op_4229"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_29_cast_fp16 = reshape(shape = var_4229, x = query_59_cast_fp16)[name = tensor("mh_q_29_cast_fp16")]; + tensor var_4231_to_fp16 = const()[name = tensor("op_4231_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_4232_cast_fp16 = mul(x = mh_q_29_cast_fp16, y = var_4231_to_fp16)[name = tensor("op_4232_cast_fp16")]; + tensor var_4235 = const()[name = tensor("op_4235"), val = tensor([1, 8, 128, 188])]; + tensor var_4236_cast_fp16 = reshape(shape = var_4235, x = key_29_cast_fp16)[name = tensor("op_4236_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_4232_cast_fp16, y = var_4236_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_4240_cast_fp16 = softmax(axis = var_4081, x = mh_w_59_cast_fp16)[name = tensor("op_4240_cast_fp16")]; + tensor var_4241 = const()[name = tensor("op_4241"), val = tensor([1, 8, 128, 188])]; + tensor var_4242_cast_fp16 = reshape(shape = var_4241, x = value_29_cast_fp16)[name = tensor("op_4242_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_4242_cast_fp16, y = var_4240_cast_fp16)[name = tensor("attn_29_cast_fp16")]; + tensor var_4245 = const()[name = tensor("op_4245"), val = tensor([1, 1024, 1, 188])]; + tensor input_385_cast_fp16 = reshape(shape = var_4245, x = attn_29_cast_fp16)[name = tensor("input_385_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 = const()[name = tensor("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(739925696)))]; + tensor obj_61_cast_fp16 = conv(bias = input_15_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, x = input_385_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_4263_to_fp16 = const()[name = tensor("op_4263_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_145_cast_fp16 = layer_norm(axes = out_145_axes_0, epsilon = var_4263_to_fp16, x = inputs_145_cast_fp16)[name = tensor("out_145_cast_fp16")]; + tensor input_387_gamma_0_to_fp16 = const()[name = tensor("input_387_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(742022912)))]; + tensor input_387_beta_0_to_fp16 = const()[name = tensor("input_387_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(742025024)))]; + tensor input_387_epsilon_0_to_fp16 = const()[name = tensor("input_387_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_387_cast_fp16 = batch_norm(beta = input_387_beta_0_to_fp16, epsilon = input_387_epsilon_0_to_fp16, gamma = input_387_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_145_cast_fp16)[name = tensor("input_387_cast_fp16")]; + tensor input_389_pad_type_0 = const()[name = tensor("input_389_pad_type_0"), val = tensor("valid")]; + tensor input_389_strides_0 = const()[name = tensor("input_389_strides_0"), val = tensor([1, 1])]; + tensor input_389_pad_0 = const()[name = tensor("input_389_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_389_dilations_0 = const()[name = tensor("input_389_dilations_0"), val = tensor([1, 1])]; + tensor input_389_groups_0 = const()[name = tensor("input_389_groups_0"), val = tensor(1)]; + tensor layers_14_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_14_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(742027136)))]; + tensor input_389_cast_fp16 = conv(dilations = input_389_dilations_0, groups = input_389_groups_0, pad = input_389_pad_0, pad_type = input_389_pad_type_0, strides = input_389_strides_0, weight = layers_14_conv_pointwise_conv1_weight_to_fp16, x = input_387_cast_fp16)[name = tensor("input_389_cast_fp16")]; + tensor input_391_split_num_splits_0 = const()[name = tensor("input_391_split_num_splits_0"), val = tensor(2)]; + tensor input_391_split_axis_0 = const()[name = tensor("input_391_split_axis_0"), val = tensor(1)]; + tensor input_391_split_cast_fp16_0, tensor input_391_split_cast_fp16_1 = split(axis = input_391_split_axis_0, num_splits = input_391_split_num_splits_0, x = input_389_cast_fp16)[name = tensor("input_391_split_cast_fp16")]; + tensor input_391_split_1_sigmoid_cast_fp16 = sigmoid(x = input_391_split_cast_fp16_1)[name = tensor("input_391_split_1_sigmoid_cast_fp16")]; + tensor input_391_cast_fp16 = mul(x = input_391_split_cast_fp16_0, y = input_391_split_1_sigmoid_cast_fp16)[name = tensor("input_391_cast_fp16")]; + tensor input_393_pad_type_0 = const()[name = tensor("input_393_pad_type_0"), val = tensor("custom")]; + tensor input_393_pad_0 = const()[name = tensor("input_393_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_393_groups_0 = const()[name = tensor("input_393_groups_0"), val = tensor(1024)]; + tensor input_393_strides_0 = const()[name = tensor("input_393_strides_0"), val = tensor([1, 1])]; + tensor input_393_dilations_0 = const()[name = tensor("input_393_dilations_0"), val = tensor([1, 1])]; + tensor const_296_to_fp16 = const()[name = tensor("const_296_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(746221504)))]; + tensor const_297_to_fp16 = const()[name = tensor("const_297_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(746240000)))]; + tensor input_395_cast_fp16 = conv(bias = const_297_to_fp16, dilations = input_393_dilations_0, groups = input_393_groups_0, pad = input_393_pad_0, pad_type = input_393_pad_type_0, strides = input_393_strides_0, weight = const_296_to_fp16, x = input_391_cast_fp16)[name = tensor("input_395_cast_fp16")]; + tensor input_397_cast_fp16 = silu(x = input_395_cast_fp16)[name = tensor("input_397_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 = const()[name = tensor("layers_14_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(746242112)))]; + 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, x = input_397_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_4311_to_fp16 = const()[name = tensor("op_4311_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_147_cast_fp16 = layer_norm(axes = out_147_axes_0, epsilon = var_4311_to_fp16, x = inputs_147_cast_fp16)[name = tensor("out_147_cast_fp16")]; + tensor input_399_gamma_0_to_fp16 = const()[name = tensor("input_399_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748339328)))]; + tensor input_399_beta_0_to_fp16 = const()[name = tensor("input_399_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748341440)))]; + tensor input_399_epsilon_0_to_fp16 = const()[name = tensor("input_399_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_399_cast_fp16 = batch_norm(beta = input_399_beta_0_to_fp16, epsilon = input_399_epsilon_0_to_fp16, gamma = input_399_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_147_cast_fp16)[name = tensor("input_399_cast_fp16")]; + tensor input_401_pad_type_0 = const()[name = tensor("input_401_pad_type_0"), val = tensor("valid")]; + tensor input_401_strides_0 = const()[name = tensor("input_401_strides_0"), val = tensor([1, 1])]; + tensor input_401_pad_0 = const()[name = tensor("input_401_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_401_dilations_0 = const()[name = tensor("input_401_dilations_0"), val = tensor([1, 1])]; + tensor input_401_groups_0 = const()[name = tensor("input_401_groups_0"), val = tensor(1)]; + tensor layers_14_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_14_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(748343552)))]; + tensor input_401_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_401_dilations_0, groups = input_401_groups_0, pad = input_401_pad_0, pad_type = input_401_pad_type_0, strides = input_401_strides_0, weight = layers_14_feed_forward2_fc1_weight_to_fp16, x = input_399_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor input_403_cast_fp16 = silu(x = input_401_cast_fp16)[name = tensor("input_403_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 var_4339_weight_0_to_fp16 = const()[name = tensor("op_4339_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(756732224)))]; + tensor var_4339_cast_fp16 = conv(bias = input_15_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 = var_4339_weight_0_to_fp16, x = input_403_cast_fp16)[name = tensor("op_4339_cast_fp16")]; + tensor inputs_149_cast_fp16 = add(x = inputs_147_cast_fp16, y = var_4339_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_4349_to_fp16 = const()[name = tensor("op_4349_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_149_cast_fp16 = layer_norm(axes = out_149_axes_0, epsilon = var_4349_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(765120896)))]; + 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(765123008)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_149_cast_fp16)[name = tensor("inputs_151_cast_fp16")]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor(3)]; + tensor out_151_axes_0 = const()[name = tensor("out_151_axes_0"), val = tensor([1])]; + tensor var_4394_to_fp16 = const()[name = tensor("op_4394_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_151_cast_fp16 = layer_norm(axes = out_151_axes_0, epsilon = var_4394_to_fp16, x = inputs_151_cast_fp16)[name = tensor("out_151_cast_fp16")]; + tensor input_405_gamma_0_to_fp16 = const()[name = tensor("input_405_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765125120)))]; + tensor input_405_beta_0_to_fp16 = const()[name = tensor("input_405_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765127232)))]; + tensor input_405_epsilon_0_to_fp16 = const()[name = tensor("input_405_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_405_cast_fp16 = batch_norm(beta = input_405_beta_0_to_fp16, epsilon = input_405_epsilon_0_to_fp16, gamma = input_405_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_151_cast_fp16)[name = tensor("input_405_cast_fp16")]; + tensor input_407_pad_type_0 = const()[name = tensor("input_407_pad_type_0"), val = tensor("valid")]; + tensor input_407_strides_0 = const()[name = tensor("input_407_strides_0"), val = tensor([1, 1])]; + tensor input_407_pad_0 = const()[name = tensor("input_407_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_407_dilations_0 = const()[name = tensor("input_407_dilations_0"), val = tensor([1, 1])]; + tensor input_407_groups_0 = const()[name = tensor("input_407_groups_0"), val = tensor(1)]; + tensor layers_15_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_15_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(765129344)))]; + tensor input_407_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_407_dilations_0, groups = input_407_groups_0, pad = input_407_pad_0, pad_type = input_407_pad_type_0, strides = input_407_strides_0, weight = layers_15_feed_forward1_fc1_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor input_409_cast_fp16 = silu(x = input_407_cast_fp16)[name = tensor("input_409_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 var_4422_weight_0_to_fp16 = const()[name = tensor("op_4422_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(773518016)))]; + tensor var_4422_cast_fp16 = conv(bias = input_15_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 = var_4422_weight_0_to_fp16, x = input_409_cast_fp16)[name = tensor("op_4422_cast_fp16")]; + tensor inputs_153_cast_fp16 = add(x = inputs_151_cast_fp16, y = var_4422_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_4432_to_fp16 = const()[name = tensor("op_4432_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_153_cast_fp16 = layer_norm(axes = out_153_axes_0, epsilon = var_4432_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(781906688)))]; + 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(781908800)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(781910912)))]; + tensor query_61_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(784008128)))]; + 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, 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 = const()[name = tensor("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(786105344)))]; + tensor value_31_cast_fp16 = conv(bias = input_15_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, x = obj_63_cast_fp16)[name = tensor("value_31_cast_fp16")]; + tensor var_4470_to_fp16 = const()[name = tensor("op_4470_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788202560)))]; + tensor query_63_cast_fp16 = add(x = query_61_cast_fp16, y = var_4470_to_fp16)[name = tensor("query_63_cast_fp16")]; + tensor var_4473_to_fp16 = const()[name = tensor("op_4473_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788204672)))]; + tensor q_with_bias_v_31_cast_fp16 = add(x = query_61_cast_fp16, y = var_4473_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 = const()[name = tensor("layers_15_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(788206784)))]; + tensor p_31_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_31_cast_fp16")]; + tensor var_4484 = const()[name = tensor("op_4484"), val = tensor([1, 8, 128, 188])]; + tensor var_4485_cast_fp16 = reshape(shape = var_4484, x = q_with_bias_v_31_cast_fp16)[name = tensor("op_4485_cast_fp16")]; + tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 8, 128, -1])]; + tensor var_4487_cast_fp16 = reshape(shape = var_4486, x = p_31_cast_fp16)[name = tensor("op_4487_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_4485_cast_fp16, y = var_4487_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_4496 = const()[name = tensor("op_4496"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_125_cast_fp16 = reshape(shape = var_4496, x = matrix_bd_123_cast_fp16)[name = tensor("matrix_bd_125_cast_fp16")]; + tensor var_4500_begin_0 = const()[name = tensor("op_4500_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4500_end_0 = const()[name = tensor("op_4500_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_4500_end_mask_0 = const()[name = tensor("op_4500_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4500_cast_fp16 = slice_by_index(begin = var_4500_begin_0, end = var_4500_end_0, end_mask = var_4500_end_mask_0, x = matrix_bd_125_cast_fp16)[name = tensor("op_4500_cast_fp16")]; + tensor var_4501 = const()[name = tensor("op_4501"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_127_cast_fp16 = reshape(shape = var_4501, x = var_4500_cast_fp16)[name = tensor("matrix_bd_127_cast_fp16")]; + tensor var_4506_begin_0 = const()[name = tensor("op_4506_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4506_end_0 = const()[name = tensor("op_4506_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_4506_end_mask_0 = const()[name = tensor("op_4506_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_4506_cast_fp16 = slice_by_index(begin = var_4506_begin_0, end = var_4506_end_0, end_mask = var_4506_end_mask_0, x = matrix_bd_127_cast_fp16)[name = tensor("op_4506_cast_fp16")]; + tensor var_4507_to_fp16 = const()[name = tensor("op_4507_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_31_cast_fp16 = mul(x = var_4506_cast_fp16, y = var_4507_to_fp16)[name = tensor("qk_mask_31_cast_fp16")]; + tensor var_4511 = const()[name = tensor("op_4511"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_31_cast_fp16 = reshape(shape = var_4511, x = query_63_cast_fp16)[name = tensor("mh_q_31_cast_fp16")]; + tensor var_4513_to_fp16 = const()[name = tensor("op_4513_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_4514_cast_fp16 = mul(x = mh_q_31_cast_fp16, y = var_4513_to_fp16)[name = tensor("op_4514_cast_fp16")]; + tensor var_4517 = const()[name = tensor("op_4517"), val = tensor([1, 8, 128, 188])]; + tensor var_4518_cast_fp16 = reshape(shape = var_4517, x = key_31_cast_fp16)[name = tensor("op_4518_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_4514_cast_fp16, y = var_4518_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_4522_cast_fp16 = softmax(axis = var_4363, x = mh_w_63_cast_fp16)[name = tensor("op_4522_cast_fp16")]; + tensor var_4523 = const()[name = tensor("op_4523"), val = tensor([1, 8, 128, 188])]; + tensor var_4524_cast_fp16 = reshape(shape = var_4523, x = value_31_cast_fp16)[name = tensor("op_4524_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_4524_cast_fp16, y = var_4522_cast_fp16)[name = tensor("attn_31_cast_fp16")]; + tensor var_4527 = const()[name = tensor("op_4527"), val = tensor([1, 1024, 1, 188])]; + tensor input_411_cast_fp16 = reshape(shape = var_4527, x = attn_31_cast_fp16)[name = tensor("input_411_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 = const()[name = tensor("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(790304000)))]; + tensor obj_65_cast_fp16 = conv(bias = input_15_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, x = input_411_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_4545_to_fp16 = const()[name = tensor("op_4545_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_155_cast_fp16 = layer_norm(axes = out_155_axes_0, epsilon = var_4545_to_fp16, x = inputs_155_cast_fp16)[name = tensor("out_155_cast_fp16")]; + tensor input_413_gamma_0_to_fp16 = const()[name = tensor("input_413_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(792401216)))]; + tensor input_413_beta_0_to_fp16 = const()[name = tensor("input_413_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(792403328)))]; + tensor input_413_epsilon_0_to_fp16 = const()[name = tensor("input_413_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_413_cast_fp16 = batch_norm(beta = input_413_beta_0_to_fp16, epsilon = input_413_epsilon_0_to_fp16, gamma = input_413_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_155_cast_fp16)[name = tensor("input_413_cast_fp16")]; + tensor input_415_pad_type_0 = const()[name = tensor("input_415_pad_type_0"), val = tensor("valid")]; + tensor input_415_strides_0 = const()[name = tensor("input_415_strides_0"), val = tensor([1, 1])]; + tensor input_415_pad_0 = const()[name = tensor("input_415_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_415_dilations_0 = const()[name = tensor("input_415_dilations_0"), val = tensor([1, 1])]; + tensor input_415_groups_0 = const()[name = tensor("input_415_groups_0"), val = tensor(1)]; + tensor layers_15_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_15_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(792405440)))]; + tensor input_415_cast_fp16 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = layers_15_conv_pointwise_conv1_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("input_415_cast_fp16")]; + tensor input_417_split_num_splits_0 = const()[name = tensor("input_417_split_num_splits_0"), val = tensor(2)]; + tensor input_417_split_axis_0 = const()[name = tensor("input_417_split_axis_0"), val = tensor(1)]; + tensor input_417_split_cast_fp16_0, tensor input_417_split_cast_fp16_1 = split(axis = input_417_split_axis_0, num_splits = input_417_split_num_splits_0, x = input_415_cast_fp16)[name = tensor("input_417_split_cast_fp16")]; + tensor input_417_split_1_sigmoid_cast_fp16 = sigmoid(x = input_417_split_cast_fp16_1)[name = tensor("input_417_split_1_sigmoid_cast_fp16")]; + tensor input_417_cast_fp16 = mul(x = input_417_split_cast_fp16_0, y = input_417_split_1_sigmoid_cast_fp16)[name = tensor("input_417_cast_fp16")]; + tensor input_419_pad_type_0 = const()[name = tensor("input_419_pad_type_0"), val = tensor("custom")]; + tensor input_419_pad_0 = const()[name = tensor("input_419_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_419_groups_0 = const()[name = tensor("input_419_groups_0"), val = tensor(1024)]; + tensor input_419_strides_0 = const()[name = tensor("input_419_strides_0"), val = tensor([1, 1])]; + tensor input_419_dilations_0 = const()[name = tensor("input_419_dilations_0"), val = tensor([1, 1])]; + tensor const_298_to_fp16 = const()[name = tensor("const_298_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(796599808)))]; + tensor const_299_to_fp16 = const()[name = tensor("const_299_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(796618304)))]; + tensor input_421_cast_fp16 = conv(bias = const_299_to_fp16, dilations = input_419_dilations_0, groups = input_419_groups_0, pad = input_419_pad_0, pad_type = input_419_pad_type_0, strides = input_419_strides_0, weight = const_298_to_fp16, x = input_417_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_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 = const()[name = tensor("layers_15_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(796620416)))]; + 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, x = input_423_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_4593_to_fp16 = const()[name = tensor("op_4593_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_157_cast_fp16 = layer_norm(axes = out_157_axes_0, epsilon = var_4593_to_fp16, x = inputs_157_cast_fp16)[name = tensor("out_157_cast_fp16")]; + tensor input_425_gamma_0_to_fp16 = const()[name = tensor("input_425_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(798717632)))]; + tensor input_425_beta_0_to_fp16 = const()[name = tensor("input_425_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(798719744)))]; + tensor input_425_epsilon_0_to_fp16 = const()[name = tensor("input_425_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_425_cast_fp16 = batch_norm(beta = input_425_beta_0_to_fp16, epsilon = input_425_epsilon_0_to_fp16, gamma = input_425_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_157_cast_fp16)[name = tensor("input_425_cast_fp16")]; + tensor input_427_pad_type_0 = const()[name = tensor("input_427_pad_type_0"), val = tensor("valid")]; + tensor input_427_strides_0 = const()[name = tensor("input_427_strides_0"), val = tensor([1, 1])]; + tensor input_427_pad_0 = const()[name = tensor("input_427_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_427_dilations_0 = const()[name = tensor("input_427_dilations_0"), val = tensor([1, 1])]; + tensor input_427_groups_0 = const()[name = tensor("input_427_groups_0"), val = tensor(1)]; + tensor layers_15_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_15_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(798721856)))]; + tensor input_427_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_427_dilations_0, groups = input_427_groups_0, pad = input_427_pad_0, pad_type = input_427_pad_type_0, strides = input_427_strides_0, weight = layers_15_feed_forward2_fc1_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("input_427_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = input_427_cast_fp16)[name = tensor("input_429_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 var_4621_weight_0_to_fp16 = const()[name = tensor("op_4621_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(807110528)))]; + tensor var_4621_cast_fp16 = conv(bias = input_15_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 = var_4621_weight_0_to_fp16, x = input_429_cast_fp16)[name = tensor("op_4621_cast_fp16")]; + tensor inputs_159_cast_fp16 = add(x = inputs_157_cast_fp16, y = var_4621_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_4631_to_fp16 = const()[name = tensor("op_4631_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_159_cast_fp16 = layer_norm(axes = out_159_axes_0, epsilon = var_4631_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(815499200)))]; + 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(815501312)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_159_cast_fp16)[name = tensor("inputs_161_cast_fp16")]; + tensor var_4645 = const()[name = tensor("op_4645"), val = tensor(3)]; + tensor out_161_axes_0 = const()[name = tensor("out_161_axes_0"), val = tensor([1])]; + tensor var_4676_to_fp16 = const()[name = tensor("op_4676_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_161_cast_fp16 = layer_norm(axes = out_161_axes_0, epsilon = var_4676_to_fp16, x = inputs_161_cast_fp16)[name = tensor("out_161_cast_fp16")]; + tensor input_431_gamma_0_to_fp16 = const()[name = tensor("input_431_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(815503424)))]; + tensor input_431_beta_0_to_fp16 = const()[name = tensor("input_431_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(815505536)))]; + tensor input_431_epsilon_0_to_fp16 = const()[name = tensor("input_431_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_431_cast_fp16 = batch_norm(beta = input_431_beta_0_to_fp16, epsilon = input_431_epsilon_0_to_fp16, gamma = input_431_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_161_cast_fp16)[name = tensor("input_431_cast_fp16")]; + tensor input_433_pad_type_0 = const()[name = tensor("input_433_pad_type_0"), val = tensor("valid")]; + tensor input_433_strides_0 = const()[name = tensor("input_433_strides_0"), val = tensor([1, 1])]; + tensor input_433_pad_0 = const()[name = tensor("input_433_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_433_dilations_0 = const()[name = tensor("input_433_dilations_0"), val = tensor([1, 1])]; + tensor input_433_groups_0 = const()[name = tensor("input_433_groups_0"), val = tensor(1)]; + tensor layers_16_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_16_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(815507648)))]; + tensor input_433_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_433_dilations_0, groups = input_433_groups_0, pad = input_433_pad_0, pad_type = input_433_pad_type_0, strides = input_433_strides_0, weight = layers_16_feed_forward1_fc1_weight_to_fp16, x = input_431_cast_fp16)[name = tensor("input_433_cast_fp16")]; + tensor input_435_cast_fp16 = silu(x = input_433_cast_fp16)[name = tensor("input_435_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 var_4704_weight_0_to_fp16 = const()[name = tensor("op_4704_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(823896320)))]; + tensor var_4704_cast_fp16 = conv(bias = input_15_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 = var_4704_weight_0_to_fp16, x = input_435_cast_fp16)[name = tensor("op_4704_cast_fp16")]; + tensor inputs_163_cast_fp16 = add(x = inputs_161_cast_fp16, y = var_4704_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_4714_to_fp16 = const()[name = tensor("op_4714_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_163_cast_fp16 = layer_norm(axes = out_163_axes_0, epsilon = var_4714_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(832284992)))]; + 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(832287104)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(832289216)))]; + tensor query_65_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(834386432)))]; + 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, 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 = const()[name = tensor("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(836483648)))]; + tensor value_33_cast_fp16 = conv(bias = input_15_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, x = obj_67_cast_fp16)[name = tensor("value_33_cast_fp16")]; + tensor var_4752_to_fp16 = const()[name = tensor("op_4752_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(838580864)))]; + tensor query_67_cast_fp16 = add(x = query_65_cast_fp16, y = var_4752_to_fp16)[name = tensor("query_67_cast_fp16")]; + tensor var_4755_to_fp16 = const()[name = tensor("op_4755_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(838582976)))]; + tensor q_with_bias_v_33_cast_fp16 = add(x = query_65_cast_fp16, y = var_4755_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 = const()[name = tensor("layers_16_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(838585088)))]; + tensor p_33_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_33_cast_fp16")]; + tensor var_4766 = const()[name = tensor("op_4766"), val = tensor([1, 8, 128, 188])]; + tensor var_4767_cast_fp16 = reshape(shape = var_4766, x = q_with_bias_v_33_cast_fp16)[name = tensor("op_4767_cast_fp16")]; + tensor var_4768 = const()[name = tensor("op_4768"), val = tensor([1, 8, 128, -1])]; + tensor var_4769_cast_fp16 = reshape(shape = var_4768, x = p_33_cast_fp16)[name = tensor("op_4769_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_4767_cast_fp16, y = var_4769_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_4778 = const()[name = tensor("op_4778"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_133_cast_fp16 = reshape(shape = var_4778, x = matrix_bd_131_cast_fp16)[name = tensor("matrix_bd_133_cast_fp16")]; + tensor var_4782_begin_0 = const()[name = tensor("op_4782_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_4782_end_0 = const()[name = tensor("op_4782_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_4782_end_mask_0 = const()[name = tensor("op_4782_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_4782_cast_fp16 = slice_by_index(begin = var_4782_begin_0, end = var_4782_end_0, end_mask = var_4782_end_mask_0, x = matrix_bd_133_cast_fp16)[name = tensor("op_4782_cast_fp16")]; + tensor var_4783 = const()[name = tensor("op_4783"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_135_cast_fp16 = reshape(shape = var_4783, x = var_4782_cast_fp16)[name = tensor("matrix_bd_135_cast_fp16")]; + tensor var_4788_begin_0 = const()[name = tensor("op_4788_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_4788_end_0 = const()[name = tensor("op_4788_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_4788_end_mask_0 = const()[name = tensor("op_4788_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_4788_cast_fp16 = slice_by_index(begin = var_4788_begin_0, end = var_4788_end_0, end_mask = var_4788_end_mask_0, x = matrix_bd_135_cast_fp16)[name = tensor("op_4788_cast_fp16")]; + tensor var_4789_to_fp16 = const()[name = tensor("op_4789_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_33_cast_fp16 = mul(x = var_4788_cast_fp16, y = var_4789_to_fp16)[name = tensor("qk_mask_33_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_33_cast_fp16 = reshape(shape = var_4793, x = query_67_cast_fp16)[name = tensor("mh_q_33_cast_fp16")]; + tensor var_4795_to_fp16 = const()[name = tensor("op_4795_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_4796_cast_fp16 = mul(x = mh_q_33_cast_fp16, y = var_4795_to_fp16)[name = tensor("op_4796_cast_fp16")]; + tensor var_4799 = const()[name = tensor("op_4799"), val = tensor([1, 8, 128, 188])]; + tensor var_4800_cast_fp16 = reshape(shape = var_4799, x = key_33_cast_fp16)[name = tensor("op_4800_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_4796_cast_fp16, y = var_4800_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_4804_cast_fp16 = softmax(axis = var_4645, x = mh_w_67_cast_fp16)[name = tensor("op_4804_cast_fp16")]; + tensor var_4805 = const()[name = tensor("op_4805"), val = tensor([1, 8, 128, 188])]; + tensor var_4806_cast_fp16 = reshape(shape = var_4805, x = value_33_cast_fp16)[name = tensor("op_4806_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_4806_cast_fp16, y = var_4804_cast_fp16)[name = tensor("attn_33_cast_fp16")]; + tensor var_4809 = const()[name = tensor("op_4809"), val = tensor([1, 1024, 1, 188])]; + tensor input_437_cast_fp16 = reshape(shape = var_4809, x = attn_33_cast_fp16)[name = tensor("input_437_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 = const()[name = tensor("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(840682304)))]; + tensor obj_69_cast_fp16 = conv(bias = input_15_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, x = input_437_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_4827_to_fp16 = const()[name = tensor("op_4827_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_165_cast_fp16 = layer_norm(axes = out_165_axes_0, epsilon = var_4827_to_fp16, x = inputs_165_cast_fp16)[name = tensor("out_165_cast_fp16")]; + tensor input_439_gamma_0_to_fp16 = const()[name = tensor("input_439_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842779520)))]; + tensor input_439_beta_0_to_fp16 = const()[name = tensor("input_439_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842781632)))]; + tensor input_439_epsilon_0_to_fp16 = const()[name = tensor("input_439_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_439_cast_fp16 = batch_norm(beta = input_439_beta_0_to_fp16, epsilon = input_439_epsilon_0_to_fp16, gamma = input_439_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_165_cast_fp16)[name = tensor("input_439_cast_fp16")]; + tensor input_441_pad_type_0 = const()[name = tensor("input_441_pad_type_0"), val = tensor("valid")]; + tensor input_441_strides_0 = const()[name = tensor("input_441_strides_0"), val = tensor([1, 1])]; + tensor input_441_pad_0 = const()[name = tensor("input_441_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_441_dilations_0 = const()[name = tensor("input_441_dilations_0"), val = tensor([1, 1])]; + tensor input_441_groups_0 = const()[name = tensor("input_441_groups_0"), val = tensor(1)]; + tensor layers_16_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_16_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(842783744)))]; + tensor input_441_cast_fp16 = conv(dilations = input_441_dilations_0, groups = input_441_groups_0, pad = input_441_pad_0, pad_type = input_441_pad_type_0, strides = input_441_strides_0, weight = layers_16_conv_pointwise_conv1_weight_to_fp16, x = input_439_cast_fp16)[name = tensor("input_441_cast_fp16")]; + tensor input_443_split_num_splits_0 = const()[name = tensor("input_443_split_num_splits_0"), val = tensor(2)]; + tensor input_443_split_axis_0 = const()[name = tensor("input_443_split_axis_0"), val = tensor(1)]; + tensor input_443_split_cast_fp16_0, tensor input_443_split_cast_fp16_1 = split(axis = input_443_split_axis_0, num_splits = input_443_split_num_splits_0, x = input_441_cast_fp16)[name = tensor("input_443_split_cast_fp16")]; + tensor input_443_split_1_sigmoid_cast_fp16 = sigmoid(x = input_443_split_cast_fp16_1)[name = tensor("input_443_split_1_sigmoid_cast_fp16")]; + tensor input_443_cast_fp16 = mul(x = input_443_split_cast_fp16_0, y = input_443_split_1_sigmoid_cast_fp16)[name = tensor("input_443_cast_fp16")]; + tensor input_445_pad_type_0 = const()[name = tensor("input_445_pad_type_0"), val = tensor("custom")]; + tensor input_445_pad_0 = const()[name = tensor("input_445_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_445_groups_0 = const()[name = tensor("input_445_groups_0"), val = tensor(1024)]; + tensor input_445_strides_0 = const()[name = tensor("input_445_strides_0"), val = tensor([1, 1])]; + tensor input_445_dilations_0 = const()[name = tensor("input_445_dilations_0"), val = tensor([1, 1])]; + tensor const_300_to_fp16 = const()[name = tensor("const_300_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846978112)))]; + tensor const_301_to_fp16 = const()[name = tensor("const_301_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846996608)))]; + tensor input_447_cast_fp16 = conv(bias = const_301_to_fp16, dilations = input_445_dilations_0, groups = input_445_groups_0, pad = input_445_pad_0, pad_type = input_445_pad_type_0, strides = input_445_strides_0, weight = const_300_to_fp16, x = input_443_cast_fp16)[name = tensor("input_447_cast_fp16")]; + tensor input_449_cast_fp16 = silu(x = input_447_cast_fp16)[name = tensor("input_449_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 = const()[name = tensor("layers_16_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(846998720)))]; + 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, x = input_449_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_4875_to_fp16 = const()[name = tensor("op_4875_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_167_cast_fp16 = layer_norm(axes = out_167_axes_0, epsilon = var_4875_to_fp16, x = inputs_167_cast_fp16)[name = tensor("out_167_cast_fp16")]; + tensor input_451_gamma_0_to_fp16 = const()[name = tensor("input_451_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(849095936)))]; + tensor input_451_beta_0_to_fp16 = const()[name = tensor("input_451_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(849098048)))]; + tensor input_451_epsilon_0_to_fp16 = const()[name = tensor("input_451_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_451_cast_fp16 = batch_norm(beta = input_451_beta_0_to_fp16, epsilon = input_451_epsilon_0_to_fp16, gamma = input_451_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_167_cast_fp16)[name = tensor("input_451_cast_fp16")]; + tensor input_453_pad_type_0 = const()[name = tensor("input_453_pad_type_0"), val = tensor("valid")]; + tensor input_453_strides_0 = const()[name = tensor("input_453_strides_0"), val = tensor([1, 1])]; + tensor input_453_pad_0 = const()[name = tensor("input_453_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_453_dilations_0 = const()[name = tensor("input_453_dilations_0"), val = tensor([1, 1])]; + tensor input_453_groups_0 = const()[name = tensor("input_453_groups_0"), val = tensor(1)]; + tensor layers_16_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_16_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(849100160)))]; + tensor input_453_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_453_dilations_0, groups = input_453_groups_0, pad = input_453_pad_0, pad_type = input_453_pad_type_0, strides = input_453_strides_0, weight = layers_16_feed_forward2_fc1_weight_to_fp16, x = input_451_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor input_455_cast_fp16 = silu(x = input_453_cast_fp16)[name = tensor("input_455_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 var_4903_weight_0_to_fp16 = const()[name = tensor("op_4903_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(857488832)))]; + tensor var_4903_cast_fp16 = conv(bias = input_15_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 = var_4903_weight_0_to_fp16, x = input_455_cast_fp16)[name = tensor("op_4903_cast_fp16")]; + tensor inputs_169_cast_fp16 = add(x = inputs_167_cast_fp16, y = var_4903_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_4913_to_fp16 = const()[name = tensor("op_4913_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_169_cast_fp16 = layer_norm(axes = out_169_axes_0, epsilon = var_4913_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(865877504)))]; + 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(865879616)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_169_cast_fp16)[name = tensor("inputs_171_cast_fp16")]; + tensor var_4927 = const()[name = tensor("op_4927"), val = tensor(3)]; + tensor out_171_axes_0 = const()[name = tensor("out_171_axes_0"), val = tensor([1])]; + tensor var_4958_to_fp16 = const()[name = tensor("op_4958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_171_cast_fp16 = layer_norm(axes = out_171_axes_0, epsilon = var_4958_to_fp16, x = inputs_171_cast_fp16)[name = tensor("out_171_cast_fp16")]; + tensor input_457_gamma_0_to_fp16 = const()[name = tensor("input_457_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(865881728)))]; + tensor input_457_beta_0_to_fp16 = const()[name = tensor("input_457_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(865883840)))]; + tensor input_457_epsilon_0_to_fp16 = const()[name = tensor("input_457_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_457_cast_fp16 = batch_norm(beta = input_457_beta_0_to_fp16, epsilon = input_457_epsilon_0_to_fp16, gamma = input_457_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_171_cast_fp16)[name = tensor("input_457_cast_fp16")]; + tensor input_459_pad_type_0 = const()[name = tensor("input_459_pad_type_0"), val = tensor("valid")]; + tensor input_459_strides_0 = const()[name = tensor("input_459_strides_0"), val = tensor([1, 1])]; + tensor input_459_pad_0 = const()[name = tensor("input_459_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_459_dilations_0 = const()[name = tensor("input_459_dilations_0"), val = tensor([1, 1])]; + tensor input_459_groups_0 = const()[name = tensor("input_459_groups_0"), val = tensor(1)]; + tensor layers_17_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_17_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(865885952)))]; + tensor input_459_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_459_dilations_0, groups = input_459_groups_0, pad = input_459_pad_0, pad_type = input_459_pad_type_0, strides = input_459_strides_0, weight = layers_17_feed_forward1_fc1_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("input_459_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = input_459_cast_fp16)[name = tensor("input_461_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 var_4986_weight_0_to_fp16 = const()[name = tensor("op_4986_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(874274624)))]; + tensor var_4986_cast_fp16 = conv(bias = input_15_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 = var_4986_weight_0_to_fp16, x = input_461_cast_fp16)[name = tensor("op_4986_cast_fp16")]; + tensor inputs_173_cast_fp16 = add(x = inputs_171_cast_fp16, y = var_4986_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_4996_to_fp16 = const()[name = tensor("op_4996_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_173_cast_fp16 = layer_norm(axes = out_173_axes_0, epsilon = var_4996_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(882663296)))]; + 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(882665408)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(882667520)))]; + tensor query_69_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(884764736)))]; + 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, 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 = const()[name = tensor("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(886861952)))]; + tensor value_35_cast_fp16 = conv(bias = input_15_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, x = obj_71_cast_fp16)[name = tensor("value_35_cast_fp16")]; + tensor var_5034_to_fp16 = const()[name = tensor("op_5034_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888959168)))]; + tensor query_71_cast_fp16 = add(x = query_69_cast_fp16, y = var_5034_to_fp16)[name = tensor("query_71_cast_fp16")]; + tensor var_5037_to_fp16 = const()[name = tensor("op_5037_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888961280)))]; + tensor q_with_bias_v_35_cast_fp16 = add(x = query_69_cast_fp16, y = var_5037_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 = const()[name = tensor("layers_17_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(888963392)))]; + tensor p_35_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_35_cast_fp16")]; + tensor var_5048 = const()[name = tensor("op_5048"), val = tensor([1, 8, 128, 188])]; + tensor var_5049_cast_fp16 = reshape(shape = var_5048, x = q_with_bias_v_35_cast_fp16)[name = tensor("op_5049_cast_fp16")]; + tensor var_5050 = const()[name = tensor("op_5050"), val = tensor([1, 8, 128, -1])]; + tensor var_5051_cast_fp16 = reshape(shape = var_5050, x = p_35_cast_fp16)[name = tensor("op_5051_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_5049_cast_fp16, y = var_5051_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_5060 = const()[name = tensor("op_5060"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_141_cast_fp16 = reshape(shape = var_5060, x = matrix_bd_139_cast_fp16)[name = tensor("matrix_bd_141_cast_fp16")]; + tensor var_5064_begin_0 = const()[name = tensor("op_5064_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5064_end_0 = const()[name = tensor("op_5064_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5064_end_mask_0 = const()[name = tensor("op_5064_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5064_cast_fp16 = slice_by_index(begin = var_5064_begin_0, end = var_5064_end_0, end_mask = var_5064_end_mask_0, x = matrix_bd_141_cast_fp16)[name = tensor("op_5064_cast_fp16")]; + tensor var_5065 = const()[name = tensor("op_5065"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_143_cast_fp16 = reshape(shape = var_5065, x = var_5064_cast_fp16)[name = tensor("matrix_bd_143_cast_fp16")]; + tensor var_5070_begin_0 = const()[name = tensor("op_5070_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5070_end_0 = const()[name = tensor("op_5070_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5070_end_mask_0 = const()[name = tensor("op_5070_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5070_cast_fp16 = slice_by_index(begin = var_5070_begin_0, end = var_5070_end_0, end_mask = var_5070_end_mask_0, x = matrix_bd_143_cast_fp16)[name = tensor("op_5070_cast_fp16")]; + tensor var_5071_to_fp16 = const()[name = tensor("op_5071_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_35_cast_fp16 = mul(x = var_5070_cast_fp16, y = var_5071_to_fp16)[name = tensor("qk_mask_35_cast_fp16")]; + tensor var_5075 = const()[name = tensor("op_5075"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_35_cast_fp16 = reshape(shape = var_5075, x = query_71_cast_fp16)[name = tensor("mh_q_35_cast_fp16")]; + tensor var_5077_to_fp16 = const()[name = tensor("op_5077_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5078_cast_fp16 = mul(x = mh_q_35_cast_fp16, y = var_5077_to_fp16)[name = tensor("op_5078_cast_fp16")]; + tensor var_5081 = const()[name = tensor("op_5081"), val = tensor([1, 8, 128, 188])]; + tensor var_5082_cast_fp16 = reshape(shape = var_5081, x = key_35_cast_fp16)[name = tensor("op_5082_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_5078_cast_fp16, y = var_5082_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_5086_cast_fp16 = softmax(axis = var_4927, x = mh_w_71_cast_fp16)[name = tensor("op_5086_cast_fp16")]; + tensor var_5087 = const()[name = tensor("op_5087"), val = tensor([1, 8, 128, 188])]; + tensor var_5088_cast_fp16 = reshape(shape = var_5087, x = value_35_cast_fp16)[name = tensor("op_5088_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_5088_cast_fp16, y = var_5086_cast_fp16)[name = tensor("attn_35_cast_fp16")]; + tensor var_5091 = const()[name = tensor("op_5091"), val = tensor([1, 1024, 1, 188])]; + tensor input_463_cast_fp16 = reshape(shape = var_5091, x = attn_35_cast_fp16)[name = tensor("input_463_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 = const()[name = tensor("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(891060608)))]; + tensor obj_73_cast_fp16 = conv(bias = input_15_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, x = input_463_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_5109_to_fp16 = const()[name = tensor("op_5109_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_175_cast_fp16 = layer_norm(axes = out_175_axes_0, epsilon = var_5109_to_fp16, x = inputs_175_cast_fp16)[name = tensor("out_175_cast_fp16")]; + tensor input_465_gamma_0_to_fp16 = const()[name = tensor("input_465_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893157824)))]; + tensor input_465_beta_0_to_fp16 = const()[name = tensor("input_465_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893159936)))]; + tensor input_465_epsilon_0_to_fp16 = const()[name = tensor("input_465_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_465_cast_fp16 = batch_norm(beta = input_465_beta_0_to_fp16, epsilon = input_465_epsilon_0_to_fp16, gamma = input_465_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_175_cast_fp16)[name = tensor("input_465_cast_fp16")]; + tensor input_467_pad_type_0 = const()[name = tensor("input_467_pad_type_0"), val = tensor("valid")]; + tensor input_467_strides_0 = const()[name = tensor("input_467_strides_0"), val = tensor([1, 1])]; + tensor input_467_pad_0 = const()[name = tensor("input_467_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_467_dilations_0 = const()[name = tensor("input_467_dilations_0"), val = tensor([1, 1])]; + tensor input_467_groups_0 = const()[name = tensor("input_467_groups_0"), val = tensor(1)]; + tensor layers_17_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_17_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(893162048)))]; + tensor input_467_cast_fp16 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = layers_17_conv_pointwise_conv1_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("input_467_cast_fp16")]; + tensor input_469_split_num_splits_0 = const()[name = tensor("input_469_split_num_splits_0"), val = tensor(2)]; + tensor input_469_split_axis_0 = const()[name = tensor("input_469_split_axis_0"), val = tensor(1)]; + tensor input_469_split_cast_fp16_0, tensor input_469_split_cast_fp16_1 = split(axis = input_469_split_axis_0, num_splits = input_469_split_num_splits_0, x = input_467_cast_fp16)[name = tensor("input_469_split_cast_fp16")]; + tensor input_469_split_1_sigmoid_cast_fp16 = sigmoid(x = input_469_split_cast_fp16_1)[name = tensor("input_469_split_1_sigmoid_cast_fp16")]; + tensor input_469_cast_fp16 = mul(x = input_469_split_cast_fp16_0, y = input_469_split_1_sigmoid_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor input_471_pad_type_0 = const()[name = tensor("input_471_pad_type_0"), val = tensor("custom")]; + tensor input_471_pad_0 = const()[name = tensor("input_471_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_471_groups_0 = const()[name = tensor("input_471_groups_0"), val = tensor(1024)]; + tensor input_471_strides_0 = const()[name = tensor("input_471_strides_0"), val = tensor([1, 1])]; + tensor input_471_dilations_0 = const()[name = tensor("input_471_dilations_0"), val = tensor([1, 1])]; + tensor const_302_to_fp16 = const()[name = tensor("const_302_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(897356416)))]; + tensor const_303_to_fp16 = const()[name = tensor("const_303_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(897374912)))]; + tensor input_473_cast_fp16 = conv(bias = const_303_to_fp16, dilations = input_471_dilations_0, groups = input_471_groups_0, pad = input_471_pad_0, pad_type = input_471_pad_type_0, strides = input_471_strides_0, weight = const_302_to_fp16, x = input_469_cast_fp16)[name = tensor("input_473_cast_fp16")]; + tensor input_475_cast_fp16 = silu(x = input_473_cast_fp16)[name = tensor("input_475_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 = const()[name = tensor("layers_17_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(897377024)))]; + 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, x = input_475_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_5157_to_fp16 = const()[name = tensor("op_5157_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_177_cast_fp16 = layer_norm(axes = out_177_axes_0, epsilon = var_5157_to_fp16, x = inputs_177_cast_fp16)[name = tensor("out_177_cast_fp16")]; + tensor input_477_gamma_0_to_fp16 = const()[name = tensor("input_477_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(899474240)))]; + tensor input_477_beta_0_to_fp16 = const()[name = tensor("input_477_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(899476352)))]; + tensor input_477_epsilon_0_to_fp16 = const()[name = tensor("input_477_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_477_cast_fp16 = batch_norm(beta = input_477_beta_0_to_fp16, epsilon = input_477_epsilon_0_to_fp16, gamma = input_477_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_177_cast_fp16)[name = tensor("input_477_cast_fp16")]; + tensor input_479_pad_type_0 = const()[name = tensor("input_479_pad_type_0"), val = tensor("valid")]; + tensor input_479_strides_0 = const()[name = tensor("input_479_strides_0"), val = tensor([1, 1])]; + tensor input_479_pad_0 = const()[name = tensor("input_479_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_479_dilations_0 = const()[name = tensor("input_479_dilations_0"), val = tensor([1, 1])]; + tensor input_479_groups_0 = const()[name = tensor("input_479_groups_0"), val = tensor(1)]; + tensor layers_17_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_17_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(899478464)))]; + tensor input_479_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_479_dilations_0, groups = input_479_groups_0, pad = input_479_pad_0, pad_type = input_479_pad_type_0, strides = input_479_strides_0, weight = layers_17_feed_forward2_fc1_weight_to_fp16, x = input_477_cast_fp16)[name = tensor("input_479_cast_fp16")]; + tensor input_481_cast_fp16 = silu(x = input_479_cast_fp16)[name = tensor("input_481_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 var_5185_weight_0_to_fp16 = const()[name = tensor("op_5185_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(907867136)))]; + tensor var_5185_cast_fp16 = conv(bias = input_15_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 = var_5185_weight_0_to_fp16, x = input_481_cast_fp16)[name = tensor("op_5185_cast_fp16")]; + tensor inputs_179_cast_fp16 = add(x = inputs_177_cast_fp16, y = var_5185_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_5195_to_fp16 = const()[name = tensor("op_5195_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_179_cast_fp16 = layer_norm(axes = out_179_axes_0, epsilon = var_5195_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(916255808)))]; + 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(916257920)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_179_cast_fp16)[name = tensor("inputs_181_cast_fp16")]; + tensor var_5209 = const()[name = tensor("op_5209"), val = tensor(3)]; + tensor out_181_axes_0 = const()[name = tensor("out_181_axes_0"), val = tensor([1])]; + tensor var_5240_to_fp16 = const()[name = tensor("op_5240_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_181_cast_fp16 = layer_norm(axes = out_181_axes_0, epsilon = var_5240_to_fp16, x = inputs_181_cast_fp16)[name = tensor("out_181_cast_fp16")]; + tensor input_483_gamma_0_to_fp16 = const()[name = tensor("input_483_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(916260032)))]; + tensor input_483_beta_0_to_fp16 = const()[name = tensor("input_483_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(916262144)))]; + tensor input_483_epsilon_0_to_fp16 = const()[name = tensor("input_483_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_483_cast_fp16 = batch_norm(beta = input_483_beta_0_to_fp16, epsilon = input_483_epsilon_0_to_fp16, gamma = input_483_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_181_cast_fp16)[name = tensor("input_483_cast_fp16")]; + tensor input_485_pad_type_0 = const()[name = tensor("input_485_pad_type_0"), val = tensor("valid")]; + tensor input_485_strides_0 = const()[name = tensor("input_485_strides_0"), val = tensor([1, 1])]; + tensor input_485_pad_0 = const()[name = tensor("input_485_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_485_dilations_0 = const()[name = tensor("input_485_dilations_0"), val = tensor([1, 1])]; + tensor input_485_groups_0 = const()[name = tensor("input_485_groups_0"), val = tensor(1)]; + tensor layers_18_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_18_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(916264256)))]; + tensor input_485_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_485_dilations_0, groups = input_485_groups_0, pad = input_485_pad_0, pad_type = input_485_pad_type_0, strides = input_485_strides_0, weight = layers_18_feed_forward1_fc1_weight_to_fp16, x = input_483_cast_fp16)[name = tensor("input_485_cast_fp16")]; + tensor input_487_cast_fp16 = silu(x = input_485_cast_fp16)[name = tensor("input_487_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 var_5268_weight_0_to_fp16 = const()[name = tensor("op_5268_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(924652928)))]; + tensor var_5268_cast_fp16 = conv(bias = input_15_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 = var_5268_weight_0_to_fp16, x = input_487_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor inputs_183_cast_fp16 = add(x = inputs_181_cast_fp16, y = var_5268_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_5278_to_fp16 = const()[name = tensor("op_5278_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_183_cast_fp16 = layer_norm(axes = out_183_axes_0, epsilon = var_5278_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(933041600)))]; + 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(933043712)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(933045824)))]; + tensor query_73_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(935143040)))]; + 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, 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 = const()[name = tensor("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(937240256)))]; + tensor value_37_cast_fp16 = conv(bias = input_15_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, x = obj_75_cast_fp16)[name = tensor("value_37_cast_fp16")]; + tensor var_5316_to_fp16 = const()[name = tensor("op_5316_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(939337472)))]; + tensor query_75_cast_fp16 = add(x = query_73_cast_fp16, y = var_5316_to_fp16)[name = tensor("query_75_cast_fp16")]; + tensor var_5319_to_fp16 = const()[name = tensor("op_5319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(939339584)))]; + tensor q_with_bias_v_37_cast_fp16 = add(x = query_73_cast_fp16, y = var_5319_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 = const()[name = tensor("layers_18_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(939341696)))]; + tensor p_37_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_37_cast_fp16")]; + tensor var_5330 = const()[name = tensor("op_5330"), val = tensor([1, 8, 128, 188])]; + tensor var_5331_cast_fp16 = reshape(shape = var_5330, x = q_with_bias_v_37_cast_fp16)[name = tensor("op_5331_cast_fp16")]; + tensor var_5332 = const()[name = tensor("op_5332"), val = tensor([1, 8, 128, -1])]; + tensor var_5333_cast_fp16 = reshape(shape = var_5332, x = p_37_cast_fp16)[name = tensor("op_5333_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_5331_cast_fp16, y = var_5333_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_5342 = const()[name = tensor("op_5342"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_149_cast_fp16 = reshape(shape = var_5342, x = matrix_bd_147_cast_fp16)[name = tensor("matrix_bd_149_cast_fp16")]; + tensor var_5346_begin_0 = const()[name = tensor("op_5346_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5346_end_0 = const()[name = tensor("op_5346_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5346_end_mask_0 = const()[name = tensor("op_5346_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5346_cast_fp16 = slice_by_index(begin = var_5346_begin_0, end = var_5346_end_0, end_mask = var_5346_end_mask_0, x = matrix_bd_149_cast_fp16)[name = tensor("op_5346_cast_fp16")]; + tensor var_5347 = const()[name = tensor("op_5347"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_151_cast_fp16 = reshape(shape = var_5347, x = var_5346_cast_fp16)[name = tensor("matrix_bd_151_cast_fp16")]; + tensor var_5352_begin_0 = const()[name = tensor("op_5352_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5352_end_0 = const()[name = tensor("op_5352_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5352_end_mask_0 = const()[name = tensor("op_5352_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5352_cast_fp16 = slice_by_index(begin = var_5352_begin_0, end = var_5352_end_0, end_mask = var_5352_end_mask_0, x = matrix_bd_151_cast_fp16)[name = tensor("op_5352_cast_fp16")]; + tensor var_5353_to_fp16 = const()[name = tensor("op_5353_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_37_cast_fp16 = mul(x = var_5352_cast_fp16, y = var_5353_to_fp16)[name = tensor("qk_mask_37_cast_fp16")]; + tensor var_5357 = const()[name = tensor("op_5357"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_37_cast_fp16 = reshape(shape = var_5357, x = query_75_cast_fp16)[name = tensor("mh_q_37_cast_fp16")]; + tensor var_5359_to_fp16 = const()[name = tensor("op_5359_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5360_cast_fp16 = mul(x = mh_q_37_cast_fp16, y = var_5359_to_fp16)[name = tensor("op_5360_cast_fp16")]; + tensor var_5363 = const()[name = tensor("op_5363"), val = tensor([1, 8, 128, 188])]; + tensor var_5364_cast_fp16 = reshape(shape = var_5363, x = key_37_cast_fp16)[name = tensor("op_5364_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_5360_cast_fp16, y = var_5364_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_5368_cast_fp16 = softmax(axis = var_5209, x = mh_w_75_cast_fp16)[name = tensor("op_5368_cast_fp16")]; + tensor var_5369 = const()[name = tensor("op_5369"), val = tensor([1, 8, 128, 188])]; + tensor var_5370_cast_fp16 = reshape(shape = var_5369, x = value_37_cast_fp16)[name = tensor("op_5370_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_5370_cast_fp16, y = var_5368_cast_fp16)[name = tensor("attn_37_cast_fp16")]; + tensor var_5373 = const()[name = tensor("op_5373"), val = tensor([1, 1024, 1, 188])]; + tensor input_489_cast_fp16 = reshape(shape = var_5373, x = attn_37_cast_fp16)[name = tensor("input_489_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 = const()[name = tensor("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(941438912)))]; + tensor obj_77_cast_fp16 = conv(bias = input_15_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, x = input_489_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_5391_to_fp16 = const()[name = tensor("op_5391_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_185_cast_fp16 = layer_norm(axes = out_185_axes_0, epsilon = var_5391_to_fp16, x = inputs_185_cast_fp16)[name = tensor("out_185_cast_fp16")]; + tensor input_491_gamma_0_to_fp16 = const()[name = tensor("input_491_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(943536128)))]; + tensor input_491_beta_0_to_fp16 = const()[name = tensor("input_491_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(943538240)))]; + tensor input_491_epsilon_0_to_fp16 = const()[name = tensor("input_491_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_491_cast_fp16 = batch_norm(beta = input_491_beta_0_to_fp16, epsilon = input_491_epsilon_0_to_fp16, gamma = input_491_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_185_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor input_493_pad_type_0 = const()[name = tensor("input_493_pad_type_0"), val = tensor("valid")]; + tensor input_493_strides_0 = const()[name = tensor("input_493_strides_0"), val = tensor([1, 1])]; + tensor input_493_pad_0 = const()[name = tensor("input_493_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_493_dilations_0 = const()[name = tensor("input_493_dilations_0"), val = tensor([1, 1])]; + tensor input_493_groups_0 = const()[name = tensor("input_493_groups_0"), val = tensor(1)]; + tensor layers_18_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_18_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(943540352)))]; + tensor input_493_cast_fp16 = conv(dilations = input_493_dilations_0, groups = input_493_groups_0, pad = input_493_pad_0, pad_type = input_493_pad_type_0, strides = input_493_strides_0, weight = layers_18_conv_pointwise_conv1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("input_493_cast_fp16")]; + tensor input_495_split_num_splits_0 = const()[name = tensor("input_495_split_num_splits_0"), val = tensor(2)]; + tensor input_495_split_axis_0 = const()[name = tensor("input_495_split_axis_0"), val = tensor(1)]; + tensor input_495_split_cast_fp16_0, tensor input_495_split_cast_fp16_1 = split(axis = input_495_split_axis_0, num_splits = input_495_split_num_splits_0, x = input_493_cast_fp16)[name = tensor("input_495_split_cast_fp16")]; + tensor input_495_split_1_sigmoid_cast_fp16 = sigmoid(x = input_495_split_cast_fp16_1)[name = tensor("input_495_split_1_sigmoid_cast_fp16")]; + tensor input_495_cast_fp16 = mul(x = input_495_split_cast_fp16_0, y = input_495_split_1_sigmoid_cast_fp16)[name = tensor("input_495_cast_fp16")]; + tensor input_497_pad_type_0 = const()[name = tensor("input_497_pad_type_0"), val = tensor("custom")]; + tensor input_497_pad_0 = const()[name = tensor("input_497_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_497_groups_0 = const()[name = tensor("input_497_groups_0"), val = tensor(1024)]; + tensor input_497_strides_0 = const()[name = tensor("input_497_strides_0"), val = tensor([1, 1])]; + tensor input_497_dilations_0 = const()[name = tensor("input_497_dilations_0"), val = tensor([1, 1])]; + tensor const_304_to_fp16 = const()[name = tensor("const_304_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947734720)))]; + tensor const_305_to_fp16 = const()[name = tensor("const_305_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947753216)))]; + tensor input_499_cast_fp16 = conv(bias = const_305_to_fp16, dilations = input_497_dilations_0, groups = input_497_groups_0, pad = input_497_pad_0, pad_type = input_497_pad_type_0, strides = input_497_strides_0, weight = const_304_to_fp16, x = input_495_cast_fp16)[name = tensor("input_499_cast_fp16")]; + tensor input_501_cast_fp16 = silu(x = input_499_cast_fp16)[name = tensor("input_501_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 = const()[name = tensor("layers_18_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(947755328)))]; + 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, x = input_501_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_5439_to_fp16 = const()[name = tensor("op_5439_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_187_cast_fp16 = layer_norm(axes = out_187_axes_0, epsilon = var_5439_to_fp16, x = inputs_187_cast_fp16)[name = tensor("out_187_cast_fp16")]; + tensor input_503_gamma_0_to_fp16 = const()[name = tensor("input_503_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(949852544)))]; + tensor input_503_beta_0_to_fp16 = const()[name = tensor("input_503_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(949854656)))]; + tensor input_503_epsilon_0_to_fp16 = const()[name = tensor("input_503_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_503_cast_fp16 = batch_norm(beta = input_503_beta_0_to_fp16, epsilon = input_503_epsilon_0_to_fp16, gamma = input_503_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_187_cast_fp16)[name = tensor("input_503_cast_fp16")]; + tensor input_505_pad_type_0 = const()[name = tensor("input_505_pad_type_0"), val = tensor("valid")]; + tensor input_505_strides_0 = const()[name = tensor("input_505_strides_0"), val = tensor([1, 1])]; + tensor input_505_pad_0 = const()[name = tensor("input_505_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_505_dilations_0 = const()[name = tensor("input_505_dilations_0"), val = tensor([1, 1])]; + tensor input_505_groups_0 = const()[name = tensor("input_505_groups_0"), val = tensor(1)]; + tensor layers_18_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_18_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(949856768)))]; + tensor input_505_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_505_dilations_0, groups = input_505_groups_0, pad = input_505_pad_0, pad_type = input_505_pad_type_0, strides = input_505_strides_0, weight = layers_18_feed_forward2_fc1_weight_to_fp16, x = input_503_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor input_507_cast_fp16 = silu(x = input_505_cast_fp16)[name = tensor("input_507_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 var_5467_weight_0_to_fp16 = const()[name = tensor("op_5467_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(958245440)))]; + tensor var_5467_cast_fp16 = conv(bias = input_15_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 = var_5467_weight_0_to_fp16, x = input_507_cast_fp16)[name = tensor("op_5467_cast_fp16")]; + tensor inputs_189_cast_fp16 = add(x = inputs_187_cast_fp16, y = var_5467_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_5477_to_fp16 = const()[name = tensor("op_5477_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_189_cast_fp16 = layer_norm(axes = out_189_axes_0, epsilon = var_5477_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(966634112)))]; + 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(966636224)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_189_cast_fp16)[name = tensor("inputs_191_cast_fp16")]; + tensor var_5491 = const()[name = tensor("op_5491"), val = tensor(3)]; + tensor out_191_axes_0 = const()[name = tensor("out_191_axes_0"), val = tensor([1])]; + tensor var_5522_to_fp16 = const()[name = tensor("op_5522_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_191_cast_fp16 = layer_norm(axes = out_191_axes_0, epsilon = var_5522_to_fp16, x = inputs_191_cast_fp16)[name = tensor("out_191_cast_fp16")]; + tensor input_509_gamma_0_to_fp16 = const()[name = tensor("input_509_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(966638336)))]; + tensor input_509_beta_0_to_fp16 = const()[name = tensor("input_509_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(966640448)))]; + tensor input_509_epsilon_0_to_fp16 = const()[name = tensor("input_509_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_509_cast_fp16 = batch_norm(beta = input_509_beta_0_to_fp16, epsilon = input_509_epsilon_0_to_fp16, gamma = input_509_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_191_cast_fp16)[name = tensor("input_509_cast_fp16")]; + tensor input_511_pad_type_0 = const()[name = tensor("input_511_pad_type_0"), val = tensor("valid")]; + tensor input_511_strides_0 = const()[name = tensor("input_511_strides_0"), val = tensor([1, 1])]; + tensor input_511_pad_0 = const()[name = tensor("input_511_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_511_dilations_0 = const()[name = tensor("input_511_dilations_0"), val = tensor([1, 1])]; + tensor input_511_groups_0 = const()[name = tensor("input_511_groups_0"), val = tensor(1)]; + tensor layers_19_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_19_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(966642560)))]; + tensor input_511_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_511_dilations_0, groups = input_511_groups_0, pad = input_511_pad_0, pad_type = input_511_pad_type_0, strides = input_511_strides_0, weight = layers_19_feed_forward1_fc1_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("input_511_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = input_511_cast_fp16)[name = tensor("input_513_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 var_5550_weight_0_to_fp16 = const()[name = tensor("op_5550_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(975031232)))]; + tensor var_5550_cast_fp16 = conv(bias = input_15_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 = var_5550_weight_0_to_fp16, x = input_513_cast_fp16)[name = tensor("op_5550_cast_fp16")]; + tensor inputs_193_cast_fp16 = add(x = inputs_191_cast_fp16, y = var_5550_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_5560_to_fp16 = const()[name = tensor("op_5560_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_193_cast_fp16 = layer_norm(axes = out_193_axes_0, epsilon = var_5560_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(983419904)))]; + 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(983422016)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(983424128)))]; + tensor query_77_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(985521344)))]; + 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, 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 = const()[name = tensor("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(987618560)))]; + tensor value_39_cast_fp16 = conv(bias = input_15_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, x = obj_79_cast_fp16)[name = tensor("value_39_cast_fp16")]; + tensor var_5598_to_fp16 = const()[name = tensor("op_5598_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(989715776)))]; + tensor query_79_cast_fp16 = add(x = query_77_cast_fp16, y = var_5598_to_fp16)[name = tensor("query_79_cast_fp16")]; + tensor var_5601_to_fp16 = const()[name = tensor("op_5601_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(989717888)))]; + tensor q_with_bias_v_39_cast_fp16 = add(x = query_77_cast_fp16, y = var_5601_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 = const()[name = tensor("layers_19_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(989720000)))]; + tensor p_39_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_39_cast_fp16")]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor([1, 8, 128, 188])]; + tensor var_5613_cast_fp16 = reshape(shape = var_5612, x = q_with_bias_v_39_cast_fp16)[name = tensor("op_5613_cast_fp16")]; + tensor var_5614 = const()[name = tensor("op_5614"), val = tensor([1, 8, 128, -1])]; + tensor var_5615_cast_fp16 = reshape(shape = var_5614, x = p_39_cast_fp16)[name = tensor("op_5615_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_5613_cast_fp16, y = var_5615_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_5624 = const()[name = tensor("op_5624"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_157_cast_fp16 = reshape(shape = var_5624, x = matrix_bd_155_cast_fp16)[name = tensor("matrix_bd_157_cast_fp16")]; + tensor var_5628_begin_0 = const()[name = tensor("op_5628_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5628_end_0 = const()[name = tensor("op_5628_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5628_end_mask_0 = const()[name = tensor("op_5628_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5628_cast_fp16 = slice_by_index(begin = var_5628_begin_0, end = var_5628_end_0, end_mask = var_5628_end_mask_0, x = matrix_bd_157_cast_fp16)[name = tensor("op_5628_cast_fp16")]; + tensor var_5629 = const()[name = tensor("op_5629"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_159_cast_fp16 = reshape(shape = var_5629, x = var_5628_cast_fp16)[name = tensor("matrix_bd_159_cast_fp16")]; + tensor var_5634_begin_0 = const()[name = tensor("op_5634_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5634_end_0 = const()[name = tensor("op_5634_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5634_end_mask_0 = const()[name = tensor("op_5634_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5634_cast_fp16 = slice_by_index(begin = var_5634_begin_0, end = var_5634_end_0, end_mask = var_5634_end_mask_0, x = matrix_bd_159_cast_fp16)[name = tensor("op_5634_cast_fp16")]; + tensor var_5635_to_fp16 = const()[name = tensor("op_5635_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_39_cast_fp16 = mul(x = var_5634_cast_fp16, y = var_5635_to_fp16)[name = tensor("qk_mask_39_cast_fp16")]; + tensor var_5639 = const()[name = tensor("op_5639"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_39_cast_fp16 = reshape(shape = var_5639, x = query_79_cast_fp16)[name = tensor("mh_q_39_cast_fp16")]; + tensor var_5641_to_fp16 = const()[name = tensor("op_5641_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5642_cast_fp16 = mul(x = mh_q_39_cast_fp16, y = var_5641_to_fp16)[name = tensor("op_5642_cast_fp16")]; + tensor var_5645 = const()[name = tensor("op_5645"), val = tensor([1, 8, 128, 188])]; + tensor var_5646_cast_fp16 = reshape(shape = var_5645, x = key_39_cast_fp16)[name = tensor("op_5646_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_5642_cast_fp16, y = var_5646_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_5650_cast_fp16 = softmax(axis = var_5491, x = mh_w_79_cast_fp16)[name = tensor("op_5650_cast_fp16")]; + tensor var_5651 = const()[name = tensor("op_5651"), val = tensor([1, 8, 128, 188])]; + tensor var_5652_cast_fp16 = reshape(shape = var_5651, x = value_39_cast_fp16)[name = tensor("op_5652_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_5652_cast_fp16, y = var_5650_cast_fp16)[name = tensor("attn_39_cast_fp16")]; + tensor var_5655 = const()[name = tensor("op_5655"), val = tensor([1, 1024, 1, 188])]; + tensor input_515_cast_fp16 = reshape(shape = var_5655, x = attn_39_cast_fp16)[name = tensor("input_515_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 = const()[name = tensor("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(991817216)))]; + tensor obj_81_cast_fp16 = conv(bias = input_15_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, x = input_515_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_5673_to_fp16 = const()[name = tensor("op_5673_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_195_cast_fp16 = layer_norm(axes = out_195_axes_0, epsilon = var_5673_to_fp16, x = inputs_195_cast_fp16)[name = tensor("out_195_cast_fp16")]; + tensor input_517_gamma_0_to_fp16 = const()[name = tensor("input_517_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(993914432)))]; + tensor input_517_beta_0_to_fp16 = const()[name = tensor("input_517_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(993916544)))]; + tensor input_517_epsilon_0_to_fp16 = const()[name = tensor("input_517_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_517_cast_fp16 = batch_norm(beta = input_517_beta_0_to_fp16, epsilon = input_517_epsilon_0_to_fp16, gamma = input_517_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_195_cast_fp16)[name = tensor("input_517_cast_fp16")]; + tensor input_519_pad_type_0 = const()[name = tensor("input_519_pad_type_0"), val = tensor("valid")]; + tensor input_519_strides_0 = const()[name = tensor("input_519_strides_0"), val = tensor([1, 1])]; + tensor input_519_pad_0 = const()[name = tensor("input_519_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_519_dilations_0 = const()[name = tensor("input_519_dilations_0"), val = tensor([1, 1])]; + tensor input_519_groups_0 = const()[name = tensor("input_519_groups_0"), val = tensor(1)]; + tensor layers_19_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_19_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(993918656)))]; + tensor input_519_cast_fp16 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = layers_19_conv_pointwise_conv1_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("input_519_cast_fp16")]; + tensor input_521_split_num_splits_0 = const()[name = tensor("input_521_split_num_splits_0"), val = tensor(2)]; + tensor input_521_split_axis_0 = const()[name = tensor("input_521_split_axis_0"), val = tensor(1)]; + tensor input_521_split_cast_fp16_0, tensor input_521_split_cast_fp16_1 = split(axis = input_521_split_axis_0, num_splits = input_521_split_num_splits_0, x = input_519_cast_fp16)[name = tensor("input_521_split_cast_fp16")]; + tensor input_521_split_1_sigmoid_cast_fp16 = sigmoid(x = input_521_split_cast_fp16_1)[name = tensor("input_521_split_1_sigmoid_cast_fp16")]; + tensor input_521_cast_fp16 = mul(x = input_521_split_cast_fp16_0, y = input_521_split_1_sigmoid_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor input_523_pad_type_0 = const()[name = tensor("input_523_pad_type_0"), val = tensor("custom")]; + tensor input_523_pad_0 = const()[name = tensor("input_523_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_523_groups_0 = const()[name = tensor("input_523_groups_0"), val = tensor(1024)]; + tensor input_523_strides_0 = const()[name = tensor("input_523_strides_0"), val = tensor([1, 1])]; + tensor input_523_dilations_0 = const()[name = tensor("input_523_dilations_0"), val = tensor([1, 1])]; + tensor const_306_to_fp16 = const()[name = tensor("const_306_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998113024)))]; + tensor const_307_to_fp16 = const()[name = tensor("const_307_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998131520)))]; + tensor input_525_cast_fp16 = conv(bias = const_307_to_fp16, dilations = input_523_dilations_0, groups = input_523_groups_0, pad = input_523_pad_0, pad_type = input_523_pad_type_0, strides = input_523_strides_0, weight = const_306_to_fp16, x = input_521_cast_fp16)[name = tensor("input_525_cast_fp16")]; + tensor input_527_cast_fp16 = silu(x = input_525_cast_fp16)[name = tensor("input_527_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 = const()[name = tensor("layers_19_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(998133632)))]; + 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, x = input_527_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_5721_to_fp16 = const()[name = tensor("op_5721_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_197_cast_fp16 = layer_norm(axes = out_197_axes_0, epsilon = var_5721_to_fp16, x = inputs_197_cast_fp16)[name = tensor("out_197_cast_fp16")]; + tensor input_529_gamma_0_to_fp16 = const()[name = tensor("input_529_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1000230848)))]; + tensor input_529_beta_0_to_fp16 = const()[name = tensor("input_529_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1000232960)))]; + tensor input_529_epsilon_0_to_fp16 = const()[name = tensor("input_529_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_529_cast_fp16 = batch_norm(beta = input_529_beta_0_to_fp16, epsilon = input_529_epsilon_0_to_fp16, gamma = input_529_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_197_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor input_531_pad_type_0 = const()[name = tensor("input_531_pad_type_0"), val = tensor("valid")]; + tensor input_531_strides_0 = const()[name = tensor("input_531_strides_0"), val = tensor([1, 1])]; + tensor input_531_pad_0 = const()[name = tensor("input_531_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_531_dilations_0 = const()[name = tensor("input_531_dilations_0"), val = tensor([1, 1])]; + tensor input_531_groups_0 = const()[name = tensor("input_531_groups_0"), val = tensor(1)]; + tensor layers_19_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_19_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1000235072)))]; + tensor input_531_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_531_dilations_0, groups = input_531_groups_0, pad = input_531_pad_0, pad_type = input_531_pad_type_0, strides = input_531_strides_0, weight = layers_19_feed_forward2_fc1_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("input_531_cast_fp16")]; + tensor input_533_cast_fp16 = silu(x = input_531_cast_fp16)[name = tensor("input_533_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 var_5749_weight_0_to_fp16 = const()[name = tensor("op_5749_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1008623744)))]; + tensor var_5749_cast_fp16 = conv(bias = input_15_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 = var_5749_weight_0_to_fp16, x = input_533_cast_fp16)[name = tensor("op_5749_cast_fp16")]; + tensor inputs_199_cast_fp16 = add(x = inputs_197_cast_fp16, y = var_5749_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_5759_to_fp16 = const()[name = tensor("op_5759_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_199_cast_fp16 = layer_norm(axes = out_199_axes_0, epsilon = var_5759_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(1017012416)))]; + 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(1017014528)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_199_cast_fp16)[name = tensor("inputs_201_cast_fp16")]; + tensor var_5773 = const()[name = tensor("op_5773"), val = tensor(3)]; + tensor out_201_axes_0 = const()[name = tensor("out_201_axes_0"), val = tensor([1])]; + tensor var_5804_to_fp16 = const()[name = tensor("op_5804_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_201_cast_fp16 = layer_norm(axes = out_201_axes_0, epsilon = var_5804_to_fp16, x = inputs_201_cast_fp16)[name = tensor("out_201_cast_fp16")]; + tensor input_535_gamma_0_to_fp16 = const()[name = tensor("input_535_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1017016640)))]; + tensor input_535_beta_0_to_fp16 = const()[name = tensor("input_535_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1017018752)))]; + tensor input_535_epsilon_0_to_fp16 = const()[name = tensor("input_535_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_535_cast_fp16 = batch_norm(beta = input_535_beta_0_to_fp16, epsilon = input_535_epsilon_0_to_fp16, gamma = input_535_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_201_cast_fp16)[name = tensor("input_535_cast_fp16")]; + tensor input_537_pad_type_0 = const()[name = tensor("input_537_pad_type_0"), val = tensor("valid")]; + tensor input_537_strides_0 = const()[name = tensor("input_537_strides_0"), val = tensor([1, 1])]; + tensor input_537_pad_0 = const()[name = tensor("input_537_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_537_dilations_0 = const()[name = tensor("input_537_dilations_0"), val = tensor([1, 1])]; + tensor input_537_groups_0 = const()[name = tensor("input_537_groups_0"), val = tensor(1)]; + tensor layers_20_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_20_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1017020864)))]; + tensor input_537_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_537_dilations_0, groups = input_537_groups_0, pad = input_537_pad_0, pad_type = input_537_pad_type_0, strides = input_537_strides_0, weight = layers_20_feed_forward1_fc1_weight_to_fp16, x = input_535_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor input_539_cast_fp16 = silu(x = input_537_cast_fp16)[name = tensor("input_539_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 var_5832_weight_0_to_fp16 = const()[name = tensor("op_5832_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1025409536)))]; + tensor var_5832_cast_fp16 = conv(bias = input_15_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 = var_5832_weight_0_to_fp16, x = input_539_cast_fp16)[name = tensor("op_5832_cast_fp16")]; + tensor inputs_203_cast_fp16 = add(x = inputs_201_cast_fp16, y = var_5832_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_5842_to_fp16 = const()[name = tensor("op_5842_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_203_cast_fp16 = layer_norm(axes = out_203_axes_0, epsilon = var_5842_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(1033798208)))]; + 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(1033800320)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1033802432)))]; + tensor query_81_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1035899648)))]; + 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, 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 = const()[name = tensor("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1037996864)))]; + tensor value_41_cast_fp16 = conv(bias = input_15_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, x = obj_83_cast_fp16)[name = tensor("value_41_cast_fp16")]; + tensor var_5880_to_fp16 = const()[name = tensor("op_5880_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040094080)))]; + tensor query_83_cast_fp16 = add(x = query_81_cast_fp16, y = var_5880_to_fp16)[name = tensor("query_83_cast_fp16")]; + tensor var_5883_to_fp16 = const()[name = tensor("op_5883_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040096192)))]; + tensor q_with_bias_v_41_cast_fp16 = add(x = query_81_cast_fp16, y = var_5883_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 = const()[name = tensor("layers_20_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1040098304)))]; + tensor p_41_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_41_cast_fp16")]; + tensor var_5894 = const()[name = tensor("op_5894"), val = tensor([1, 8, 128, 188])]; + tensor var_5895_cast_fp16 = reshape(shape = var_5894, x = q_with_bias_v_41_cast_fp16)[name = tensor("op_5895_cast_fp16")]; + tensor var_5896 = const()[name = tensor("op_5896"), val = tensor([1, 8, 128, -1])]; + tensor var_5897_cast_fp16 = reshape(shape = var_5896, x = p_41_cast_fp16)[name = tensor("op_5897_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_5895_cast_fp16, y = var_5897_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_5906 = const()[name = tensor("op_5906"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_165_cast_fp16 = reshape(shape = var_5906, x = matrix_bd_163_cast_fp16)[name = tensor("matrix_bd_165_cast_fp16")]; + tensor var_5910_begin_0 = const()[name = tensor("op_5910_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_5910_end_0 = const()[name = tensor("op_5910_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_5910_end_mask_0 = const()[name = tensor("op_5910_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_5910_cast_fp16 = slice_by_index(begin = var_5910_begin_0, end = var_5910_end_0, end_mask = var_5910_end_mask_0, x = matrix_bd_165_cast_fp16)[name = tensor("op_5910_cast_fp16")]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_167_cast_fp16 = reshape(shape = var_5911, x = var_5910_cast_fp16)[name = tensor("matrix_bd_167_cast_fp16")]; + tensor var_5916_begin_0 = const()[name = tensor("op_5916_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_5916_end_0 = const()[name = tensor("op_5916_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_5916_end_mask_0 = const()[name = tensor("op_5916_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_5916_cast_fp16 = slice_by_index(begin = var_5916_begin_0, end = var_5916_end_0, end_mask = var_5916_end_mask_0, x = matrix_bd_167_cast_fp16)[name = tensor("op_5916_cast_fp16")]; + tensor var_5917_to_fp16 = const()[name = tensor("op_5917_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_41_cast_fp16 = mul(x = var_5916_cast_fp16, y = var_5917_to_fp16)[name = tensor("qk_mask_41_cast_fp16")]; + tensor var_5921 = const()[name = tensor("op_5921"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_41_cast_fp16 = reshape(shape = var_5921, x = query_83_cast_fp16)[name = tensor("mh_q_41_cast_fp16")]; + tensor var_5923_to_fp16 = const()[name = tensor("op_5923_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_5924_cast_fp16 = mul(x = mh_q_41_cast_fp16, y = var_5923_to_fp16)[name = tensor("op_5924_cast_fp16")]; + tensor var_5927 = const()[name = tensor("op_5927"), val = tensor([1, 8, 128, 188])]; + tensor var_5928_cast_fp16 = reshape(shape = var_5927, x = key_41_cast_fp16)[name = tensor("op_5928_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_5924_cast_fp16, y = var_5928_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_5932_cast_fp16 = softmax(axis = var_5773, x = mh_w_83_cast_fp16)[name = tensor("op_5932_cast_fp16")]; + tensor var_5933 = const()[name = tensor("op_5933"), val = tensor([1, 8, 128, 188])]; + tensor var_5934_cast_fp16 = reshape(shape = var_5933, x = value_41_cast_fp16)[name = tensor("op_5934_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_5934_cast_fp16, y = var_5932_cast_fp16)[name = tensor("attn_41_cast_fp16")]; + tensor var_5937 = const()[name = tensor("op_5937"), val = tensor([1, 1024, 1, 188])]; + tensor input_541_cast_fp16 = reshape(shape = var_5937, x = attn_41_cast_fp16)[name = tensor("input_541_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 = const()[name = tensor("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1042195520)))]; + tensor obj_85_cast_fp16 = conv(bias = input_15_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, x = input_541_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_5955_to_fp16 = const()[name = tensor("op_5955_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_205_cast_fp16 = layer_norm(axes = out_205_axes_0, epsilon = var_5955_to_fp16, x = inputs_205_cast_fp16)[name = tensor("out_205_cast_fp16")]; + tensor input_543_gamma_0_to_fp16 = const()[name = tensor("input_543_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044292736)))]; + tensor input_543_beta_0_to_fp16 = const()[name = tensor("input_543_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044294848)))]; + tensor input_543_epsilon_0_to_fp16 = const()[name = tensor("input_543_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_543_cast_fp16 = batch_norm(beta = input_543_beta_0_to_fp16, epsilon = input_543_epsilon_0_to_fp16, gamma = input_543_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_205_cast_fp16)[name = tensor("input_543_cast_fp16")]; + tensor input_545_pad_type_0 = const()[name = tensor("input_545_pad_type_0"), val = tensor("valid")]; + tensor input_545_strides_0 = const()[name = tensor("input_545_strides_0"), val = tensor([1, 1])]; + tensor input_545_pad_0 = const()[name = tensor("input_545_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_545_dilations_0 = const()[name = tensor("input_545_dilations_0"), val = tensor([1, 1])]; + tensor input_545_groups_0 = const()[name = tensor("input_545_groups_0"), val = tensor(1)]; + tensor layers_20_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_20_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1044296960)))]; + tensor input_545_cast_fp16 = conv(dilations = input_545_dilations_0, groups = input_545_groups_0, pad = input_545_pad_0, pad_type = input_545_pad_type_0, strides = input_545_strides_0, weight = layers_20_conv_pointwise_conv1_weight_to_fp16, x = input_543_cast_fp16)[name = tensor("input_545_cast_fp16")]; + tensor input_547_split_num_splits_0 = const()[name = tensor("input_547_split_num_splits_0"), val = tensor(2)]; + tensor input_547_split_axis_0 = const()[name = tensor("input_547_split_axis_0"), val = tensor(1)]; + tensor input_547_split_cast_fp16_0, tensor input_547_split_cast_fp16_1 = split(axis = input_547_split_axis_0, num_splits = input_547_split_num_splits_0, x = input_545_cast_fp16)[name = tensor("input_547_split_cast_fp16")]; + tensor input_547_split_1_sigmoid_cast_fp16 = sigmoid(x = input_547_split_cast_fp16_1)[name = tensor("input_547_split_1_sigmoid_cast_fp16")]; + tensor input_547_cast_fp16 = mul(x = input_547_split_cast_fp16_0, y = input_547_split_1_sigmoid_cast_fp16)[name = tensor("input_547_cast_fp16")]; + tensor input_549_pad_type_0 = const()[name = tensor("input_549_pad_type_0"), val = tensor("custom")]; + tensor input_549_pad_0 = const()[name = tensor("input_549_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_549_groups_0 = const()[name = tensor("input_549_groups_0"), val = tensor(1024)]; + tensor input_549_strides_0 = const()[name = tensor("input_549_strides_0"), val = tensor([1, 1])]; + tensor input_549_dilations_0 = const()[name = tensor("input_549_dilations_0"), val = tensor([1, 1])]; + tensor const_308_to_fp16 = const()[name = tensor("const_308_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1048491328)))]; + tensor const_309_to_fp16 = const()[name = tensor("const_309_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1048509824)))]; + tensor input_551_cast_fp16 = conv(bias = const_309_to_fp16, dilations = input_549_dilations_0, groups = input_549_groups_0, pad = input_549_pad_0, pad_type = input_549_pad_type_0, strides = input_549_strides_0, weight = const_308_to_fp16, x = input_547_cast_fp16)[name = tensor("input_551_cast_fp16")]; + tensor input_553_cast_fp16 = silu(x = input_551_cast_fp16)[name = tensor("input_553_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 = const()[name = tensor("layers_20_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1048511936)))]; + 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, x = input_553_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_6003_to_fp16 = const()[name = tensor("op_6003_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_207_cast_fp16 = layer_norm(axes = out_207_axes_0, epsilon = var_6003_to_fp16, x = inputs_207_cast_fp16)[name = tensor("out_207_cast_fp16")]; + tensor input_555_gamma_0_to_fp16 = const()[name = tensor("input_555_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050609152)))]; + tensor input_555_beta_0_to_fp16 = const()[name = tensor("input_555_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050611264)))]; + tensor input_555_epsilon_0_to_fp16 = const()[name = tensor("input_555_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_555_cast_fp16 = batch_norm(beta = input_555_beta_0_to_fp16, epsilon = input_555_epsilon_0_to_fp16, gamma = input_555_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_207_cast_fp16)[name = tensor("input_555_cast_fp16")]; + tensor input_557_pad_type_0 = const()[name = tensor("input_557_pad_type_0"), val = tensor("valid")]; + tensor input_557_strides_0 = const()[name = tensor("input_557_strides_0"), val = tensor([1, 1])]; + tensor input_557_pad_0 = const()[name = tensor("input_557_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_557_dilations_0 = const()[name = tensor("input_557_dilations_0"), val = tensor([1, 1])]; + tensor input_557_groups_0 = const()[name = tensor("input_557_groups_0"), val = tensor(1)]; + tensor layers_20_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_20_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1050613376)))]; + tensor input_557_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_557_dilations_0, groups = input_557_groups_0, pad = input_557_pad_0, pad_type = input_557_pad_type_0, strides = input_557_strides_0, weight = layers_20_feed_forward2_fc1_weight_to_fp16, x = input_555_cast_fp16)[name = tensor("input_557_cast_fp16")]; + tensor input_559_cast_fp16 = silu(x = input_557_cast_fp16)[name = tensor("input_559_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 var_6031_weight_0_to_fp16 = const()[name = tensor("op_6031_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1059002048)))]; + tensor var_6031_cast_fp16 = conv(bias = input_15_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 = var_6031_weight_0_to_fp16, x = input_559_cast_fp16)[name = tensor("op_6031_cast_fp16")]; + tensor inputs_209_cast_fp16 = add(x = inputs_207_cast_fp16, y = var_6031_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_6041_to_fp16 = const()[name = tensor("op_6041_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_209_cast_fp16 = layer_norm(axes = out_209_axes_0, epsilon = var_6041_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(1067390720)))]; + 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(1067392832)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_209_cast_fp16)[name = tensor("inputs_211_cast_fp16")]; + tensor var_6055 = const()[name = tensor("op_6055"), val = tensor(3)]; + tensor out_211_axes_0 = const()[name = tensor("out_211_axes_0"), val = tensor([1])]; + tensor var_6086_to_fp16 = const()[name = tensor("op_6086_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_211_cast_fp16 = layer_norm(axes = out_211_axes_0, epsilon = var_6086_to_fp16, x = inputs_211_cast_fp16)[name = tensor("out_211_cast_fp16")]; + tensor input_561_gamma_0_to_fp16 = const()[name = tensor("input_561_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1067394944)))]; + tensor input_561_beta_0_to_fp16 = const()[name = tensor("input_561_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1067397056)))]; + tensor input_561_epsilon_0_to_fp16 = const()[name = tensor("input_561_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_561_cast_fp16 = batch_norm(beta = input_561_beta_0_to_fp16, epsilon = input_561_epsilon_0_to_fp16, gamma = input_561_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_211_cast_fp16)[name = tensor("input_561_cast_fp16")]; + tensor input_563_pad_type_0 = const()[name = tensor("input_563_pad_type_0"), val = tensor("valid")]; + tensor input_563_strides_0 = const()[name = tensor("input_563_strides_0"), val = tensor([1, 1])]; + tensor input_563_pad_0 = const()[name = tensor("input_563_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_563_dilations_0 = const()[name = tensor("input_563_dilations_0"), val = tensor([1, 1])]; + tensor input_563_groups_0 = const()[name = tensor("input_563_groups_0"), val = tensor(1)]; + tensor layers_21_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_21_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1067399168)))]; + tensor input_563_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_563_dilations_0, groups = input_563_groups_0, pad = input_563_pad_0, pad_type = input_563_pad_type_0, strides = input_563_strides_0, weight = layers_21_feed_forward1_fc1_weight_to_fp16, x = input_561_cast_fp16)[name = tensor("input_563_cast_fp16")]; + tensor input_565_cast_fp16 = silu(x = input_563_cast_fp16)[name = tensor("input_565_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 var_6114_weight_0_to_fp16 = const()[name = tensor("op_6114_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1075787840)))]; + tensor var_6114_cast_fp16 = conv(bias = input_15_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 = var_6114_weight_0_to_fp16, x = input_565_cast_fp16)[name = tensor("op_6114_cast_fp16")]; + tensor inputs_213_cast_fp16 = add(x = inputs_211_cast_fp16, y = var_6114_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_6124_to_fp16 = const()[name = tensor("op_6124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_213_cast_fp16 = layer_norm(axes = out_213_axes_0, epsilon = var_6124_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(1084176512)))]; + 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(1084178624)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1084180736)))]; + tensor query_85_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1086277952)))]; + 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, 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 = const()[name = tensor("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088375168)))]; + tensor value_43_cast_fp16 = conv(bias = input_15_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, x = obj_87_cast_fp16)[name = tensor("value_43_cast_fp16")]; + tensor var_6162_to_fp16 = const()[name = tensor("op_6162_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090472384)))]; + tensor query_87_cast_fp16 = add(x = query_85_cast_fp16, y = var_6162_to_fp16)[name = tensor("query_87_cast_fp16")]; + tensor var_6165_to_fp16 = const()[name = tensor("op_6165_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090474496)))]; + tensor q_with_bias_v_43_cast_fp16 = add(x = query_85_cast_fp16, y = var_6165_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 = const()[name = tensor("layers_21_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1090476608)))]; + tensor p_43_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_43_cast_fp16")]; + tensor var_6176 = const()[name = tensor("op_6176"), val = tensor([1, 8, 128, 188])]; + tensor var_6177_cast_fp16 = reshape(shape = var_6176, x = q_with_bias_v_43_cast_fp16)[name = tensor("op_6177_cast_fp16")]; + tensor var_6178 = const()[name = tensor("op_6178"), val = tensor([1, 8, 128, -1])]; + tensor var_6179_cast_fp16 = reshape(shape = var_6178, x = p_43_cast_fp16)[name = tensor("op_6179_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_6177_cast_fp16, y = var_6179_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_6188 = const()[name = tensor("op_6188"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_173_cast_fp16 = reshape(shape = var_6188, x = matrix_bd_171_cast_fp16)[name = tensor("matrix_bd_173_cast_fp16")]; + tensor var_6192_begin_0 = const()[name = tensor("op_6192_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_6192_end_0 = const()[name = tensor("op_6192_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_6192_end_mask_0 = const()[name = tensor("op_6192_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6192_cast_fp16 = slice_by_index(begin = var_6192_begin_0, end = var_6192_end_0, end_mask = var_6192_end_mask_0, x = matrix_bd_173_cast_fp16)[name = tensor("op_6192_cast_fp16")]; + tensor var_6193 = const()[name = tensor("op_6193"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_175_cast_fp16 = reshape(shape = var_6193, x = var_6192_cast_fp16)[name = tensor("matrix_bd_175_cast_fp16")]; + tensor var_6198_begin_0 = const()[name = tensor("op_6198_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6198_end_0 = const()[name = tensor("op_6198_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_6198_end_mask_0 = const()[name = tensor("op_6198_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6198_cast_fp16 = slice_by_index(begin = var_6198_begin_0, end = var_6198_end_0, end_mask = var_6198_end_mask_0, x = matrix_bd_175_cast_fp16)[name = tensor("op_6198_cast_fp16")]; + tensor var_6199_to_fp16 = const()[name = tensor("op_6199_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_43_cast_fp16 = mul(x = var_6198_cast_fp16, y = var_6199_to_fp16)[name = tensor("qk_mask_43_cast_fp16")]; + tensor var_6203 = const()[name = tensor("op_6203"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_43_cast_fp16 = reshape(shape = var_6203, x = query_87_cast_fp16)[name = tensor("mh_q_43_cast_fp16")]; + tensor var_6205_to_fp16 = const()[name = tensor("op_6205_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_6206_cast_fp16 = mul(x = mh_q_43_cast_fp16, y = var_6205_to_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor var_6209 = const()[name = tensor("op_6209"), val = tensor([1, 8, 128, 188])]; + tensor var_6210_cast_fp16 = reshape(shape = var_6209, x = key_43_cast_fp16)[name = tensor("op_6210_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_6206_cast_fp16, y = var_6210_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_6214_cast_fp16 = softmax(axis = var_6055, x = mh_w_87_cast_fp16)[name = tensor("op_6214_cast_fp16")]; + tensor var_6215 = const()[name = tensor("op_6215"), val = tensor([1, 8, 128, 188])]; + tensor var_6216_cast_fp16 = reshape(shape = var_6215, x = value_43_cast_fp16)[name = tensor("op_6216_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_6216_cast_fp16, y = var_6214_cast_fp16)[name = tensor("attn_43_cast_fp16")]; + tensor var_6219 = const()[name = tensor("op_6219"), val = tensor([1, 1024, 1, 188])]; + tensor input_567_cast_fp16 = reshape(shape = var_6219, x = attn_43_cast_fp16)[name = tensor("input_567_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 = const()[name = tensor("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1092573824)))]; + tensor obj_89_cast_fp16 = conv(bias = input_15_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, x = input_567_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_6237_to_fp16 = const()[name = tensor("op_6237_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_215_cast_fp16 = layer_norm(axes = out_215_axes_0, epsilon = var_6237_to_fp16, x = inputs_215_cast_fp16)[name = tensor("out_215_cast_fp16")]; + tensor input_569_gamma_0_to_fp16 = const()[name = tensor("input_569_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1094671040)))]; + tensor input_569_beta_0_to_fp16 = const()[name = tensor("input_569_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1094673152)))]; + tensor input_569_epsilon_0_to_fp16 = const()[name = tensor("input_569_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_569_cast_fp16 = batch_norm(beta = input_569_beta_0_to_fp16, epsilon = input_569_epsilon_0_to_fp16, gamma = input_569_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_215_cast_fp16)[name = tensor("input_569_cast_fp16")]; + tensor input_571_pad_type_0 = const()[name = tensor("input_571_pad_type_0"), val = tensor("valid")]; + tensor input_571_strides_0 = const()[name = tensor("input_571_strides_0"), val = tensor([1, 1])]; + tensor input_571_pad_0 = const()[name = tensor("input_571_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_571_dilations_0 = const()[name = tensor("input_571_dilations_0"), val = tensor([1, 1])]; + tensor input_571_groups_0 = const()[name = tensor("input_571_groups_0"), val = tensor(1)]; + tensor layers_21_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_21_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1094675264)))]; + tensor input_571_cast_fp16 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = layers_21_conv_pointwise_conv1_weight_to_fp16, x = input_569_cast_fp16)[name = tensor("input_571_cast_fp16")]; + tensor input_573_split_num_splits_0 = const()[name = tensor("input_573_split_num_splits_0"), val = tensor(2)]; + tensor input_573_split_axis_0 = const()[name = tensor("input_573_split_axis_0"), val = tensor(1)]; + tensor input_573_split_cast_fp16_0, tensor input_573_split_cast_fp16_1 = split(axis = input_573_split_axis_0, num_splits = input_573_split_num_splits_0, x = input_571_cast_fp16)[name = tensor("input_573_split_cast_fp16")]; + tensor input_573_split_1_sigmoid_cast_fp16 = sigmoid(x = input_573_split_cast_fp16_1)[name = tensor("input_573_split_1_sigmoid_cast_fp16")]; + tensor input_573_cast_fp16 = mul(x = input_573_split_cast_fp16_0, y = input_573_split_1_sigmoid_cast_fp16)[name = tensor("input_573_cast_fp16")]; + tensor input_575_pad_type_0 = const()[name = tensor("input_575_pad_type_0"), val = tensor("custom")]; + tensor input_575_pad_0 = const()[name = tensor("input_575_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_575_groups_0 = const()[name = tensor("input_575_groups_0"), val = tensor(1024)]; + tensor input_575_strides_0 = const()[name = tensor("input_575_strides_0"), val = tensor([1, 1])]; + tensor input_575_dilations_0 = const()[name = tensor("input_575_dilations_0"), val = tensor([1, 1])]; + tensor const_310_to_fp16 = const()[name = tensor("const_310_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1098869632)))]; + tensor const_311_to_fp16 = const()[name = tensor("const_311_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1098888128)))]; + tensor input_577_cast_fp16 = conv(bias = const_311_to_fp16, dilations = input_575_dilations_0, groups = input_575_groups_0, pad = input_575_pad_0, pad_type = input_575_pad_type_0, strides = input_575_strides_0, weight = const_310_to_fp16, x = input_573_cast_fp16)[name = tensor("input_577_cast_fp16")]; + tensor input_579_cast_fp16 = silu(x = input_577_cast_fp16)[name = tensor("input_579_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 = const()[name = tensor("layers_21_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1098890240)))]; + 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, x = input_579_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_6285_to_fp16 = const()[name = tensor("op_6285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_217_cast_fp16 = layer_norm(axes = out_217_axes_0, epsilon = var_6285_to_fp16, x = inputs_217_cast_fp16)[name = tensor("out_217_cast_fp16")]; + tensor input_581_gamma_0_to_fp16 = const()[name = tensor("input_581_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1100987456)))]; + tensor input_581_beta_0_to_fp16 = const()[name = tensor("input_581_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1100989568)))]; + tensor input_581_epsilon_0_to_fp16 = const()[name = tensor("input_581_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_581_cast_fp16 = batch_norm(beta = input_581_beta_0_to_fp16, epsilon = input_581_epsilon_0_to_fp16, gamma = input_581_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_217_cast_fp16)[name = tensor("input_581_cast_fp16")]; + tensor input_583_pad_type_0 = const()[name = tensor("input_583_pad_type_0"), val = tensor("valid")]; + tensor input_583_strides_0 = const()[name = tensor("input_583_strides_0"), val = tensor([1, 1])]; + tensor input_583_pad_0 = const()[name = tensor("input_583_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_583_dilations_0 = const()[name = tensor("input_583_dilations_0"), val = tensor([1, 1])]; + tensor input_583_groups_0 = const()[name = tensor("input_583_groups_0"), val = tensor(1)]; + tensor layers_21_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_21_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1100991680)))]; + tensor input_583_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_583_dilations_0, groups = input_583_groups_0, pad = input_583_pad_0, pad_type = input_583_pad_type_0, strides = input_583_strides_0, weight = layers_21_feed_forward2_fc1_weight_to_fp16, x = input_581_cast_fp16)[name = tensor("input_583_cast_fp16")]; + tensor input_585_cast_fp16 = silu(x = input_583_cast_fp16)[name = tensor("input_585_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 var_6313_weight_0_to_fp16 = const()[name = tensor("op_6313_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1109380352)))]; + tensor var_6313_cast_fp16 = conv(bias = input_15_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 = var_6313_weight_0_to_fp16, x = input_585_cast_fp16)[name = tensor("op_6313_cast_fp16")]; + tensor inputs_219_cast_fp16 = add(x = inputs_217_cast_fp16, y = var_6313_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_6323_to_fp16 = const()[name = tensor("op_6323_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_219_cast_fp16 = layer_norm(axes = out_219_axes_0, epsilon = var_6323_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(1117769024)))]; + 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(1117771136)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_219_cast_fp16)[name = tensor("inputs_221_cast_fp16")]; + tensor var_6337 = const()[name = tensor("op_6337"), val = tensor(3)]; + tensor out_221_axes_0 = const()[name = tensor("out_221_axes_0"), val = tensor([1])]; + tensor var_6368_to_fp16 = const()[name = tensor("op_6368_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_221_cast_fp16 = layer_norm(axes = out_221_axes_0, epsilon = var_6368_to_fp16, x = inputs_221_cast_fp16)[name = tensor("out_221_cast_fp16")]; + tensor input_587_gamma_0_to_fp16 = const()[name = tensor("input_587_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1117773248)))]; + tensor input_587_beta_0_to_fp16 = const()[name = tensor("input_587_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1117775360)))]; + tensor input_587_epsilon_0_to_fp16 = const()[name = tensor("input_587_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_587_cast_fp16 = batch_norm(beta = input_587_beta_0_to_fp16, epsilon = input_587_epsilon_0_to_fp16, gamma = input_587_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_221_cast_fp16)[name = tensor("input_587_cast_fp16")]; + tensor input_589_pad_type_0 = const()[name = tensor("input_589_pad_type_0"), val = tensor("valid")]; + tensor input_589_strides_0 = const()[name = tensor("input_589_strides_0"), val = tensor([1, 1])]; + tensor input_589_pad_0 = const()[name = tensor("input_589_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_589_dilations_0 = const()[name = tensor("input_589_dilations_0"), val = tensor([1, 1])]; + tensor input_589_groups_0 = const()[name = tensor("input_589_groups_0"), val = tensor(1)]; + tensor layers_22_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_22_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1117777472)))]; + tensor input_589_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_589_dilations_0, groups = input_589_groups_0, pad = input_589_pad_0, pad_type = input_589_pad_type_0, strides = input_589_strides_0, weight = layers_22_feed_forward1_fc1_weight_to_fp16, x = input_587_cast_fp16)[name = tensor("input_589_cast_fp16")]; + tensor input_591_cast_fp16 = silu(x = input_589_cast_fp16)[name = tensor("input_591_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 var_6396_weight_0_to_fp16 = const()[name = tensor("op_6396_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1126166144)))]; + tensor var_6396_cast_fp16 = conv(bias = input_15_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 = var_6396_weight_0_to_fp16, x = input_591_cast_fp16)[name = tensor("op_6396_cast_fp16")]; + tensor inputs_223_cast_fp16 = add(x = inputs_221_cast_fp16, y = var_6396_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_6406_to_fp16 = const()[name = tensor("op_6406_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_223_cast_fp16 = layer_norm(axes = out_223_axes_0, epsilon = var_6406_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(1134554816)))]; + 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(1134556928)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1134559040)))]; + tensor query_89_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1136656256)))]; + 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, 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 = const()[name = tensor("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1138753472)))]; + tensor value_45_cast_fp16 = conv(bias = input_15_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, x = obj_91_cast_fp16)[name = tensor("value_45_cast_fp16")]; + tensor var_6444_to_fp16 = const()[name = tensor("op_6444_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1140850688)))]; + tensor query_91_cast_fp16 = add(x = query_89_cast_fp16, y = var_6444_to_fp16)[name = tensor("query_91_cast_fp16")]; + tensor var_6447_to_fp16 = const()[name = tensor("op_6447_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1140852800)))]; + tensor q_with_bias_v_45_cast_fp16 = add(x = query_89_cast_fp16, y = var_6447_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 = const()[name = tensor("layers_22_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1140854912)))]; + tensor p_45_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_45_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([1, 8, 128, 188])]; + tensor var_6459_cast_fp16 = reshape(shape = var_6458, x = q_with_bias_v_45_cast_fp16)[name = tensor("op_6459_cast_fp16")]; + tensor var_6460 = const()[name = tensor("op_6460"), val = tensor([1, 8, 128, -1])]; + tensor var_6461_cast_fp16 = reshape(shape = var_6460, x = p_45_cast_fp16)[name = tensor("op_6461_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_6459_cast_fp16, y = var_6461_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_6470 = const()[name = tensor("op_6470"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_181_cast_fp16 = reshape(shape = var_6470, x = matrix_bd_179_cast_fp16)[name = tensor("matrix_bd_181_cast_fp16")]; + tensor var_6474_begin_0 = const()[name = tensor("op_6474_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_6474_end_0 = const()[name = tensor("op_6474_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_6474_end_mask_0 = const()[name = tensor("op_6474_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6474_cast_fp16 = slice_by_index(begin = var_6474_begin_0, end = var_6474_end_0, end_mask = var_6474_end_mask_0, x = matrix_bd_181_cast_fp16)[name = tensor("op_6474_cast_fp16")]; + tensor var_6475 = const()[name = tensor("op_6475"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_183_cast_fp16 = reshape(shape = var_6475, x = var_6474_cast_fp16)[name = tensor("matrix_bd_183_cast_fp16")]; + tensor var_6480_begin_0 = const()[name = tensor("op_6480_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6480_end_0 = const()[name = tensor("op_6480_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_6480_end_mask_0 = const()[name = tensor("op_6480_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6480_cast_fp16 = slice_by_index(begin = var_6480_begin_0, end = var_6480_end_0, end_mask = var_6480_end_mask_0, x = matrix_bd_183_cast_fp16)[name = tensor("op_6480_cast_fp16")]; + tensor var_6481_to_fp16 = const()[name = tensor("op_6481_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_45_cast_fp16 = mul(x = var_6480_cast_fp16, y = var_6481_to_fp16)[name = tensor("qk_mask_45_cast_fp16")]; + tensor var_6485 = const()[name = tensor("op_6485"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_45_cast_fp16 = reshape(shape = var_6485, x = query_91_cast_fp16)[name = tensor("mh_q_45_cast_fp16")]; + tensor var_6487_to_fp16 = const()[name = tensor("op_6487_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_6488_cast_fp16 = mul(x = mh_q_45_cast_fp16, y = var_6487_to_fp16)[name = tensor("op_6488_cast_fp16")]; + tensor var_6491 = const()[name = tensor("op_6491"), val = tensor([1, 8, 128, 188])]; + tensor var_6492_cast_fp16 = reshape(shape = var_6491, x = key_45_cast_fp16)[name = tensor("op_6492_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_6488_cast_fp16, y = var_6492_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_6496_cast_fp16 = softmax(axis = var_6337, x = mh_w_91_cast_fp16)[name = tensor("op_6496_cast_fp16")]; + tensor var_6497 = const()[name = tensor("op_6497"), val = tensor([1, 8, 128, 188])]; + tensor var_6498_cast_fp16 = reshape(shape = var_6497, x = value_45_cast_fp16)[name = tensor("op_6498_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_6498_cast_fp16, y = var_6496_cast_fp16)[name = tensor("attn_45_cast_fp16")]; + tensor var_6501 = const()[name = tensor("op_6501"), val = tensor([1, 1024, 1, 188])]; + tensor input_593_cast_fp16 = reshape(shape = var_6501, x = attn_45_cast_fp16)[name = tensor("input_593_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 = const()[name = tensor("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1142952128)))]; + tensor obj_93_cast_fp16 = conv(bias = input_15_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, x = input_593_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_6519_to_fp16 = const()[name = tensor("op_6519_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_225_cast_fp16 = layer_norm(axes = out_225_axes_0, epsilon = var_6519_to_fp16, x = inputs_225_cast_fp16)[name = tensor("out_225_cast_fp16")]; + tensor input_595_gamma_0_to_fp16 = const()[name = tensor("input_595_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1145049344)))]; + tensor input_595_beta_0_to_fp16 = const()[name = tensor("input_595_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1145051456)))]; + tensor input_595_epsilon_0_to_fp16 = const()[name = tensor("input_595_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_595_cast_fp16 = batch_norm(beta = input_595_beta_0_to_fp16, epsilon = input_595_epsilon_0_to_fp16, gamma = input_595_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_225_cast_fp16)[name = tensor("input_595_cast_fp16")]; + tensor input_597_pad_type_0 = const()[name = tensor("input_597_pad_type_0"), val = tensor("valid")]; + tensor input_597_strides_0 = const()[name = tensor("input_597_strides_0"), val = tensor([1, 1])]; + tensor input_597_pad_0 = const()[name = tensor("input_597_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_597_dilations_0 = const()[name = tensor("input_597_dilations_0"), val = tensor([1, 1])]; + tensor input_597_groups_0 = const()[name = tensor("input_597_groups_0"), val = tensor(1)]; + tensor layers_22_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_22_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1145053568)))]; + tensor input_597_cast_fp16 = conv(dilations = input_597_dilations_0, groups = input_597_groups_0, pad = input_597_pad_0, pad_type = input_597_pad_type_0, strides = input_597_strides_0, weight = layers_22_conv_pointwise_conv1_weight_to_fp16, x = input_595_cast_fp16)[name = tensor("input_597_cast_fp16")]; + tensor input_599_split_num_splits_0 = const()[name = tensor("input_599_split_num_splits_0"), val = tensor(2)]; + tensor input_599_split_axis_0 = const()[name = tensor("input_599_split_axis_0"), val = tensor(1)]; + tensor input_599_split_cast_fp16_0, tensor input_599_split_cast_fp16_1 = split(axis = input_599_split_axis_0, num_splits = input_599_split_num_splits_0, x = input_597_cast_fp16)[name = tensor("input_599_split_cast_fp16")]; + tensor input_599_split_1_sigmoid_cast_fp16 = sigmoid(x = input_599_split_cast_fp16_1)[name = tensor("input_599_split_1_sigmoid_cast_fp16")]; + tensor input_599_cast_fp16 = mul(x = input_599_split_cast_fp16_0, y = input_599_split_1_sigmoid_cast_fp16)[name = tensor("input_599_cast_fp16")]; + tensor input_601_pad_type_0 = const()[name = tensor("input_601_pad_type_0"), val = tensor("custom")]; + tensor input_601_pad_0 = const()[name = tensor("input_601_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_601_groups_0 = const()[name = tensor("input_601_groups_0"), val = tensor(1024)]; + tensor input_601_strides_0 = const()[name = tensor("input_601_strides_0"), val = tensor([1, 1])]; + tensor input_601_dilations_0 = const()[name = tensor("input_601_dilations_0"), val = tensor([1, 1])]; + tensor const_312_to_fp16 = const()[name = tensor("const_312_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1149247936)))]; + tensor const_313_to_fp16 = const()[name = tensor("const_313_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1149266432)))]; + tensor input_603_cast_fp16 = conv(bias = const_313_to_fp16, dilations = input_601_dilations_0, groups = input_601_groups_0, pad = input_601_pad_0, pad_type = input_601_pad_type_0, strides = input_601_strides_0, weight = const_312_to_fp16, x = input_599_cast_fp16)[name = tensor("input_603_cast_fp16")]; + tensor input_605_cast_fp16 = silu(x = input_603_cast_fp16)[name = tensor("input_605_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 = const()[name = tensor("layers_22_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1149268544)))]; + 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, x = input_605_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_6567_to_fp16 = const()[name = tensor("op_6567_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_227_cast_fp16 = layer_norm(axes = out_227_axes_0, epsilon = var_6567_to_fp16, x = inputs_227_cast_fp16)[name = tensor("out_227_cast_fp16")]; + tensor input_607_gamma_0_to_fp16 = const()[name = tensor("input_607_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1151365760)))]; + tensor input_607_beta_0_to_fp16 = const()[name = tensor("input_607_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1151367872)))]; + tensor input_607_epsilon_0_to_fp16 = const()[name = tensor("input_607_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_607_cast_fp16 = batch_norm(beta = input_607_beta_0_to_fp16, epsilon = input_607_epsilon_0_to_fp16, gamma = input_607_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_227_cast_fp16)[name = tensor("input_607_cast_fp16")]; + tensor input_609_pad_type_0 = const()[name = tensor("input_609_pad_type_0"), val = tensor("valid")]; + tensor input_609_strides_0 = const()[name = tensor("input_609_strides_0"), val = tensor([1, 1])]; + tensor input_609_pad_0 = const()[name = tensor("input_609_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_609_dilations_0 = const()[name = tensor("input_609_dilations_0"), val = tensor([1, 1])]; + tensor input_609_groups_0 = const()[name = tensor("input_609_groups_0"), val = tensor(1)]; + tensor layers_22_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_22_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1151369984)))]; + tensor input_609_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_609_dilations_0, groups = input_609_groups_0, pad = input_609_pad_0, pad_type = input_609_pad_type_0, strides = input_609_strides_0, weight = layers_22_feed_forward2_fc1_weight_to_fp16, x = input_607_cast_fp16)[name = tensor("input_609_cast_fp16")]; + tensor input_611_cast_fp16 = silu(x = input_609_cast_fp16)[name = tensor("input_611_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 var_6595_weight_0_to_fp16 = const()[name = tensor("op_6595_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1159758656)))]; + tensor var_6595_cast_fp16 = conv(bias = input_15_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 = var_6595_weight_0_to_fp16, x = input_611_cast_fp16)[name = tensor("op_6595_cast_fp16")]; + tensor inputs_229_cast_fp16 = add(x = inputs_227_cast_fp16, y = var_6595_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_6605_to_fp16 = const()[name = tensor("op_6605_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_229_cast_fp16 = layer_norm(axes = out_229_axes_0, epsilon = var_6605_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(1168147328)))]; + 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(1168149440)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_229_cast_fp16)[name = tensor("inputs_231_cast_fp16")]; + tensor var_6619 = const()[name = tensor("op_6619"), val = tensor(3)]; + tensor out_231_axes_0 = const()[name = tensor("out_231_axes_0"), val = tensor([1])]; + tensor var_6650_to_fp16 = const()[name = tensor("op_6650_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_231_cast_fp16 = layer_norm(axes = out_231_axes_0, epsilon = var_6650_to_fp16, x = inputs_231_cast_fp16)[name = tensor("out_231_cast_fp16")]; + tensor input_613_gamma_0_to_fp16 = const()[name = tensor("input_613_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168151552)))]; + tensor input_613_beta_0_to_fp16 = const()[name = tensor("input_613_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168153664)))]; + tensor input_613_epsilon_0_to_fp16 = const()[name = tensor("input_613_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_613_cast_fp16 = batch_norm(beta = input_613_beta_0_to_fp16, epsilon = input_613_epsilon_0_to_fp16, gamma = input_613_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_231_cast_fp16)[name = tensor("input_613_cast_fp16")]; + tensor input_615_pad_type_0 = const()[name = tensor("input_615_pad_type_0"), val = tensor("valid")]; + tensor input_615_strides_0 = const()[name = tensor("input_615_strides_0"), val = tensor([1, 1])]; + tensor input_615_pad_0 = const()[name = tensor("input_615_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_615_dilations_0 = const()[name = tensor("input_615_dilations_0"), val = tensor([1, 1])]; + tensor input_615_groups_0 = const()[name = tensor("input_615_groups_0"), val = tensor(1)]; + tensor layers_23_feed_forward1_fc1_weight_to_fp16 = const()[name = tensor("layers_23_feed_forward1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1168155776)))]; + tensor input_615_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_615_dilations_0, groups = input_615_groups_0, pad = input_615_pad_0, pad_type = input_615_pad_type_0, strides = input_615_strides_0, weight = layers_23_feed_forward1_fc1_weight_to_fp16, x = input_613_cast_fp16)[name = tensor("input_615_cast_fp16")]; + tensor input_617_cast_fp16 = silu(x = input_615_cast_fp16)[name = tensor("input_617_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 var_6678_weight_0_to_fp16 = const()[name = tensor("op_6678_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1176544448)))]; + tensor var_6678_cast_fp16 = conv(bias = input_15_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 = var_6678_weight_0_to_fp16, x = input_617_cast_fp16)[name = tensor("op_6678_cast_fp16")]; + tensor inputs_233_cast_fp16 = add(x = inputs_231_cast_fp16, y = var_6678_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_6688_to_fp16 = const()[name = tensor("op_6688_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_233_cast_fp16 = layer_norm(axes = out_233_axes_0, epsilon = var_6688_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(1184933120)))]; + 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(1184935232)))]; + 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_15_mean_0_to_fp16, variance = input_15_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 = const()[name = tensor("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1184937344)))]; + tensor query_93_cast_fp16 = conv(bias = input_15_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, 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 = const()[name = tensor("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1187034560)))]; + 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, 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 = const()[name = tensor("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1189131776)))]; + tensor value_cast_fp16 = conv(bias = input_15_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, x = obj_95_cast_fp16)[name = tensor("value_cast_fp16")]; + tensor var_6726_to_fp16 = const()[name = tensor("op_6726_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1191228992)))]; + tensor query_cast_fp16 = add(x = query_93_cast_fp16, y = var_6726_to_fp16)[name = tensor("query_cast_fp16")]; + tensor var_6729_to_fp16 = const()[name = tensor("op_6729_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1191231104)))]; + tensor q_with_bias_v_cast_fp16 = add(x = query_93_cast_fp16, y = var_6729_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 = const()[name = tensor("layers_23_self_attn_linear_pos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1191233216)))]; + tensor p_cast_fp16 = conv(bias = input_15_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, x = pos_enc_to_fp16)[name = tensor("p_cast_fp16")]; + tensor var_6740 = const()[name = tensor("op_6740"), val = tensor([1, 8, 128, 188])]; + tensor var_6741_cast_fp16 = reshape(shape = var_6740, x = q_with_bias_v_cast_fp16)[name = tensor("op_6741_cast_fp16")]; + tensor var_6742 = const()[name = tensor("op_6742"), val = tensor([1, 8, 128, -1])]; + tensor var_6743_cast_fp16 = reshape(shape = var_6742, x = p_cast_fp16)[name = tensor("op_6743_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_6741_cast_fp16, y = var_6743_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_6752 = const()[name = tensor("op_6752"), val = tensor([1, 8, -1, 188])]; + tensor matrix_bd_189_cast_fp16 = reshape(shape = var_6752, x = matrix_bd_187_cast_fp16)[name = tensor("matrix_bd_189_cast_fp16")]; + tensor var_6756_begin_0 = const()[name = tensor("op_6756_begin_0"), val = tensor([0, 0, 1, 0])]; + tensor var_6756_end_0 = const()[name = tensor("op_6756_end_0"), val = tensor([1, 8, 376, 188])]; + tensor var_6756_end_mask_0 = const()[name = tensor("op_6756_end_mask_0"), val = tensor([true, true, true, true])]; + tensor var_6756_cast_fp16 = slice_by_index(begin = var_6756_begin_0, end = var_6756_end_0, end_mask = var_6756_end_mask_0, x = matrix_bd_189_cast_fp16)[name = tensor("op_6756_cast_fp16")]; + tensor var_6757 = const()[name = tensor("op_6757"), val = tensor([1, 8, 188, 375])]; + tensor matrix_bd_cast_fp16 = reshape(shape = var_6757, x = var_6756_cast_fp16)[name = tensor("matrix_bd_cast_fp16")]; + tensor var_6762_begin_0 = const()[name = tensor("op_6762_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6762_end_0 = const()[name = tensor("op_6762_end_0"), val = tensor([1, 8, 188, 188])]; + tensor var_6762_end_mask_0 = const()[name = tensor("op_6762_end_mask_0"), val = tensor([true, true, true, false])]; + tensor var_6762_cast_fp16 = slice_by_index(begin = var_6762_begin_0, end = var_6762_end_0, end_mask = var_6762_end_mask_0, x = matrix_bd_cast_fp16)[name = tensor("op_6762_cast_fp16")]; + tensor var_6763_to_fp16 = const()[name = tensor("op_6763_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor qk_mask_cast_fp16 = mul(x = var_6762_cast_fp16, y = var_6763_to_fp16)[name = tensor("qk_mask_cast_fp16")]; + tensor var_6767 = const()[name = tensor("op_6767"), val = tensor([1, 8, 128, 188])]; + tensor mh_q_cast_fp16 = reshape(shape = var_6767, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; + tensor var_6769_to_fp16 = const()[name = tensor("op_6769_to_fp16"), val = tensor(0x1.6ap-4)]; + tensor var_6770_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_6769_to_fp16)[name = tensor("op_6770_cast_fp16")]; + tensor var_6773 = const()[name = tensor("op_6773"), val = tensor([1, 8, 128, 188])]; + tensor var_6774_cast_fp16 = reshape(shape = var_6773, x = key_cast_fp16)[name = tensor("op_6774_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_6770_cast_fp16, y = var_6774_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_6778_cast_fp16 = softmax(axis = var_6619, x = mh_w_cast_fp16)[name = tensor("op_6778_cast_fp16")]; + tensor var_6779 = const()[name = tensor("op_6779"), val = tensor([1, 8, 128, 188])]; + tensor var_6780_cast_fp16 = reshape(shape = var_6779, x = value_cast_fp16)[name = tensor("op_6780_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_6780_cast_fp16, y = var_6778_cast_fp16)[name = tensor("attn_cast_fp16")]; + tensor var_6783 = const()[name = tensor("op_6783"), val = tensor([1, 1024, 1, 188])]; + tensor input_619_cast_fp16 = reshape(shape = var_6783, x = attn_cast_fp16)[name = tensor("input_619_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 = const()[name = tensor("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1193330432)))]; + tensor obj_cast_fp16 = conv(bias = input_15_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, x = input_619_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_6801_to_fp16 = const()[name = tensor("op_6801_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_235_cast_fp16 = layer_norm(axes = out_235_axes_0, epsilon = var_6801_to_fp16, x = inputs_235_cast_fp16)[name = tensor("out_235_cast_fp16")]; + tensor input_621_gamma_0_to_fp16 = const()[name = tensor("input_621_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195427648)))]; + tensor input_621_beta_0_to_fp16 = const()[name = tensor("input_621_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195429760)))]; + tensor input_621_epsilon_0_to_fp16 = const()[name = tensor("input_621_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_621_cast_fp16 = batch_norm(beta = input_621_beta_0_to_fp16, epsilon = input_621_epsilon_0_to_fp16, gamma = input_621_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_235_cast_fp16)[name = tensor("input_621_cast_fp16")]; + tensor input_623_pad_type_0 = const()[name = tensor("input_623_pad_type_0"), val = tensor("valid")]; + tensor input_623_strides_0 = const()[name = tensor("input_623_strides_0"), val = tensor([1, 1])]; + tensor input_623_pad_0 = const()[name = tensor("input_623_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_623_dilations_0 = const()[name = tensor("input_623_dilations_0"), val = tensor([1, 1])]; + tensor input_623_groups_0 = const()[name = tensor("input_623_groups_0"), val = tensor(1)]; + tensor layers_23_conv_pointwise_conv1_weight_to_fp16 = const()[name = tensor("layers_23_conv_pointwise_conv1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1195431872)))]; + tensor input_623_cast_fp16 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = layers_23_conv_pointwise_conv1_weight_to_fp16, x = input_621_cast_fp16)[name = tensor("input_623_cast_fp16")]; + tensor input_625_split_num_splits_0 = const()[name = tensor("input_625_split_num_splits_0"), val = tensor(2)]; + tensor input_625_split_axis_0 = const()[name = tensor("input_625_split_axis_0"), val = tensor(1)]; + tensor input_625_split_cast_fp16_0, tensor input_625_split_cast_fp16_1 = split(axis = input_625_split_axis_0, num_splits = input_625_split_num_splits_0, x = input_623_cast_fp16)[name = tensor("input_625_split_cast_fp16")]; + tensor input_625_split_1_sigmoid_cast_fp16 = sigmoid(x = input_625_split_cast_fp16_1)[name = tensor("input_625_split_1_sigmoid_cast_fp16")]; + tensor input_625_cast_fp16 = mul(x = input_625_split_cast_fp16_0, y = input_625_split_1_sigmoid_cast_fp16)[name = tensor("input_625_cast_fp16")]; + tensor input_627_pad_type_0 = const()[name = tensor("input_627_pad_type_0"), val = tensor("custom")]; + tensor input_627_pad_0 = const()[name = tensor("input_627_pad_0"), val = tensor([0, 0, 4, 4])]; + tensor input_627_groups_0 = const()[name = tensor("input_627_groups_0"), val = tensor(1024)]; + tensor input_627_strides_0 = const()[name = tensor("input_627_strides_0"), val = tensor([1, 1])]; + tensor input_627_dilations_0 = const()[name = tensor("input_627_dilations_0"), val = tensor([1, 1])]; + tensor const_314_to_fp16 = const()[name = tensor("const_314_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1199626240)))]; + tensor const_315_to_fp16 = const()[name = tensor("const_315_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1199644736)))]; + tensor input_629_cast_fp16 = conv(bias = const_315_to_fp16, dilations = input_627_dilations_0, groups = input_627_groups_0, pad = input_627_pad_0, pad_type = input_627_pad_type_0, strides = input_627_strides_0, weight = const_314_to_fp16, x = input_625_cast_fp16)[name = tensor("input_629_cast_fp16")]; + tensor input_631_cast_fp16 = silu(x = input_629_cast_fp16)[name = tensor("input_631_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 = const()[name = tensor("layers_23_conv_pointwise_conv2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1199646848)))]; + 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, x = input_631_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_6849_to_fp16 = const()[name = tensor("op_6849_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_237_cast_fp16 = layer_norm(axes = out_237_axes_0, epsilon = var_6849_to_fp16, x = inputs_237_cast_fp16)[name = tensor("out_237_cast_fp16")]; + tensor input_633_gamma_0_to_fp16 = const()[name = tensor("input_633_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201744064)))]; + tensor input_633_beta_0_to_fp16 = const()[name = tensor("input_633_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201746176)))]; + tensor input_633_epsilon_0_to_fp16 = const()[name = tensor("input_633_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_633_cast_fp16 = batch_norm(beta = input_633_beta_0_to_fp16, epsilon = input_633_epsilon_0_to_fp16, gamma = input_633_gamma_0_to_fp16, mean = input_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_237_cast_fp16)[name = tensor("input_633_cast_fp16")]; + tensor input_635_pad_type_0 = const()[name = tensor("input_635_pad_type_0"), val = tensor("valid")]; + tensor input_635_strides_0 = const()[name = tensor("input_635_strides_0"), val = tensor([1, 1])]; + tensor input_635_pad_0 = const()[name = tensor("input_635_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor input_635_dilations_0 = const()[name = tensor("input_635_dilations_0"), val = tensor([1, 1])]; + tensor input_635_groups_0 = const()[name = tensor("input_635_groups_0"), val = tensor(1)]; + tensor layers_23_feed_forward2_fc1_weight_to_fp16 = const()[name = tensor("layers_23_feed_forward2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1201748288)))]; + tensor input_635_cast_fp16 = conv(bias = layers_0_feed_forward1_fc1_bias_to_fp16, dilations = input_635_dilations_0, groups = input_635_groups_0, pad = input_635_pad_0, pad_type = input_635_pad_type_0, strides = input_635_strides_0, weight = layers_23_feed_forward2_fc1_weight_to_fp16, x = input_633_cast_fp16)[name = tensor("input_635_cast_fp16")]; + tensor input_637_cast_fp16 = silu(x = input_635_cast_fp16)[name = tensor("input_637_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 var_6877_weight_0_to_fp16 = const()[name = tensor("op_6877_weight_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1210136960)))]; + tensor var_6877_cast_fp16 = conv(bias = input_15_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 = var_6877_weight_0_to_fp16, x = input_637_cast_fp16)[name = tensor("op_6877_cast_fp16")]; + tensor inputs_cast_fp16 = add(x = inputs_237_cast_fp16, y = var_6877_cast_fp16)[name = tensor("inputs_cast_fp16")]; + tensor out_239_axes_0 = const()[name = tensor("out_239_axes_0"), val = tensor([1])]; + tensor var_6887_to_fp16 = const()[name = tensor("op_6887_to_fp16"), val = tensor(0x1.5p-17)]; + tensor out_239_cast_fp16 = layer_norm(axes = out_239_axes_0, epsilon = var_6887_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(1218525632)))]; + 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(1218527744)))]; + 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_15_mean_0_to_fp16, variance = input_15_variance_0_to_fp16, x = out_239_cast_fp16)[name = tensor("encoder_output_embeds_type_fp32_cast_fp16")]; + tensor var_6907_pad_type_0 = const()[name = tensor("op_6907_pad_type_0"), val = tensor("valid")]; + tensor var_6907_strides_0 = const()[name = tensor("op_6907_strides_0"), val = tensor([1, 1])]; + tensor var_6907_pad_0 = const()[name = tensor("op_6907_pad_0"), val = tensor([0, 0, 0, 0])]; + tensor var_6907_dilations_0 = const()[name = tensor("op_6907_dilations_0"), val = tensor([1, 1])]; + tensor var_6907_groups_0 = const()[name = tensor("op_6907_groups_0"), val = tensor(1)]; + tensor joint_projection_weight_to_fp16 = const()[name = tensor("joint_projection_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1218529856)))]; + 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(1219840640)))]; + tensor joint_projected_encoder_output_embeds = conv(bias = joint_projection_bias_to_fp16, dilations = var_6907_dilations_0, groups = var_6907_groups_0, pad = var_6907_pad_0, pad_type = var_6907_pad_type_0, strides = var_6907_strides_0, weight = joint_projection_weight_to_fp16, x = encoder_output_embeds)[name = tensor("op_6907_cast_fp16")]; + } -> (encoder_output_embeds, joint_projected_encoder_output_embeds); +} \ No newline at end of file