program(1.3) [buildInfo = dict({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { func main(tensor attention_mask, tensor mel_features) { tensor var_669 = const()[name = string("op_669"), val = tensor([1, 128, 30, 100])]; string mel_features_to_fp16_dtype_0 = const()[name = string("mel_features_to_fp16_dtype_0"), val = string("fp16")]; tensor mel_features_to_fp16 = cast(dtype = mel_features_to_fp16_dtype_0, x = mel_features)[name = string("cast_78")]; tensor var_670_cast_fp16 = reshape(shape = var_669, x = mel_features_to_fp16)[name = string("op_670_cast_fp16")]; tensor var_675 = const()[name = string("op_675"), val = tensor([0, 2, 1, 3])]; tensor var_681 = const()[name = string("op_681"), val = tensor([30, 1, 128, 100])]; tensor var_676_cast_fp16 = transpose(perm = var_675, x = var_670_cast_fp16)[name = string("transpose_163")]; tensor input_1_cast_fp16 = reshape(shape = var_681, x = var_676_cast_fp16)[name = string("input_1_cast_fp16")]; string var_694_pad_type_0 = const()[name = string("op_694_pad_type_0"), val = string("custom")]; tensor var_694_pad_0 = const()[name = string("op_694_pad_0"), val = tensor([1, 1, 1, 1])]; tensor var_694_strides_0 = const()[name = string("op_694_strides_0"), val = tensor([2, 2])]; tensor var_694_dilations_0 = const()[name = string("op_694_dilations_0"), val = tensor([1, 1])]; int32 var_694_groups_0 = const()[name = string("op_694_groups_0"), val = int32(1)]; tensor tower_conv2d1_weight_to_fp16 = const()[name = string("tower_conv2d1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; tensor tower_conv2d1_bias_to_fp16 = const()[name = string("tower_conv2d1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8768)))]; tensor var_694_cast_fp16 = conv(bias = tower_conv2d1_bias_to_fp16, dilations = var_694_dilations_0, groups = var_694_groups_0, pad = var_694_pad_0, pad_type = var_694_pad_type_0, strides = var_694_strides_0, weight = tower_conv2d1_weight_to_fp16, x = input_1_cast_fp16)[name = string("op_694_cast_fp16")]; string input_3_mode_0 = const()[name = string("input_3_mode_0"), val = string("EXACT")]; tensor input_3_cast_fp16 = gelu(mode = input_3_mode_0, x = var_694_cast_fp16)[name = string("input_3_cast_fp16")]; string var_708_pad_type_0 = const()[name = string("op_708_pad_type_0"), val = string("custom")]; tensor var_708_pad_0 = const()[name = string("op_708_pad_0"), val = tensor([1, 1, 1, 1])]; tensor var_708_strides_0 = const()[name = string("op_708_strides_0"), val = tensor([2, 2])]; tensor var_708_dilations_0 = const()[name = string("op_708_dilations_0"), val = tensor([1, 1])]; int32 var_708_groups_0 = const()[name = string("op_708_groups_0"), val = int32(1)]; tensor tower_conv2d2_weight_to_fp16 = const()[name = string("tower_conv2d2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9792)))]; tensor tower_conv2d2_bias_to_fp16 = const()[name = string("tower_conv2d2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4157056)))]; tensor var_708_cast_fp16 = conv(bias = tower_conv2d2_bias_to_fp16, dilations = var_708_dilations_0, groups = var_708_groups_0, pad = var_708_pad_0, pad_type = var_708_pad_type_0, strides = var_708_strides_0, weight = tower_conv2d2_weight_to_fp16, x = input_3_cast_fp16)[name = string("op_708_cast_fp16")]; string input_5_mode_0 = const()[name = string("input_5_mode_0"), val = string("EXACT")]; tensor input_5_cast_fp16 = gelu(mode = input_5_mode_0, x = var_708_cast_fp16)[name = string("input_5_cast_fp16")]; string var_722_pad_type_0 = const()[name = string("op_722_pad_type_0"), val = string("custom")]; tensor var_722_pad_0 = const()[name = string("op_722_pad_0"), val = tensor([1, 1, 1, 1])]; tensor var_722_strides_0 = const()[name = string("op_722_strides_0"), val = tensor([2, 2])]; tensor var_722_dilations_0 = const()[name = string("op_722_dilations_0"), val = tensor([1, 1])]; int32 var_722_groups_0 = const()[name = string("op_722_groups_0"), val = int32(1)]; tensor tower_conv2d3_weight_to_fp16 = const()[name = string("tower_conv2d3_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4158080)))]; tensor tower_conv2d3_bias_to_fp16 = const()[name = string("tower_conv2d3_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8305344)))]; tensor var_722_cast_fp16 = conv(bias = tower_conv2d3_bias_to_fp16, dilations = var_722_dilations_0, groups = var_722_groups_0, pad = var_722_pad_0, pad_type = var_722_pad_type_0, strides = var_722_strides_0, weight = tower_conv2d3_weight_to_fp16, x = input_5_cast_fp16)[name = string("op_722_cast_fp16")]; string hidden_1_mode_0 = const()[name = string("hidden_1_mode_0"), val = string("EXACT")]; tensor hidden_1_cast_fp16 = gelu(mode = hidden_1_mode_0, x = var_722_cast_fp16)[name = string("hidden_1_cast_fp16")]; tensor var_729 = const()[name = string("op_729"), val = tensor([0, 3, 1, 2])]; tensor var_734 = const()[name = string("op_734"), val = tensor([30, 13, 7680])]; tensor var_730_cast_fp16 = transpose(perm = var_729, x = hidden_1_cast_fp16)[name = string("transpose_162")]; tensor input_7_cast_fp16 = reshape(shape = var_734, x = var_730_cast_fp16)[name = string("input_7_cast_fp16")]; tensor tower_conv_out_weight_to_fp16 = const()[name = string("tower_conv_out_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8306368)))]; tensor linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22068992)))]; tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = tower_conv_out_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_0_cast_fp16")]; tensor const_0_to_fp16 = const()[name = string("const_0_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22070848)))]; tensor hidden_5_cast_fp16 = add(x = linear_0_cast_fp16, y = const_0_to_fp16)[name = string("hidden_5_cast_fp16")]; tensor var_754 = const()[name = string("op_754"), val = tensor([1, 390, 896])]; tensor input_9_cast_fp16 = reshape(shape = var_754, x = hidden_5_cast_fp16)[name = string("input_9_cast_fp16")]; tensor input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor([-1])]; tensor tower_layers_0_self_attn_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_0_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22094208)))]; tensor tower_layers_0_self_attn_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_0_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22096064)))]; fp16 var_757_to_fp16 = const()[name = string("op_757_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = tower_layers_0_self_attn_layer_norm_bias_to_fp16, epsilon = var_757_to_fp16, gamma = tower_layers_0_self_attn_layer_norm_weight_to_fp16, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")]; tensor tower_layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("tower_layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(22097920)))]; tensor tower_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("tower_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23703616)))]; tensor linear_1_cast_fp16 = linear(bias = tower_layers_0_self_attn_q_proj_bias_to_fp16, weight = tower_layers_0_self_attn_q_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_1_cast_fp16")]; tensor var_770 = const()[name = string("op_770"), val = tensor([1, 390, 14, 64])]; tensor query_1_cast_fp16 = reshape(shape = var_770, x = linear_1_cast_fp16)[name = string("query_1_cast_fp16")]; tensor tower_layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("tower_layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23705472)))]; tensor tower_layers_0_self_attn_k_proj_bias_to_fp16 = const()[name = string("tower_layers_0_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25311168)))]; tensor linear_2_cast_fp16 = linear(bias = tower_layers_0_self_attn_k_proj_bias_to_fp16, weight = tower_layers_0_self_attn_k_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_2_cast_fp16")]; tensor var_779 = const()[name = string("op_779"), val = tensor([1, 390, 14, 64])]; tensor key_1_cast_fp16 = reshape(shape = var_779, x = linear_2_cast_fp16)[name = string("key_1_cast_fp16")]; tensor tower_layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("tower_layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25313024)))]; tensor tower_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("tower_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26918720)))]; tensor linear_3_cast_fp16 = linear(bias = tower_layers_0_self_attn_v_proj_bias_to_fp16, weight = tower_layers_0_self_attn_v_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_3_cast_fp16")]; tensor var_788 = const()[name = string("op_788"), val = tensor([1, 390, 14, 64])]; tensor value_1_cast_fp16 = reshape(shape = var_788, x = linear_3_cast_fp16)[name = string("value_1_cast_fp16")]; tensor var_806 = const()[name = string("op_806"), val = tensor([0, 2, 1, 3])]; bool var_811_transpose_x_0 = const()[name = string("op_811_transpose_x_0"), val = bool(false)]; bool var_811_transpose_y_0 = const()[name = string("op_811_transpose_y_0"), val = bool(false)]; tensor transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_55 = transpose(perm = transpose_55_perm_0, x = key_1_cast_fp16)[name = string("transpose_159")]; tensor transpose_54 = transpose(perm = transpose_54_perm_0, x = query_1_cast_fp16)[name = string("transpose_160")]; tensor var_811_cast_fp16 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = transpose_54, y = transpose_55)[name = string("op_811_cast_fp16")]; fp16 var_812_to_fp16 = const()[name = string("op_812_to_fp16"), val = fp16(0x1p-3)]; tensor scores_1_cast_fp16 = mul(x = var_811_cast_fp16, y = var_812_to_fp16)[name = string("scores_1_cast_fp16")]; string attention_mask_to_fp16_dtype_0 = const()[name = string("attention_mask_to_fp16_dtype_0"), val = string("fp16")]; tensor attention_mask_to_fp16 = cast(dtype = attention_mask_to_fp16_dtype_0, x = attention_mask)[name = string("cast_77")]; tensor scores_3_cast_fp16 = add(x = scores_1_cast_fp16, y = attention_mask_to_fp16)[name = string("scores_3_cast_fp16")]; int32 var_821 = const()[name = string("op_821"), val = int32(-1)]; tensor var_823_cast_fp16 = softmax(axis = var_821, x = scores_3_cast_fp16)[name = string("op_823_cast_fp16")]; bool output_1_transpose_x_0 = const()[name = string("output_1_transpose_x_0"), val = bool(false)]; bool output_1_transpose_y_0 = const()[name = string("output_1_transpose_y_0"), val = bool(false)]; tensor value_3_cast_fp16 = transpose(perm = var_806, x = value_1_cast_fp16)[name = string("transpose_161")]; tensor output_1_cast_fp16 = matmul(transpose_x = output_1_transpose_x_0, transpose_y = output_1_transpose_y_0, x = var_823_cast_fp16, y = value_3_cast_fp16)[name = string("output_1_cast_fp16")]; tensor var_834 = const()[name = string("op_834"), val = tensor([0, 2, 1, 3])]; tensor var_839 = const()[name = string("op_839"), val = tensor([1, 390, 896])]; tensor var_835_cast_fp16 = transpose(perm = var_834, x = output_1_cast_fp16)[name = string("transpose_158")]; tensor input_13_cast_fp16 = reshape(shape = var_839, x = var_835_cast_fp16)[name = string("input_13_cast_fp16")]; tensor tower_layers_0_self_attn_out_proj_weight_to_fp16 = const()[name = string("tower_layers_0_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26920576)))]; tensor tower_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = string("tower_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28526272)))]; tensor linear_4_cast_fp16 = linear(bias = tower_layers_0_self_attn_out_proj_bias_to_fp16, weight = tower_layers_0_self_attn_out_proj_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_4_cast_fp16")]; tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = string("input_15_cast_fp16")]; tensor input_17_axes_0 = const()[name = string("input_17_axes_0"), val = tensor([-1])]; tensor tower_layers_0_final_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_0_final_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28528128)))]; tensor tower_layers_0_final_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_0_final_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28529984)))]; fp16 var_847_to_fp16 = const()[name = string("op_847_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = tower_layers_0_final_layer_norm_bias_to_fp16, epsilon = var_847_to_fp16, gamma = tower_layers_0_final_layer_norm_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; tensor tower_layers_0_fc1_weight_to_fp16 = const()[name = string("tower_layers_0_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28531840)))]; tensor tower_layers_0_fc1_bias_to_fp16 = const()[name = string("tower_layers_0_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34954432)))]; tensor linear_5_cast_fp16 = linear(bias = tower_layers_0_fc1_bias_to_fp16, weight = tower_layers_0_fc1_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_5_cast_fp16")]; string input_21_mode_0 = const()[name = string("input_21_mode_0"), val = string("EXACT")]; tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_5_cast_fp16)[name = string("input_21_cast_fp16")]; tensor tower_layers_0_fc2_weight_to_fp16 = const()[name = string("tower_layers_0_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34961664)))]; tensor tower_layers_0_fc2_bias_to_fp16 = const()[name = string("tower_layers_0_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41384256)))]; tensor linear_6_cast_fp16 = linear(bias = tower_layers_0_fc2_bias_to_fp16, weight = tower_layers_0_fc2_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_6_cast_fp16")]; tensor input_23_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = string("input_23_cast_fp16")]; tensor input_25_axes_0 = const()[name = string("input_25_axes_0"), val = tensor([-1])]; tensor tower_layers_1_self_attn_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_1_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41386112)))]; tensor tower_layers_1_self_attn_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_1_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41387968)))]; fp16 var_864_to_fp16 = const()[name = string("op_864_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_25_cast_fp16 = layer_norm(axes = input_25_axes_0, beta = tower_layers_1_self_attn_layer_norm_bias_to_fp16, epsilon = var_864_to_fp16, gamma = tower_layers_1_self_attn_layer_norm_weight_to_fp16, x = input_23_cast_fp16)[name = string("input_25_cast_fp16")]; tensor tower_layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("tower_layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41389824)))]; tensor tower_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("tower_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42995520)))]; tensor linear_7_cast_fp16 = linear(bias = tower_layers_1_self_attn_q_proj_bias_to_fp16, weight = tower_layers_1_self_attn_q_proj_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_7_cast_fp16")]; tensor var_877 = const()[name = string("op_877"), val = tensor([1, 390, 14, 64])]; tensor query_5_cast_fp16 = reshape(shape = var_877, x = linear_7_cast_fp16)[name = string("query_5_cast_fp16")]; tensor tower_layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("tower_layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42997376)))]; tensor tower_layers_1_self_attn_k_proj_bias_to_fp16 = const()[name = string("tower_layers_1_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44603072)))]; tensor linear_8_cast_fp16 = linear(bias = tower_layers_1_self_attn_k_proj_bias_to_fp16, weight = tower_layers_1_self_attn_k_proj_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_8_cast_fp16")]; tensor var_886 = const()[name = string("op_886"), val = tensor([1, 390, 14, 64])]; tensor key_5_cast_fp16 = reshape(shape = var_886, x = linear_8_cast_fp16)[name = string("key_5_cast_fp16")]; tensor tower_layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("tower_layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44604928)))]; tensor tower_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("tower_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46210624)))]; tensor linear_9_cast_fp16 = linear(bias = tower_layers_1_self_attn_v_proj_bias_to_fp16, weight = tower_layers_1_self_attn_v_proj_weight_to_fp16, x = input_25_cast_fp16)[name = string("linear_9_cast_fp16")]; tensor var_895 = const()[name = string("op_895"), val = tensor([1, 390, 14, 64])]; tensor value_5_cast_fp16 = reshape(shape = var_895, x = linear_9_cast_fp16)[name = string("value_5_cast_fp16")]; tensor var_913 = const()[name = string("op_913"), val = tensor([0, 2, 1, 3])]; bool var_918_transpose_x_0 = const()[name = string("op_918_transpose_x_0"), val = bool(false)]; bool var_918_transpose_y_0 = const()[name = string("op_918_transpose_y_0"), val = bool(false)]; tensor transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_57 = transpose(perm = transpose_57_perm_0, x = key_5_cast_fp16)[name = string("transpose_155")]; tensor transpose_56 = transpose(perm = transpose_56_perm_0, x = query_5_cast_fp16)[name = string("transpose_156")]; tensor var_918_cast_fp16 = matmul(transpose_x = var_918_transpose_x_0, transpose_y = var_918_transpose_y_0, x = transpose_56, y = transpose_57)[name = string("op_918_cast_fp16")]; fp16 var_919_to_fp16 = const()[name = string("op_919_to_fp16"), val = fp16(0x1p-3)]; tensor scores_7_cast_fp16 = mul(x = var_918_cast_fp16, y = var_919_to_fp16)[name = string("scores_7_cast_fp16")]; tensor scores_9_cast_fp16 = add(x = scores_7_cast_fp16, y = attention_mask_to_fp16)[name = string("scores_9_cast_fp16")]; int32 var_928 = const()[name = string("op_928"), val = int32(-1)]; tensor var_930_cast_fp16 = softmax(axis = var_928, x = scores_9_cast_fp16)[name = string("op_930_cast_fp16")]; bool output_3_transpose_x_0 = const()[name = string("output_3_transpose_x_0"), val = bool(false)]; bool output_3_transpose_y_0 = const()[name = string("output_3_transpose_y_0"), val = bool(false)]; tensor value_7_cast_fp16 = transpose(perm = var_913, x = value_5_cast_fp16)[name = string("transpose_157")]; tensor output_3_cast_fp16 = matmul(transpose_x = output_3_transpose_x_0, transpose_y = output_3_transpose_y_0, x = var_930_cast_fp16, y = value_7_cast_fp16)[name = string("output_3_cast_fp16")]; tensor var_941 = const()[name = string("op_941"), val = tensor([0, 2, 1, 3])]; tensor var_946 = const()[name = string("op_946"), val = tensor([1, 390, 896])]; tensor var_942_cast_fp16 = transpose(perm = var_941, x = output_3_cast_fp16)[name = string("transpose_154")]; tensor input_27_cast_fp16 = reshape(shape = var_946, x = var_942_cast_fp16)[name = string("input_27_cast_fp16")]; tensor tower_layers_1_self_attn_out_proj_weight_to_fp16 = const()[name = string("tower_layers_1_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(46212480)))]; tensor tower_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = string("tower_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47818176)))]; tensor linear_10_cast_fp16 = linear(bias = tower_layers_1_self_attn_out_proj_bias_to_fp16, weight = tower_layers_1_self_attn_out_proj_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_10_cast_fp16")]; tensor input_29_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = string("input_29_cast_fp16")]; tensor input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor([-1])]; tensor tower_layers_1_final_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_1_final_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47820032)))]; tensor tower_layers_1_final_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_1_final_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47821888)))]; fp16 var_954_to_fp16 = const()[name = string("op_954_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = tower_layers_1_final_layer_norm_bias_to_fp16, epsilon = var_954_to_fp16, gamma = tower_layers_1_final_layer_norm_weight_to_fp16, x = input_29_cast_fp16)[name = string("input_31_cast_fp16")]; tensor tower_layers_1_fc1_weight_to_fp16 = const()[name = string("tower_layers_1_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47823744)))]; tensor tower_layers_1_fc1_bias_to_fp16 = const()[name = string("tower_layers_1_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54246336)))]; tensor linear_11_cast_fp16 = linear(bias = tower_layers_1_fc1_bias_to_fp16, weight = tower_layers_1_fc1_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_11_cast_fp16")]; string input_35_mode_0 = const()[name = string("input_35_mode_0"), val = string("EXACT")]; tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_11_cast_fp16)[name = string("input_35_cast_fp16")]; tensor tower_layers_1_fc2_weight_to_fp16 = const()[name = string("tower_layers_1_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54253568)))]; tensor tower_layers_1_fc2_bias_to_fp16 = const()[name = string("tower_layers_1_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60676160)))]; tensor linear_12_cast_fp16 = linear(bias = tower_layers_1_fc2_bias_to_fp16, weight = tower_layers_1_fc2_weight_to_fp16, x = input_35_cast_fp16)[name = string("linear_12_cast_fp16")]; tensor input_37_cast_fp16 = add(x = input_29_cast_fp16, y = linear_12_cast_fp16)[name = string("input_37_cast_fp16")]; tensor input_39_axes_0 = const()[name = string("input_39_axes_0"), val = tensor([-1])]; tensor tower_layers_2_self_attn_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_2_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60678016)))]; tensor tower_layers_2_self_attn_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_2_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60679872)))]; fp16 var_971_to_fp16 = const()[name = string("op_971_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = tower_layers_2_self_attn_layer_norm_bias_to_fp16, epsilon = var_971_to_fp16, gamma = tower_layers_2_self_attn_layer_norm_weight_to_fp16, x = input_37_cast_fp16)[name = string("input_39_cast_fp16")]; tensor tower_layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("tower_layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60681728)))]; tensor tower_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("tower_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62287424)))]; tensor linear_13_cast_fp16 = linear(bias = tower_layers_2_self_attn_q_proj_bias_to_fp16, weight = tower_layers_2_self_attn_q_proj_weight_to_fp16, x = input_39_cast_fp16)[name = string("linear_13_cast_fp16")]; tensor var_984 = const()[name = string("op_984"), val = tensor([1, 390, 14, 64])]; tensor query_9_cast_fp16 = reshape(shape = var_984, x = linear_13_cast_fp16)[name = string("query_9_cast_fp16")]; tensor tower_layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("tower_layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62289280)))]; tensor tower_layers_2_self_attn_k_proj_bias_to_fp16 = const()[name = string("tower_layers_2_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63894976)))]; tensor linear_14_cast_fp16 = linear(bias = tower_layers_2_self_attn_k_proj_bias_to_fp16, weight = tower_layers_2_self_attn_k_proj_weight_to_fp16, x = input_39_cast_fp16)[name = string("linear_14_cast_fp16")]; tensor var_993 = const()[name = string("op_993"), val = tensor([1, 390, 14, 64])]; tensor key_9_cast_fp16 = reshape(shape = var_993, x = linear_14_cast_fp16)[name = string("key_9_cast_fp16")]; tensor tower_layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("tower_layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63896832)))]; tensor tower_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("tower_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65502528)))]; tensor linear_15_cast_fp16 = linear(bias = tower_layers_2_self_attn_v_proj_bias_to_fp16, weight = tower_layers_2_self_attn_v_proj_weight_to_fp16, x = input_39_cast_fp16)[name = string("linear_15_cast_fp16")]; tensor var_1002 = const()[name = string("op_1002"), val = tensor([1, 390, 14, 64])]; tensor value_9_cast_fp16 = reshape(shape = var_1002, x = linear_15_cast_fp16)[name = string("value_9_cast_fp16")]; tensor var_1020 = const()[name = string("op_1020"), val = tensor([0, 2, 1, 3])]; bool var_1025_transpose_x_0 = const()[name = string("op_1025_transpose_x_0"), val = bool(false)]; bool var_1025_transpose_y_0 = const()[name = string("op_1025_transpose_y_0"), val = bool(false)]; tensor transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_59 = transpose(perm = transpose_59_perm_0, x = key_9_cast_fp16)[name = string("transpose_151")]; tensor transpose_58 = transpose(perm = transpose_58_perm_0, x = query_9_cast_fp16)[name = string("transpose_152")]; tensor var_1025_cast_fp16 = matmul(transpose_x = var_1025_transpose_x_0, transpose_y = var_1025_transpose_y_0, x = transpose_58, y = transpose_59)[name = string("op_1025_cast_fp16")]; fp16 var_1026_to_fp16 = const()[name = string("op_1026_to_fp16"), val = fp16(0x1p-3)]; tensor scores_13_cast_fp16 = mul(x = var_1025_cast_fp16, y = var_1026_to_fp16)[name = string("scores_13_cast_fp16")]; tensor scores_15_cast_fp16 = add(x = scores_13_cast_fp16, y = attention_mask_to_fp16)[name = string("scores_15_cast_fp16")]; int32 var_1035 = const()[name = string("op_1035"), val = int32(-1)]; tensor var_1037_cast_fp16 = softmax(axis = var_1035, x = scores_15_cast_fp16)[name = string("op_1037_cast_fp16")]; bool output_5_transpose_x_0 = const()[name = string("output_5_transpose_x_0"), val = bool(false)]; bool output_5_transpose_y_0 = const()[name = string("output_5_transpose_y_0"), val = bool(false)]; tensor value_11_cast_fp16 = transpose(perm = var_1020, x = value_9_cast_fp16)[name = string("transpose_153")]; tensor output_5_cast_fp16 = matmul(transpose_x = output_5_transpose_x_0, transpose_y = output_5_transpose_y_0, x = var_1037_cast_fp16, y = value_11_cast_fp16)[name = string("output_5_cast_fp16")]; tensor var_1048 = const()[name = string("op_1048"), val = tensor([0, 2, 1, 3])]; tensor var_1053 = const()[name = string("op_1053"), val = tensor([1, 390, 896])]; tensor var_1049_cast_fp16 = transpose(perm = var_1048, x = output_5_cast_fp16)[name = string("transpose_150")]; tensor input_41_cast_fp16 = reshape(shape = var_1053, x = var_1049_cast_fp16)[name = string("input_41_cast_fp16")]; tensor tower_layers_2_self_attn_out_proj_weight_to_fp16 = const()[name = string("tower_layers_2_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65504384)))]; tensor tower_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = string("tower_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67110080)))]; tensor linear_16_cast_fp16 = linear(bias = tower_layers_2_self_attn_out_proj_bias_to_fp16, weight = tower_layers_2_self_attn_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_16_cast_fp16")]; tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = string("input_43_cast_fp16")]; tensor input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor([-1])]; tensor tower_layers_2_final_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_2_final_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67111936)))]; tensor tower_layers_2_final_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_2_final_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67113792)))]; fp16 var_1061_to_fp16 = const()[name = string("op_1061_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = tower_layers_2_final_layer_norm_bias_to_fp16, epsilon = var_1061_to_fp16, gamma = tower_layers_2_final_layer_norm_weight_to_fp16, x = input_43_cast_fp16)[name = string("input_45_cast_fp16")]; tensor tower_layers_2_fc1_weight_to_fp16 = const()[name = string("tower_layers_2_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67115648)))]; tensor tower_layers_2_fc1_bias_to_fp16 = const()[name = string("tower_layers_2_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73538240)))]; tensor linear_17_cast_fp16 = linear(bias = tower_layers_2_fc1_bias_to_fp16, weight = tower_layers_2_fc1_weight_to_fp16, x = input_45_cast_fp16)[name = string("linear_17_cast_fp16")]; string input_49_mode_0 = const()[name = string("input_49_mode_0"), val = string("EXACT")]; tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_17_cast_fp16)[name = string("input_49_cast_fp16")]; tensor tower_layers_2_fc2_weight_to_fp16 = const()[name = string("tower_layers_2_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73545472)))]; tensor tower_layers_2_fc2_bias_to_fp16 = const()[name = string("tower_layers_2_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79968064)))]; tensor linear_18_cast_fp16 = linear(bias = tower_layers_2_fc2_bias_to_fp16, weight = tower_layers_2_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_18_cast_fp16")]; tensor input_51_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = string("input_51_cast_fp16")]; tensor input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor([-1])]; tensor tower_layers_3_self_attn_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_3_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79969920)))]; tensor tower_layers_3_self_attn_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_3_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79971776)))]; fp16 var_1078_to_fp16 = const()[name = string("op_1078_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = tower_layers_3_self_attn_layer_norm_bias_to_fp16, epsilon = var_1078_to_fp16, gamma = tower_layers_3_self_attn_layer_norm_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")]; tensor tower_layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("tower_layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79973632)))]; tensor tower_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("tower_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81579328)))]; tensor linear_19_cast_fp16 = linear(bias = tower_layers_3_self_attn_q_proj_bias_to_fp16, weight = tower_layers_3_self_attn_q_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_19_cast_fp16")]; tensor var_1091 = const()[name = string("op_1091"), val = tensor([1, 390, 14, 64])]; tensor query_13_cast_fp16 = reshape(shape = var_1091, x = linear_19_cast_fp16)[name = string("query_13_cast_fp16")]; tensor tower_layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("tower_layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81581184)))]; tensor tower_layers_3_self_attn_k_proj_bias_to_fp16 = const()[name = string("tower_layers_3_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83186880)))]; tensor linear_20_cast_fp16 = linear(bias = tower_layers_3_self_attn_k_proj_bias_to_fp16, weight = tower_layers_3_self_attn_k_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_20_cast_fp16")]; tensor var_1100 = const()[name = string("op_1100"), val = tensor([1, 390, 14, 64])]; tensor key_13_cast_fp16 = reshape(shape = var_1100, x = linear_20_cast_fp16)[name = string("key_13_cast_fp16")]; tensor tower_layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("tower_layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83188736)))]; tensor tower_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("tower_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84794432)))]; tensor linear_21_cast_fp16 = linear(bias = tower_layers_3_self_attn_v_proj_bias_to_fp16, weight = tower_layers_3_self_attn_v_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_21_cast_fp16")]; tensor var_1109 = const()[name = string("op_1109"), val = tensor([1, 390, 14, 64])]; tensor value_13_cast_fp16 = reshape(shape = var_1109, x = linear_21_cast_fp16)[name = string("value_13_cast_fp16")]; tensor var_1127 = const()[name = string("op_1127"), val = tensor([0, 2, 1, 3])]; bool var_1132_transpose_x_0 = const()[name = string("op_1132_transpose_x_0"), val = bool(false)]; bool var_1132_transpose_y_0 = const()[name = string("op_1132_transpose_y_0"), val = bool(false)]; tensor transpose_60_perm_0 = const()[name = string("transpose_60_perm_0"), val = tensor([0, 2, -3, -1])]; tensor transpose_61_perm_0 = const()[name = string("transpose_61_perm_0"), val = tensor([0, 2, -1, -3])]; tensor transpose_61 = transpose(perm = transpose_61_perm_0, x = key_13_cast_fp16)[name = string("transpose_147")]; tensor transpose_60 = transpose(perm = transpose_60_perm_0, x = query_13_cast_fp16)[name = string("transpose_148")]; tensor var_1132_cast_fp16 = matmul(transpose_x = var_1132_transpose_x_0, transpose_y = var_1132_transpose_y_0, x = transpose_60, y = transpose_61)[name = string("op_1132_cast_fp16")]; fp16 var_1133_to_fp16 = const()[name = string("op_1133_to_fp16"), val = fp16(0x1p-3)]; tensor scores_19_cast_fp16 = mul(x = var_1132_cast_fp16, y = var_1133_to_fp16)[name = string("scores_19_cast_fp16")]; tensor scores_21_cast_fp16 = add(x = scores_19_cast_fp16, y = attention_mask_to_fp16)[name = string("scores_21_cast_fp16")]; int32 var_1142 = const()[name = string("op_1142"), val = int32(-1)]; tensor var_1144_cast_fp16 = softmax(axis = var_1142, x = scores_21_cast_fp16)[name = string("op_1144_cast_fp16")]; bool output_7_transpose_x_0 = const()[name = string("output_7_transpose_x_0"), val = bool(false)]; bool output_7_transpose_y_0 = const()[name = string("output_7_transpose_y_0"), val = bool(false)]; tensor value_15_cast_fp16 = transpose(perm = var_1127, x = value_13_cast_fp16)[name = string("transpose_149")]; tensor output_7_cast_fp16 = matmul(transpose_x = output_7_transpose_x_0, transpose_y = output_7_transpose_y_0, x = var_1144_cast_fp16, y = value_15_cast_fp16)[name = string("output_7_cast_fp16")]; tensor var_1155 = const()[name = string("op_1155"), val = tensor([0, 2, 1, 3])]; tensor var_1160 = const()[name = string("op_1160"), val = tensor([1, 390, 896])]; tensor var_1156_cast_fp16 = transpose(perm = var_1155, x = output_7_cast_fp16)[name = string("transpose_146")]; tensor input_55_cast_fp16 = reshape(shape = var_1160, x = var_1156_cast_fp16)[name = string("input_55_cast_fp16")]; tensor tower_layers_3_self_attn_out_proj_weight_to_fp16 = const()[name = string("tower_layers_3_self_attn_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84796288)))]; tensor tower_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = string("tower_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86401984)))]; tensor linear_22_cast_fp16 = linear(bias = tower_layers_3_self_attn_out_proj_bias_to_fp16, weight = tower_layers_3_self_attn_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = string("linear_22_cast_fp16")]; tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_22_cast_fp16)[name = string("input_57_cast_fp16")]; tensor input_59_axes_0 = const()[name = string("input_59_axes_0"), val = tensor([-1])]; tensor tower_layers_3_final_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_3_final_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86403840)))]; tensor tower_layers_3_final_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_3_final_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86405696)))]; fp16 var_1168_to_fp16 = const()[name = string("op_1168_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = tower_layers_3_final_layer_norm_bias_to_fp16, epsilon = var_1168_to_fp16, gamma = tower_layers_3_final_layer_norm_weight_to_fp16, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")]; tensor tower_layers_3_fc1_weight_to_fp16 = const()[name = string("tower_layers_3_fc1_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86407552)))]; tensor tower_layers_3_fc1_bias_to_fp16 = const()[name = string("tower_layers_3_fc1_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92830144)))]; tensor linear_23_cast_fp16 = linear(bias = tower_layers_3_fc1_bias_to_fp16, weight = tower_layers_3_fc1_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_23_cast_fp16")]; string input_63_mode_0 = const()[name = string("input_63_mode_0"), val = string("EXACT")]; tensor input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = linear_23_cast_fp16)[name = string("input_63_cast_fp16")]; tensor tower_layers_3_fc2_weight_to_fp16 = const()[name = string("tower_layers_3_fc2_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92837376)))]; tensor tower_layers_3_fc2_bias_to_fp16 = const()[name = string("tower_layers_3_fc2_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99259968)))]; tensor linear_24_cast_fp16 = linear(bias = tower_layers_3_fc2_bias_to_fp16, weight = tower_layers_3_fc2_weight_to_fp16, x = input_63_cast_fp16)[name = string("linear_24_cast_fp16")]; tensor input_65_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = string("input_65_cast_fp16")]; tensor input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor([-1])]; tensor tower_layers_4_self_attn_layer_norm_weight_to_fp16 = const()[name = string("tower_layers_4_self_attn_layer_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99261824)))]; tensor tower_layers_4_self_attn_layer_norm_bias_to_fp16 = const()[name = string("tower_layers_4_self_attn_layer_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99263680)))]; fp16 var_1185_to_fp16 = const()[name = string("op_1185_to_fp16"), val = fp16(0x1.5p-17)]; tensor input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = tower_layers_4_self_attn_layer_norm_bias_to_fp16, epsilon = var_1185_to_fp16, gamma = tower_layers_4_self_attn_layer_norm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")]; tensor tower_layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("tower_layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99265536)))]; tensor tower_layers_4_self_attn_q_proj_bias_to_fp16 = const()[name = string("tower_layers_4_self_attn_q_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100871232)))]; tensor linear_25_cast_fp16 = linear(bias = tower_layers_4_self_attn_q_proj_bias_to_fp16, weight = tower_layers_4_self_attn_q_proj_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_25_cast_fp16")]; tensor var_1198 = const()[name = string("op_1198"), val = tensor([1, 390, 14, 64])]; tensor query_17_cast_fp16 = reshape(shape = var_1198, x = linear_25_cast_fp16)[name = string("query_17_cast_fp16")]; tensor tower_layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("tower_layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100873088)))]; tensor tower_layers_4_self_attn_k_proj_bias_to_fp16 = const()[name = string("tower_layers_4_self_attn_k_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102478784)))]; tensor linear_26_cast_fp16 = linear(bias = tower_layers_4_self_attn_k_proj_bias_to_fp16, weight = tower_layers_4_self_attn_k_proj_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_26_cast_fp16")]; tensor var_1207 = const()[name = string("op_1207"), val = tensor([1, 390, 14, 64])]; tensor key_17_cast_fp16 = reshape(shape = var_1207, x = linear_26_cast_fp16)[name = string("key_17_cast_fp16")]; tensor tower_layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("tower_layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102480640)))]; tensor tower_layers_4_self_attn_v_proj_bias_to_fp16 = const()[name = string("tower_layers_4_self_attn_v_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104086336)))]; tensor linear_27_cast_fp16 = linear(bias = tower_layers_4_self_attn_v_proj_bias_to_fp16, weight = tower_layers_4_self_attn_v_proj_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_27_cast_fp16")]; tensor