| program(1.0) | |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3405.2.1"}, {"coremlc-version", "3404.23.1"}, {"coremltools-component-torch", "2.6.0+cu124"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})] | |
| { | |
| func main<ios17>(tensor<int32, [1, 77]> input_text) { | |
| tensor<int32, []> var_18 = const()[name = tensor<string, []>("op_18"), val = tensor<int32, []>(-1)]; | |
| tensor<int32, []> token_emb_1_axis_0 = const()[name = tensor<string, []>("token_emb_1_axis_0"), val = tensor<int32, []>(0)]; | |
| tensor<int32, []> token_emb_1_batch_dims_0 = const()[name = tensor<string, []>("token_emb_1_batch_dims_0"), val = tensor<int32, []>(0)]; | |
| tensor<bool, []> token_emb_1_validate_indices_0 = const()[name = tensor<string, []>("token_emb_1_validate_indices_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [49408, 512]> text_encoder_embedding_layer_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_embedding_layer_weight_to_fp16"), val = tensor<fp16, [49408, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))]; | |
| tensor<fp16, [1, 77, 512]> token_emb_1_cast_fp16 = gather(axis = token_emb_1_axis_0, batch_dims = token_emb_1_batch_dims_0, indices = input_text, validate_indices = token_emb_1_validate_indices_0, x = text_encoder_embedding_layer_weight_to_fp16)[name = tensor<string, []>("token_emb_1_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> const_1_to_fp16 = const()[name = tensor<string, []>("const_1_to_fp16"), val = tensor<fp16, [1, 77, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50593920)))]; | |
| tensor<fp16, [1, 77, 512]> input_1_cast_fp16 = add(x = token_emb_1_cast_fp16, y = const_1_to_fp16)[name = tensor<string, []>("input_1_cast_fp16")]; | |
| tensor<int32, [1]> var_67_axes_0 = const()[name = tensor<string, []>("op_67_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50672832)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50673920)))]; | |
| tensor<fp16, []> var_5_to_fp16 = const()[name = tensor<string, []>("op_5_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 77, 512]> var_67_cast_fp16 = layer_norm(axes = var_67_axes_0, beta = text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16, x = input_1_cast_fp16)[name = tensor<string, []>("op_67_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50675008)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52247936)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_0_cast_fp16 = linear(bias = text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_67_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")]; | |
| tensor<int32, [5]> var_79 = const()[name = tensor<string, []>("op_79"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_1_cast_fp16 = reshape(shape = var_79, x = linear_0_cast_fp16)[name = tensor<string, []>("qkv_1_cast_fp16")]; | |
| tensor<int32, [5]> var_81_perm_0 = const()[name = tensor<string, []>("op_81_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_1_begin_0 = const()[name = tensor<string, []>("query_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_1_end_0 = const()[name = tensor<string, []>("query_1_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_1_end_mask_0 = const()[name = tensor<string, []>("query_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_1_squeeze_mask_0 = const()[name = tensor<string, []>("query_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_81_cast_fp16 = transpose(perm = var_81_perm_0, x = qkv_1_cast_fp16)[name = tensor<string, []>("transpose_25")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_1_cast_fp16 = slice_by_index(begin = query_1_begin_0, end = query_1_end_0, end_mask = query_1_end_mask_0, squeeze_mask = query_1_squeeze_mask_0, x = var_81_cast_fp16)[name = tensor<string, []>("query_1_cast_fp16")]; | |
| tensor<int32, [5]> key_1_begin_0 = const()[name = tensor<string, []>("key_1_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_1_end_0 = const()[name = tensor<string, []>("key_1_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_1_end_mask_0 = const()[name = tensor<string, []>("key_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_1_squeeze_mask_0 = const()[name = tensor<string, []>("key_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_1_cast_fp16 = slice_by_index(begin = key_1_begin_0, end = key_1_end_0, end_mask = key_1_end_mask_0, squeeze_mask = key_1_squeeze_mask_0, x = var_81_cast_fp16)[name = tensor<string, []>("key_1_cast_fp16")]; | |
| tensor<int32, [5]> value_1_begin_0 = const()[name = tensor<string, []>("value_1_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_1_end_0 = const()[name = tensor<string, []>("value_1_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_1_end_mask_0 = const()[name = tensor<string, []>("value_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_1_squeeze_mask_0 = const()[name = tensor<string, []>("value_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_1_cast_fp16 = slice_by_index(begin = value_1_begin_0, end = value_1_end_0, end_mask = value_1_end_mask_0, squeeze_mask = value_1_squeeze_mask_0, x = var_81_cast_fp16)[name = tensor<string, []>("value_1_cast_fp16")]; | |
| tensor<fp16, []> var_92_to_fp16 = const()[name = tensor<string, []>("op_92_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_3_cast_fp16 = mul(x = query_1_cast_fp16, y = var_92_to_fp16)[name = tensor<string, []>("query_3_cast_fp16")]; | |
| tensor<bool, []> attn_1_transpose_x_1 = const()[name = tensor<string, []>("attn_1_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_1_transpose_y_1 = const()[name = tensor<string, []>("attn_1_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_1, transpose_y = attn_1_transpose_y_1, x = query_3_cast_fp16, y = key_1_cast_fp16)[name = tensor<string, []>("attn_1_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_1_cast_fp16 = softmax(axis = var_18, x = attn_1_cast_fp16)[name = tensor<string, []>("attn_as_float_1_cast_fp16")]; | |
| tensor<bool, []> out_1_transpose_x_0 = const()[name = tensor<string, []>("out_1_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_1_transpose_y_0 = const()[name = tensor<string, []>("out_1_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_as_float_1_cast_fp16, y = value_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")]; | |
| tensor<int32, [4]> var_101_perm_0 = const()[name = tensor<string, []>("op_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_102 = const()[name = tensor<string, []>("op_102"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_101_cast_fp16 = transpose(perm = var_101_perm_0, x = out_1_cast_fp16)[name = tensor<string, []>("transpose_24")]; | |
| tensor<fp16, [1, 77, 512]> input_9_cast_fp16 = reshape(shape = var_102, x = var_101_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52251072)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52775424)))]; | |
| tensor<fp16, [1, 77, 512]> linear_1_cast_fp16 = linear(bias = text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_9_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_5_cast_fp16 = add(x = linear_1_cast_fp16, y = input_1_cast_fp16)[name = tensor<string, []>("x_5_cast_fp16")]; | |
| tensor<int32, [1]> var_116_axes_0 = const()[name = tensor<string, []>("op_116_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52776512)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52777600)))]; | |
| tensor<fp16, [1, 77, 512]> var_116_cast_fp16 = layer_norm(axes = var_116_axes_0, beta = text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16, x = x_5_cast_fp16)[name = tensor<string, []>("op_116_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(52778688)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54875904)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_2_cast_fp16 = linear(bias = text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16, x = var_116_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")]; | |
| tensor<string, []> input_19_mode_0 = const()[name = tensor<string, []>("input_19_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_2_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(54880064)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56977280)))]; | |
| tensor<fp16, [1, 77, 512]> linear_3_cast_fp16 = linear(bias = text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_7_cast_fp16 = add(x = x_5_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("x_7_cast_fp16")]; | |
| tensor<int32, [1]> var_143_axes_0 = const()[name = tensor<string, []>("op_143_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56978368)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56979456)))]; | |
| tensor<fp16, [1, 77, 512]> var_143_cast_fp16 = layer_norm(axes = var_143_axes_0, beta = text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16, x = x_7_cast_fp16)[name = tensor<string, []>("op_143_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(56980544)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58553472)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_4_cast_fp16 = linear(bias = text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_143_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")]; | |
| tensor<int32, [5]> var_155 = const()[name = tensor<string, []>("op_155"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_5_cast_fp16 = reshape(shape = var_155, x = linear_4_cast_fp16)[name = tensor<string, []>("qkv_5_cast_fp16")]; | |
| tensor<int32, [5]> var_157_perm_0 = const()[name = tensor<string, []>("op_157_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_5_begin_0 = const()[name = tensor<string, []>("query_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_5_end_0 = const()[name = tensor<string, []>("query_5_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_5_end_mask_0 = const()[name = tensor<string, []>("query_5_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_5_squeeze_mask_0 = const()[name = tensor<string, []>("query_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_157_cast_fp16 = transpose(perm = var_157_perm_0, x = qkv_5_cast_fp16)[name = tensor<string, []>("transpose_23")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_5_cast_fp16 = slice_by_index(begin = query_5_begin_0, end = query_5_end_0, end_mask = query_5_end_mask_0, squeeze_mask = query_5_squeeze_mask_0, x = var_157_cast_fp16)[name = tensor<string, []>("query_5_cast_fp16")]; | |
| tensor<int32, [5]> key_5_begin_0 = const()[name = tensor<string, []>("key_5_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_5_end_0 = const()[name = tensor<string, []>("key_5_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_5_end_mask_0 = const()[name = tensor<string, []>("key_5_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_5_squeeze_mask_0 = const()[name = tensor<string, []>("key_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_5_cast_fp16 = slice_by_index(begin = key_5_begin_0, end = key_5_end_0, end_mask = key_5_end_mask_0, squeeze_mask = key_5_squeeze_mask_0, x = var_157_cast_fp16)[name = tensor<string, []>("key_5_cast_fp16")]; | |
| tensor<int32, [5]> value_3_begin_0 = const()[name = tensor<string, []>("value_3_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_3_end_0 = const()[name = tensor<string, []>("value_3_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_3_end_mask_0 = const()[name = tensor<string, []>("value_3_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_3_squeeze_mask_0 = const()[name = tensor<string, []>("value_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_3_cast_fp16 = slice_by_index(begin = value_3_begin_0, end = value_3_end_0, end_mask = value_3_end_mask_0, squeeze_mask = value_3_squeeze_mask_0, x = var_157_cast_fp16)[name = tensor<string, []>("value_3_cast_fp16")]; | |
| tensor<fp16, []> var_168_to_fp16 = const()[name = tensor<string, []>("op_168_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_7_cast_fp16 = mul(x = query_5_cast_fp16, y = var_168_to_fp16)[name = tensor<string, []>("query_7_cast_fp16")]; | |
| tensor<bool, []> attn_5_transpose_x_1 = const()[name = tensor<string, []>("attn_5_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_5_transpose_y_1 = const()[name = tensor<string, []>("attn_5_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_1, transpose_y = attn_5_transpose_y_1, x = query_7_cast_fp16, y = key_5_cast_fp16)[name = tensor<string, []>("attn_5_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_3_cast_fp16 = softmax(axis = var_18, x = attn_5_cast_fp16)[name = tensor<string, []>("attn_as_float_3_cast_fp16")]; | |
| tensor<bool, []> out_3_transpose_x_0 = const()[name = tensor<string, []>("out_3_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_3_transpose_y_0 = const()[name = tensor<string, []>("out_3_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_3_cast_fp16 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_as_float_3_cast_fp16, y = value_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")]; | |
| tensor<int32, [4]> var_177_perm_0 = const()[name = tensor<string, []>("op_177_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_178 = const()[name = tensor<string, []>("op_178"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_177_cast_fp16 = transpose(perm = var_177_perm_0, x = out_3_cast_fp16)[name = tensor<string, []>("transpose_22")]; | |
| tensor<fp16, [1, 77, 512]> input_31_cast_fp16 = reshape(shape = var_178, x = var_177_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(58556608)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59080960)))]; | |
| tensor<fp16, [1, 77, 512]> linear_5_cast_fp16 = linear(bias = text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_11_cast_fp16 = add(x = linear_5_cast_fp16, y = x_7_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")]; | |
| tensor<int32, [1]> var_192_axes_0 = const()[name = tensor<string, []>("op_192_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59082048)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59083136)))]; | |
| tensor<fp16, [1, 77, 512]> var_192_cast_fp16 = layer_norm(axes = var_192_axes_0, beta = text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16, x = x_11_cast_fp16)[name = tensor<string, []>("op_192_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(59084224)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61181440)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_6_cast_fp16 = linear(bias = text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16, x = var_192_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")]; | |
| tensor<string, []> input_41_mode_0 = const()[name = tensor<string, []>("input_41_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_6_cast_fp16)[name = tensor<string, []>("input_41_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61185600)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63282816)))]; | |
| tensor<fp16, [1, 77, 512]> linear_7_cast_fp16 = linear(bias = text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16, x = input_41_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_7_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")]; | |
| tensor<int32, [1]> var_219_axes_0 = const()[name = tensor<string, []>("op_219_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63283904)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63284992)))]; | |
| tensor<fp16, [1, 77, 512]> var_219_cast_fp16 = layer_norm(axes = var_219_axes_0, beta = text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16, x = x_13_cast_fp16)[name = tensor<string, []>("op_219_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63286080)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64859008)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_8_cast_fp16 = linear(bias = text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_219_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")]; | |
| tensor<int32, [5]> var_231 = const()[name = tensor<string, []>("op_231"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_9_cast_fp16 = reshape(shape = var_231, x = linear_8_cast_fp16)[name = tensor<string, []>("qkv_9_cast_fp16")]; | |
| tensor<int32, [5]> var_233_perm_0 = const()[name = tensor<string, []>("op_233_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_9_begin_0 = const()[name = tensor<string, []>("query_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_9_end_0 = const()[name = tensor<string, []>("query_9_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_9_end_mask_0 = const()[name = tensor<string, []>("query_9_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_9_squeeze_mask_0 = const()[name = tensor<string, []>("query_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_233_cast_fp16 = transpose(perm = var_233_perm_0, x = qkv_9_cast_fp16)[name = tensor<string, []>("transpose_21")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_9_cast_fp16 = slice_by_index(begin = query_9_begin_0, end = query_9_end_0, end_mask = query_9_end_mask_0, squeeze_mask = query_9_squeeze_mask_0, x = var_233_cast_fp16)[name = tensor<string, []>("query_9_cast_fp16")]; | |
| tensor<int32, [5]> key_9_begin_0 = const()[name = tensor<string, []>("key_9_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_9_end_0 = const()[name = tensor<string, []>("key_9_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_9_end_mask_0 = const()[name = tensor<string, []>("key_9_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_9_squeeze_mask_0 = const()[name = tensor<string, []>("key_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_9_cast_fp16 = slice_by_index(begin = key_9_begin_0, end = key_9_end_0, end_mask = key_9_end_mask_0, squeeze_mask = key_9_squeeze_mask_0, x = var_233_cast_fp16)[name = tensor<string, []>("key_9_cast_fp16")]; | |
| tensor<int32, [5]> value_5_begin_0 = const()[name = tensor<string, []>("value_5_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_5_end_0 = const()[name = tensor<string, []>("value_5_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_5_end_mask_0 = const()[name = tensor<string, []>("value_5_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_5_squeeze_mask_0 = const()[name = tensor<string, []>("value_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_5_cast_fp16 = slice_by_index(begin = value_5_begin_0, end = value_5_end_0, end_mask = value_5_end_mask_0, squeeze_mask = value_5_squeeze_mask_0, x = var_233_cast_fp16)[name = tensor<string, []>("value_5_cast_fp16")]; | |
| tensor<fp16, []> var_244_to_fp16 = const()[name = tensor<string, []>("op_244_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_11_cast_fp16 = mul(x = query_9_cast_fp16, y = var_244_to_fp16)[name = tensor<string, []>("query_11_cast_fp16")]; | |
| tensor<bool, []> attn_9_transpose_x_1 = const()[name = tensor<string, []>("attn_9_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_9_transpose_y_1 = const()[name = tensor<string, []>("attn_9_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_1, transpose_y = attn_9_transpose_y_1, x = query_11_cast_fp16, y = key_9_cast_fp16)[name = tensor<string, []>("attn_9_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_5_cast_fp16 = softmax(axis = var_18, x = attn_9_cast_fp16)[name = tensor<string, []>("attn_as_float_5_cast_fp16")]; | |
| tensor<bool, []> out_5_transpose_x_0 = const()[name = tensor<string, []>("out_5_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_5_transpose_y_0 = const()[name = tensor<string, []>("out_5_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_as_float_5_cast_fp16, y = value_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")]; | |
| tensor<int32, [4]> var_253_perm_0 = const()[name = tensor<string, []>("op_253_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_254 = const()[name = tensor<string, []>("op_254"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_253_cast_fp16 = transpose(perm = var_253_perm_0, x = out_5_cast_fp16)[name = tensor<string, []>("transpose_20")]; | |
| tensor<fp16, [1, 77, 512]> input_53_cast_fp16 = reshape(shape = var_254, x = var_253_cast_fp16)[name = tensor<string, []>("input_53_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64862144)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65386496)))]; | |
| tensor<fp16, [1, 77, 512]> linear_9_cast_fp16 = linear(bias = text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_17_cast_fp16 = add(x = linear_9_cast_fp16, y = x_13_cast_fp16)[name = tensor<string, []>("x_17_cast_fp16")]; | |
| tensor<int32, [1]> var_268_axes_0 = const()[name = tensor<string, []>("op_268_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65387584)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65388672)))]; | |
| tensor<fp16, [1, 77, 512]> var_268_cast_fp16 = layer_norm(axes = var_268_axes_0, beta = text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16, x = x_17_cast_fp16)[name = tensor<string, []>("op_268_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65389760)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67486976)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_10_cast_fp16 = linear(bias = text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16, x = var_268_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")]; | |
| tensor<string, []> input_63_mode_0 = const()[name = tensor<string, []>("input_63_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_63_cast_fp16 = gelu(mode = input_63_mode_0, x = linear_10_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67491136)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69588352)))]; | |
| tensor<fp16, [1, 77, 512]> linear_11_cast_fp16 = linear(bias = text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16, x = input_63_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_19_cast_fp16 = add(x = x_17_cast_fp16, y = linear_11_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")]; | |
| tensor<int32, [1]> var_295_axes_0 = const()[name = tensor<string, []>("op_295_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69589440)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69590528)))]; | |
| tensor<fp16, [1, 77, 512]> var_295_cast_fp16 = layer_norm(axes = var_295_axes_0, beta = text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16, x = x_19_cast_fp16)[name = tensor<string, []>("op_295_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69591616)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71164544)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_12_cast_fp16 = linear(bias = text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_295_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")]; | |
| tensor<int32, [5]> var_307 = const()[name = tensor<string, []>("op_307"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_13_cast_fp16 = reshape(shape = var_307, x = linear_12_cast_fp16)[name = tensor<string, []>("qkv_13_cast_fp16")]; | |
| tensor<int32, [5]> var_309_perm_0 = const()[name = tensor<string, []>("op_309_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_13_begin_0 = const()[name = tensor<string, []>("query_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_13_end_0 = const()[name = tensor<string, []>("query_13_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_13_end_mask_0 = const()[name = tensor<string, []>("query_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_13_squeeze_mask_0 = const()[name = tensor<string, []>("query_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_309_cast_fp16 = transpose(perm = var_309_perm_0, x = qkv_13_cast_fp16)[name = tensor<string, []>("transpose_19")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_13_cast_fp16 = slice_by_index(begin = query_13_begin_0, end = query_13_end_0, end_mask = query_13_end_mask_0, squeeze_mask = query_13_squeeze_mask_0, x = var_309_cast_fp16)[name = tensor<string, []>("query_13_cast_fp16")]; | |
| tensor<int32, [5]> key_13_begin_0 = const()[name = tensor<string, []>("key_13_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_13_end_0 = const()[name = tensor<string, []>("key_13_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_13_end_mask_0 = const()[name = tensor<string, []>("key_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_13_squeeze_mask_0 = const()[name = tensor<string, []>("key_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_13_cast_fp16 = slice_by_index(begin = key_13_begin_0, end = key_13_end_0, end_mask = key_13_end_mask_0, squeeze_mask = key_13_squeeze_mask_0, x = var_309_cast_fp16)[name = tensor<string, []>("key_13_cast_fp16")]; | |
| tensor<int32, [5]> value_7_begin_0 = const()[name = tensor<string, []>("value_7_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_7_end_0 = const()[name = tensor<string, []>("value_7_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_7_end_mask_0 = const()[name = tensor<string, []>("value_7_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_7_squeeze_mask_0 = const()[name = tensor<string, []>("value_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_7_cast_fp16 = slice_by_index(begin = value_7_begin_0, end = value_7_end_0, end_mask = value_7_end_mask_0, squeeze_mask = value_7_squeeze_mask_0, x = var_309_cast_fp16)[name = tensor<string, []>("value_7_cast_fp16")]; | |
| tensor<fp16, []> var_320_to_fp16 = const()[name = tensor<string, []>("op_320_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_15_cast_fp16 = mul(x = query_13_cast_fp16, y = var_320_to_fp16)[name = tensor<string, []>("query_15_cast_fp16")]; | |
| tensor<bool, []> attn_13_transpose_x_1 = const()[name = tensor<string, []>("attn_13_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_13_transpose_y_1 = const()[name = tensor<string, []>("attn_13_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_1, transpose_y = attn_13_transpose_y_1, x = query_15_cast_fp16, y = key_13_cast_fp16)[name = tensor<string, []>("attn_13_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_7_cast_fp16 = softmax(axis = var_18, x = attn_13_cast_fp16)[name = tensor<string, []>("attn_as_float_7_cast_fp16")]; | |
| tensor<bool, []> out_7_transpose_x_0 = const()[name = tensor<string, []>("out_7_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_7_transpose_y_0 = const()[name = tensor<string, []>("out_7_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_7_cast_fp16 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_as_float_7_cast_fp16, y = value_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")]; | |
| tensor<int32, [4]> var_329_perm_0 = const()[name = tensor<string, []>("op_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_330 = const()[name = tensor<string, []>("op_330"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_329_cast_fp16 = transpose(perm = var_329_perm_0, x = out_7_cast_fp16)[name = tensor<string, []>("transpose_18")]; | |
| tensor<fp16, [1, 77, 512]> input_75_cast_fp16 = reshape(shape = var_330, x = var_329_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71167680)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71692032)))]; | |
| tensor<fp16, [1, 77, 512]> linear_13_cast_fp16 = linear(bias = text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_23_cast_fp16 = add(x = linear_13_cast_fp16, y = x_19_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")]; | |
| tensor<int32, [1]> var_344_axes_0 = const()[name = tensor<string, []>("op_344_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71693120)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71694208)))]; | |
| tensor<fp16, [1, 77, 512]> var_344_cast_fp16 = layer_norm(axes = var_344_axes_0, beta = text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16, x = x_23_cast_fp16)[name = tensor<string, []>("op_344_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(71695296)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73792512)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_14_cast_fp16 = linear(bias = text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16, x = var_344_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")]; | |
| tensor<string, []> input_85_mode_0 = const()[name = tensor<string, []>("input_85_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_14_cast_fp16)[name = tensor<string, []>("input_85_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73796672)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75893888)))]; | |
| tensor<fp16, [1, 77, 512]> linear_15_cast_fp16 = linear(bias = text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16, x = input_85_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_15_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")]; | |
| tensor<int32, [1]> var_371_axes_0 = const()[name = tensor<string, []>("op_371_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75894976)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75896064)))]; | |
| tensor<fp16, [1, 77, 512]> var_371_cast_fp16 = layer_norm(axes = var_371_axes_0, beta = text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16, x = x_25_cast_fp16)[name = tensor<string, []>("op_371_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(75897152)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77470080)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_16_cast_fp16 = linear(bias = text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_371_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")]; | |
| tensor<int32, [5]> var_383 = const()[name = tensor<string, []>("op_383"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_17_cast_fp16 = reshape(shape = var_383, x = linear_16_cast_fp16)[name = tensor<string, []>("qkv_17_cast_fp16")]; | |
| tensor<int32, [5]> var_385_perm_0 = const()[name = tensor<string, []>("op_385_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_17_begin_0 = const()[name = tensor<string, []>("query_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_17_end_0 = const()[name = tensor<string, []>("query_17_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_17_end_mask_0 = const()[name = tensor<string, []>("query_17_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_17_squeeze_mask_0 = const()[name = tensor<string, []>("query_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_385_cast_fp16 = transpose(perm = var_385_perm_0, x = qkv_17_cast_fp16)[name = tensor<string, []>("transpose_17")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_17_cast_fp16 = slice_by_index(begin = query_17_begin_0, end = query_17_end_0, end_mask = query_17_end_mask_0, squeeze_mask = query_17_squeeze_mask_0, x = var_385_cast_fp16)[name = tensor<string, []>("query_17_cast_fp16")]; | |
| tensor<int32, [5]> key_17_begin_0 = const()[name = tensor<string, []>("key_17_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_17_end_0 = const()[name = tensor<string, []>("key_17_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_17_end_mask_0 = const()[name = tensor<string, []>("key_17_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_17_squeeze_mask_0 = const()[name = tensor<string, []>("key_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_17_cast_fp16 = slice_by_index(begin = key_17_begin_0, end = key_17_end_0, end_mask = key_17_end_mask_0, squeeze_mask = key_17_squeeze_mask_0, x = var_385_cast_fp16)[name = tensor<string, []>("key_17_cast_fp16")]; | |
| tensor<int32, [5]> value_9_begin_0 = const()[name = tensor<string, []>("value_9_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_9_end_0 = const()[name = tensor<string, []>("value_9_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_9_end_mask_0 = const()[name = tensor<string, []>("value_9_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_9_squeeze_mask_0 = const()[name = tensor<string, []>("value_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_9_cast_fp16 = slice_by_index(begin = value_9_begin_0, end = value_9_end_0, end_mask = value_9_end_mask_0, squeeze_mask = value_9_squeeze_mask_0, x = var_385_cast_fp16)[name = tensor<string, []>("value_9_cast_fp16")]; | |
| tensor<fp16, []> var_396_to_fp16 = const()[name = tensor<string, []>("op_396_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_19_cast_fp16 = mul(x = query_17_cast_fp16, y = var_396_to_fp16)[name = tensor<string, []>("query_19_cast_fp16")]; | |
| tensor<bool, []> attn_17_transpose_x_1 = const()[name = tensor<string, []>("attn_17_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_17_transpose_y_1 = const()[name = tensor<string, []>("attn_17_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_17_cast_fp16 = matmul(transpose_x = attn_17_transpose_x_1, transpose_y = attn_17_transpose_y_1, x = query_19_cast_fp16, y = key_17_cast_fp16)[name = tensor<string, []>("attn_17_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_9_cast_fp16 = softmax(axis = var_18, x = attn_17_cast_fp16)[name = tensor<string, []>("attn_as_float_9_cast_fp16")]; | |
| tensor<bool, []> out_9_transpose_x_0 = const()[name = tensor<string, []>("out_9_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_9_transpose_y_0 = const()[name = tensor<string, []>("out_9_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_as_float_9_cast_fp16, y = value_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")]; | |
| tensor<int32, [4]> var_405_perm_0 = const()[name = tensor<string, []>("op_405_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_406 = const()[name = tensor<string, []>("op_406"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_405_cast_fp16 = transpose(perm = var_405_perm_0, x = out_9_cast_fp16)[name = tensor<string, []>("transpose_16")]; | |
| tensor<fp16, [1, 77, 512]> input_97_cast_fp16 = reshape(shape = var_406, x = var_405_cast_fp16)[name = tensor<string, []>("input_97_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77473216)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77997568)))]; | |
| tensor<fp16, [1, 77, 512]> linear_17_cast_fp16 = linear(bias = text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_97_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_29_cast_fp16 = add(x = linear_17_cast_fp16, y = x_25_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")]; | |
| tensor<int32, [1]> var_420_axes_0 = const()[name = tensor<string, []>("op_420_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77998656)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77999744)))]; | |
| tensor<fp16, [1, 77, 512]> var_420_cast_fp16 = layer_norm(axes = var_420_axes_0, beta = text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16, x = x_29_cast_fp16)[name = tensor<string, []>("op_420_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78000832)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80098048)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_18_cast_fp16 = linear(bias = text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16, x = var_420_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")]; | |
| tensor<string, []> input_107_mode_0 = const()[name = tensor<string, []>("input_107_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_18_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(80102208)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82199424)))]; | |
| tensor<fp16, [1, 77, 512]> linear_19_cast_fp16 = linear(bias = text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16, x = input_107_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_31_cast_fp16 = add(x = x_29_cast_fp16, y = linear_19_cast_fp16)[name = tensor<string, []>("x_31_cast_fp16")]; | |
| tensor<int32, [1]> var_447_axes_0 = const()[name = tensor<string, []>("op_447_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82200512)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82201600)))]; | |
| tensor<fp16, [1, 77, 512]> var_447_cast_fp16 = layer_norm(axes = var_447_axes_0, beta = text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16, x = x_31_cast_fp16)[name = tensor<string, []>("op_447_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82202688)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83775616)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_20_cast_fp16 = linear(bias = text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_447_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")]; | |
| tensor<int32, [5]> var_459 = const()[name = tensor<string, []>("op_459"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_21_cast_fp16 = reshape(shape = var_459, x = linear_20_cast_fp16)[name = tensor<string, []>("qkv_21_cast_fp16")]; | |
| tensor<int32, [5]> var_461_perm_0 = const()[name = tensor<string, []>("op_461_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_21_begin_0 = const()[name = tensor<string, []>("query_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_21_end_0 = const()[name = tensor<string, []>("query_21_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_21_end_mask_0 = const()[name = tensor<string, []>("query_21_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_21_squeeze_mask_0 = const()[name = tensor<string, []>("query_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_461_cast_fp16 = transpose(perm = var_461_perm_0, x = qkv_21_cast_fp16)[name = tensor<string, []>("transpose_15")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_21_cast_fp16 = slice_by_index(begin = query_21_begin_0, end = query_21_end_0, end_mask = query_21_end_mask_0, squeeze_mask = query_21_squeeze_mask_0, x = var_461_cast_fp16)[name = tensor<string, []>("query_21_cast_fp16")]; | |
| tensor<int32, [5]> key_21_begin_0 = const()[name = tensor<string, []>("key_21_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_21_end_0 = const()[name = tensor<string, []>("key_21_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_21_end_mask_0 = const()[name = tensor<string, []>("key_21_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_21_squeeze_mask_0 = const()[name = tensor<string, []>("key_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_21_cast_fp16 = slice_by_index(begin = key_21_begin_0, end = key_21_end_0, end_mask = key_21_end_mask_0, squeeze_mask = key_21_squeeze_mask_0, x = var_461_cast_fp16)[name = tensor<string, []>("key_21_cast_fp16")]; | |
| tensor<int32, [5]> value_11_begin_0 = const()[name = tensor<string, []>("value_11_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_11_end_0 = const()[name = tensor<string, []>("value_11_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_11_end_mask_0 = const()[name = tensor<string, []>("value_11_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_11_squeeze_mask_0 = const()[name = tensor<string, []>("value_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_11_cast_fp16 = slice_by_index(begin = value_11_begin_0, end = value_11_end_0, end_mask = value_11_end_mask_0, squeeze_mask = value_11_squeeze_mask_0, x = var_461_cast_fp16)[name = tensor<string, []>("value_11_cast_fp16")]; | |
| tensor<fp16, []> var_472_to_fp16 = const()[name = tensor<string, []>("op_472_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_23_cast_fp16 = mul(x = query_21_cast_fp16, y = var_472_to_fp16)[name = tensor<string, []>("query_23_cast_fp16")]; | |
| tensor<bool, []> attn_21_transpose_x_1 = const()[name = tensor<string, []>("attn_21_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_21_transpose_y_1 = const()[name = tensor<string, []>("attn_21_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_21_cast_fp16 = matmul(transpose_x = attn_21_transpose_x_1, transpose_y = attn_21_transpose_y_1, x = query_23_cast_fp16, y = key_21_cast_fp16)[name = tensor<string, []>("attn_21_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_11_cast_fp16 = softmax(axis = var_18, x = attn_21_cast_fp16)[name = tensor<string, []>("attn_as_float_11_cast_fp16")]; | |
| tensor<bool, []> out_11_transpose_x_0 = const()[name = tensor<string, []>("out_11_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_11_transpose_y_0 = const()[name = tensor<string, []>("out_11_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_11_cast_fp16 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_as_float_11_cast_fp16, y = value_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")]; | |
| tensor<int32, [4]> var_481_perm_0 = const()[name = tensor<string, []>("op_481_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_482 = const()[name = tensor<string, []>("op_482"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_481_cast_fp16 = transpose(perm = var_481_perm_0, x = out_11_cast_fp16)[name = tensor<string, []>("transpose_14")]; | |
| tensor<fp16, [1, 77, 512]> input_119_cast_fp16 = reshape(shape = var_482, x = var_481_cast_fp16)[name = tensor<string, []>("input_119_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83778752)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84303104)))]; | |
| tensor<fp16, [1, 77, 512]> linear_21_cast_fp16 = linear(bias = text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_119_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_35_cast_fp16 = add(x = linear_21_cast_fp16, y = x_31_cast_fp16)[name = tensor<string, []>("x_35_cast_fp16")]; | |
| tensor<int32, [1]> var_496_axes_0 = const()[name = tensor<string, []>("op_496_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84304192)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84305280)))]; | |
| tensor<fp16, [1, 77, 512]> var_496_cast_fp16 = layer_norm(axes = var_496_axes_0, beta = text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16, x = x_35_cast_fp16)[name = tensor<string, []>("op_496_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84306368)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86403584)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_22_cast_fp16 = linear(bias = text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16, x = var_496_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")]; | |
| tensor<string, []> input_129_mode_0 = const()[name = tensor<string, []>("input_129_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = linear_22_cast_fp16)[name = tensor<string, []>("input_129_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86407744)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88504960)))]; | |
| tensor<fp16, [1, 77, 512]> linear_23_cast_fp16 = linear(bias = text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16, x = input_129_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_23_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")]; | |
| tensor<int32, [1]> var_523_axes_0 = const()[name = tensor<string, []>("op_523_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88506048)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88507136)))]; | |
| tensor<fp16, [1, 77, 512]> var_523_cast_fp16 = layer_norm(axes = var_523_axes_0, beta = text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16, x = x_37_cast_fp16)[name = tensor<string, []>("op_523_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88508224)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90081152)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_24_cast_fp16 = linear(bias = text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_523_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")]; | |
| tensor<int32, [5]> var_535 = const()[name = tensor<string, []>("op_535"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_25_cast_fp16 = reshape(shape = var_535, x = linear_24_cast_fp16)[name = tensor<string, []>("qkv_25_cast_fp16")]; | |
| tensor<int32, [5]> var_537_perm_0 = const()[name = tensor<string, []>("op_537_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_25_begin_0 = const()[name = tensor<string, []>("query_25_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_25_end_0 = const()[name = tensor<string, []>("query_25_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_25_end_mask_0 = const()[name = tensor<string, []>("query_25_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_25_squeeze_mask_0 = const()[name = tensor<string, []>("query_25_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_537_cast_fp16 = transpose(perm = var_537_perm_0, x = qkv_25_cast_fp16)[name = tensor<string, []>("transpose_13")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_25_cast_fp16 = slice_by_index(begin = query_25_begin_0, end = query_25_end_0, end_mask = query_25_end_mask_0, squeeze_mask = query_25_squeeze_mask_0, x = var_537_cast_fp16)[name = tensor<string, []>("query_25_cast_fp16")]; | |
| tensor<int32, [5]> key_25_begin_0 = const()[name = tensor<string, []>("key_25_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_25_end_0 = const()[name = tensor<string, []>("key_25_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_25_end_mask_0 = const()[name = tensor<string, []>("key_25_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_25_squeeze_mask_0 = const()[name = tensor<string, []>("key_25_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_25_cast_fp16 = slice_by_index(begin = key_25_begin_0, end = key_25_end_0, end_mask = key_25_end_mask_0, squeeze_mask = key_25_squeeze_mask_0, x = var_537_cast_fp16)[name = tensor<string, []>("key_25_cast_fp16")]; | |
| tensor<int32, [5]> value_13_begin_0 = const()[name = tensor<string, []>("value_13_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_13_end_0 = const()[name = tensor<string, []>("value_13_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_13_end_mask_0 = const()[name = tensor<string, []>("value_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_13_squeeze_mask_0 = const()[name = tensor<string, []>("value_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_13_cast_fp16 = slice_by_index(begin = value_13_begin_0, end = value_13_end_0, end_mask = value_13_end_mask_0, squeeze_mask = value_13_squeeze_mask_0, x = var_537_cast_fp16)[name = tensor<string, []>("value_13_cast_fp16")]; | |
| tensor<fp16, []> var_548_to_fp16 = const()[name = tensor<string, []>("op_548_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_27_cast_fp16 = mul(x = query_25_cast_fp16, y = var_548_to_fp16)[name = tensor<string, []>("query_27_cast_fp16")]; | |
| tensor<bool, []> attn_25_transpose_x_1 = const()[name = tensor<string, []>("attn_25_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_25_transpose_y_1 = const()[name = tensor<string, []>("attn_25_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_25_cast_fp16 = matmul(transpose_x = attn_25_transpose_x_1, transpose_y = attn_25_transpose_y_1, x = query_27_cast_fp16, y = key_25_cast_fp16)[name = tensor<string, []>("attn_25_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_13_cast_fp16 = softmax(axis = var_18, x = attn_25_cast_fp16)[name = tensor<string, []>("attn_as_float_13_cast_fp16")]; | |
| tensor<bool, []> out_13_transpose_x_0 = const()[name = tensor<string, []>("out_13_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_13_transpose_y_0 = const()[name = tensor<string, []>("out_13_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = attn_as_float_13_cast_fp16, y = value_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")]; | |
| tensor<int32, [4]> var_557_perm_0 = const()[name = tensor<string, []>("op_557_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_558 = const()[name = tensor<string, []>("op_558"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_557_cast_fp16 = transpose(perm = var_557_perm_0, x = out_13_cast_fp16)[name = tensor<string, []>("transpose_12")]; | |
| tensor<fp16, [1, 77, 512]> input_141_cast_fp16 = reshape(shape = var_558, x = var_557_cast_fp16)[name = tensor<string, []>("input_141_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90084288)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90608640)))]; | |
| tensor<fp16, [1, 77, 512]> linear_25_cast_fp16 = linear(bias = text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_141_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_41_cast_fp16 = add(x = linear_25_cast_fp16, y = x_37_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")]; | |
| tensor<int32, [1]> var_572_axes_0 = const()[name = tensor<string, []>("op_572_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90609728)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90610816)))]; | |
| tensor<fp16, [1, 77, 512]> var_572_cast_fp16 = layer_norm(axes = var_572_axes_0, beta = text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16, x = x_41_cast_fp16)[name = tensor<string, []>("op_572_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90611904)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92709120)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_26_cast_fp16 = linear(bias = text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16, x = var_572_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")]; | |
| tensor<string, []> input_151_mode_0 = const()[name = tensor<string, []>("input_151_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_26_cast_fp16)[name = tensor<string, []>("input_151_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92713280)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94810496)))]; | |
| tensor<fp16, [1, 77, 512]> linear_27_cast_fp16 = linear(bias = text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16, x = input_151_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_43_cast_fp16 = add(x = x_41_cast_fp16, y = linear_27_cast_fp16)[name = tensor<string, []>("x_43_cast_fp16")]; | |
| tensor<int32, [1]> var_599_axes_0 = const()[name = tensor<string, []>("op_599_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94811584)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94812672)))]; | |
| tensor<fp16, [1, 77, 512]> var_599_cast_fp16 = layer_norm(axes = var_599_axes_0, beta = text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16, x = x_43_cast_fp16)[name = tensor<string, []>("op_599_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(94813760)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96386688)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_28_cast_fp16 = linear(bias = text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_599_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")]; | |
| tensor<int32, [5]> var_611 = const()[name = tensor<string, []>("op_611"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_29_cast_fp16 = reshape(shape = var_611, x = linear_28_cast_fp16)[name = tensor<string, []>("qkv_29_cast_fp16")]; | |
| tensor<int32, [5]> var_613_perm_0 = const()[name = tensor<string, []>("op_613_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_29_begin_0 = const()[name = tensor<string, []>("query_29_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_29_end_0 = const()[name = tensor<string, []>("query_29_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_29_end_mask_0 = const()[name = tensor<string, []>("query_29_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_29_squeeze_mask_0 = const()[name = tensor<string, []>("query_29_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_613_cast_fp16 = transpose(perm = var_613_perm_0, x = qkv_29_cast_fp16)[name = tensor<string, []>("transpose_11")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_29_cast_fp16 = slice_by_index(begin = query_29_begin_0, end = query_29_end_0, end_mask = query_29_end_mask_0, squeeze_mask = query_29_squeeze_mask_0, x = var_613_cast_fp16)[name = tensor<string, []>("query_29_cast_fp16")]; | |
| tensor<int32, [5]> key_29_begin_0 = const()[name = tensor<string, []>("key_29_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_29_end_0 = const()[name = tensor<string, []>("key_29_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_29_end_mask_0 = const()[name = tensor<string, []>("key_29_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_29_squeeze_mask_0 = const()[name = tensor<string, []>("key_29_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_29_cast_fp16 = slice_by_index(begin = key_29_begin_0, end = key_29_end_0, end_mask = key_29_end_mask_0, squeeze_mask = key_29_squeeze_mask_0, x = var_613_cast_fp16)[name = tensor<string, []>("key_29_cast_fp16")]; | |
| tensor<int32, [5]> value_15_begin_0 = const()[name = tensor<string, []>("value_15_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_15_end_0 = const()[name = tensor<string, []>("value_15_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_15_end_mask_0 = const()[name = tensor<string, []>("value_15_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_15_squeeze_mask_0 = const()[name = tensor<string, []>("value_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_15_cast_fp16 = slice_by_index(begin = value_15_begin_0, end = value_15_end_0, end_mask = value_15_end_mask_0, squeeze_mask = value_15_squeeze_mask_0, x = var_613_cast_fp16)[name = tensor<string, []>("value_15_cast_fp16")]; | |
| tensor<fp16, []> var_624_to_fp16 = const()[name = tensor<string, []>("op_624_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_31_cast_fp16 = mul(x = query_29_cast_fp16, y = var_624_to_fp16)[name = tensor<string, []>("query_31_cast_fp16")]; | |
| tensor<bool, []> attn_29_transpose_x_1 = const()[name = tensor<string, []>("attn_29_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_29_transpose_y_1 = const()[name = tensor<string, []>("attn_29_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_29_cast_fp16 = matmul(transpose_x = attn_29_transpose_x_1, transpose_y = attn_29_transpose_y_1, x = query_31_cast_fp16, y = key_29_cast_fp16)[name = tensor<string, []>("attn_29_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_15_cast_fp16 = softmax(axis = var_18, x = attn_29_cast_fp16)[name = tensor<string, []>("attn_as_float_15_cast_fp16")]; | |
| tensor<bool, []> out_15_transpose_x_0 = const()[name = tensor<string, []>("out_15_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_15_transpose_y_0 = const()[name = tensor<string, []>("out_15_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_15_cast_fp16 = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = attn_as_float_15_cast_fp16, y = value_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")]; | |
| tensor<int32, [4]> var_633_perm_0 = const()[name = tensor<string, []>("op_633_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_634 = const()[name = tensor<string, []>("op_634"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_633_cast_fp16 = transpose(perm = var_633_perm_0, x = out_15_cast_fp16)[name = tensor<string, []>("transpose_10")]; | |
| tensor<fp16, [1, 77, 512]> input_163_cast_fp16 = reshape(shape = var_634, x = var_633_cast_fp16)[name = tensor<string, []>("input_163_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96389824)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96914176)))]; | |
| tensor<fp16, [1, 77, 512]> linear_29_cast_fp16 = linear(bias = text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_47_cast_fp16 = add(x = linear_29_cast_fp16, y = x_43_cast_fp16)[name = tensor<string, []>("x_47_cast_fp16")]; | |
| tensor<int32, [1]> var_648_axes_0 = const()[name = tensor<string, []>("op_648_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96915264)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96916352)))]; | |
| tensor<fp16, [1, 77, 512]> var_648_cast_fp16 = layer_norm(axes = var_648_axes_0, beta = text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16, x = x_47_cast_fp16)[name = tensor<string, []>("op_648_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(96917440)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99014656)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_30_cast_fp16 = linear(bias = text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16, x = var_648_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")]; | |
| tensor<string, []> input_173_mode_0 = const()[name = tensor<string, []>("input_173_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_173_cast_fp16 = gelu(mode = input_173_mode_0, x = linear_30_cast_fp16)[name = tensor<string, []>("input_173_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99018816)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101116032)))]; | |
| tensor<fp16, [1, 77, 512]> linear_31_cast_fp16 = linear(bias = text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16, x = input_173_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_31_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")]; | |
| tensor<int32, [1]> var_675_axes_0 = const()[name = tensor<string, []>("op_675_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101117120)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101118208)))]; | |
| tensor<fp16, [1, 77, 512]> var_675_cast_fp16 = layer_norm(axes = var_675_axes_0, beta = text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16, x = x_49_cast_fp16)[name = tensor<string, []>("op_675_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101119296)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102692224)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_32_cast_fp16 = linear(bias = text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_675_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")]; | |
| tensor<int32, [5]> var_687 = const()[name = tensor<string, []>("op_687"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_33_cast_fp16 = reshape(shape = var_687, x = linear_32_cast_fp16)[name = tensor<string, []>("qkv_33_cast_fp16")]; | |
| tensor<int32, [5]> var_689_perm_0 = const()[name = tensor<string, []>("op_689_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_33_begin_0 = const()[name = tensor<string, []>("query_33_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_33_end_0 = const()[name = tensor<string, []>("query_33_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_33_end_mask_0 = const()[name = tensor<string, []>("query_33_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_33_squeeze_mask_0 = const()[name = tensor<string, []>("query_33_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_689_cast_fp16 = transpose(perm = var_689_perm_0, x = qkv_33_cast_fp16)[name = tensor<string, []>("transpose_9")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_33_cast_fp16 = slice_by_index(begin = query_33_begin_0, end = query_33_end_0, end_mask = query_33_end_mask_0, squeeze_mask = query_33_squeeze_mask_0, x = var_689_cast_fp16)[name = tensor<string, []>("query_33_cast_fp16")]; | |
| tensor<int32, [5]> key_33_begin_0 = const()[name = tensor<string, []>("key_33_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_33_end_0 = const()[name = tensor<string, []>("key_33_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_33_end_mask_0 = const()[name = tensor<string, []>("key_33_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_33_squeeze_mask_0 = const()[name = tensor<string, []>("key_33_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_33_cast_fp16 = slice_by_index(begin = key_33_begin_0, end = key_33_end_0, end_mask = key_33_end_mask_0, squeeze_mask = key_33_squeeze_mask_0, x = var_689_cast_fp16)[name = tensor<string, []>("key_33_cast_fp16")]; | |
| tensor<int32, [5]> value_17_begin_0 = const()[name = tensor<string, []>("value_17_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_17_end_0 = const()[name = tensor<string, []>("value_17_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_17_end_mask_0 = const()[name = tensor<string, []>("value_17_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_17_squeeze_mask_0 = const()[name = tensor<string, []>("value_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_17_cast_fp16 = slice_by_index(begin = value_17_begin_0, end = value_17_end_0, end_mask = value_17_end_mask_0, squeeze_mask = value_17_squeeze_mask_0, x = var_689_cast_fp16)[name = tensor<string, []>("value_17_cast_fp16")]; | |
| tensor<fp16, []> var_700_to_fp16 = const()[name = tensor<string, []>("op_700_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_35_cast_fp16 = mul(x = query_33_cast_fp16, y = var_700_to_fp16)[name = tensor<string, []>("query_35_cast_fp16")]; | |
| tensor<bool, []> attn_33_transpose_x_1 = const()[name = tensor<string, []>("attn_33_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_33_transpose_y_1 = const()[name = tensor<string, []>("attn_33_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_33_cast_fp16 = matmul(transpose_x = attn_33_transpose_x_1, transpose_y = attn_33_transpose_y_1, x = query_35_cast_fp16, y = key_33_cast_fp16)[name = tensor<string, []>("attn_33_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_17_cast_fp16 = softmax(axis = var_18, x = attn_33_cast_fp16)[name = tensor<string, []>("attn_as_float_17_cast_fp16")]; | |
| tensor<bool, []> out_17_transpose_x_0 = const()[name = tensor<string, []>("out_17_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_17_transpose_y_0 = const()[name = tensor<string, []>("out_17_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = attn_as_float_17_cast_fp16, y = value_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")]; | |
| tensor<int32, [4]> var_709_perm_0 = const()[name = tensor<string, []>("op_709_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_710 = const()[name = tensor<string, []>("op_710"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_709_cast_fp16 = transpose(perm = var_709_perm_0, x = out_17_cast_fp16)[name = tensor<string, []>("transpose_8")]; | |
| tensor<fp16, [1, 77, 512]> input_185_cast_fp16 = reshape(shape = var_710, x = var_709_cast_fp16)[name = tensor<string, []>("input_185_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(102695360)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103219712)))]; | |
| tensor<fp16, [1, 77, 512]> linear_33_cast_fp16 = linear(bias = text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_53_cast_fp16 = add(x = linear_33_cast_fp16, y = x_49_cast_fp16)[name = tensor<string, []>("x_53_cast_fp16")]; | |
| tensor<int32, [1]> var_724_axes_0 = const()[name = tensor<string, []>("op_724_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103220800)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103221888)))]; | |
| tensor<fp16, [1, 77, 512]> var_724_cast_fp16 = layer_norm(axes = var_724_axes_0, beta = text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16, x = x_53_cast_fp16)[name = tensor<string, []>("op_724_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(103222976)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105320192)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_34_cast_fp16 = linear(bias = text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16, x = var_724_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")]; | |
| tensor<string, []> input_195_mode_0 = const()[name = tensor<string, []>("input_195_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_195_cast_fp16 = gelu(mode = input_195_mode_0, x = linear_34_cast_fp16)[name = tensor<string, []>("input_195_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105324352)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107421568)))]; | |
| tensor<fp16, [1, 77, 512]> linear_35_cast_fp16 = linear(bias = text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16, x = input_195_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_55_cast_fp16 = add(x = x_53_cast_fp16, y = linear_35_cast_fp16)[name = tensor<string, []>("x_55_cast_fp16")]; | |
| tensor<int32, [1]> var_751_axes_0 = const()[name = tensor<string, []>("op_751_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107422656)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107423744)))]; | |
| tensor<fp16, [1, 77, 512]> var_751_cast_fp16 = layer_norm(axes = var_751_axes_0, beta = text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16, x = x_55_cast_fp16)[name = tensor<string, []>("op_751_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(107424832)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(108997760)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_36_cast_fp16 = linear(bias = text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_751_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")]; | |
| tensor<int32, [5]> var_763 = const()[name = tensor<string, []>("op_763"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_37_cast_fp16 = reshape(shape = var_763, x = linear_36_cast_fp16)[name = tensor<string, []>("qkv_37_cast_fp16")]; | |
| tensor<int32, [5]> var_765_perm_0 = const()[name = tensor<string, []>("op_765_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_37_begin_0 = const()[name = tensor<string, []>("query_37_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_37_end_0 = const()[name = tensor<string, []>("query_37_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_37_end_mask_0 = const()[name = tensor<string, []>("query_37_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_37_squeeze_mask_0 = const()[name = tensor<string, []>("query_37_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_765_cast_fp16 = transpose(perm = var_765_perm_0, x = qkv_37_cast_fp16)[name = tensor<string, []>("transpose_7")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_37_cast_fp16 = slice_by_index(begin = query_37_begin_0, end = query_37_end_0, end_mask = query_37_end_mask_0, squeeze_mask = query_37_squeeze_mask_0, x = var_765_cast_fp16)[name = tensor<string, []>("query_37_cast_fp16")]; | |
| tensor<int32, [5]> key_37_begin_0 = const()[name = tensor<string, []>("key_37_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_37_end_0 = const()[name = tensor<string, []>("key_37_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_37_end_mask_0 = const()[name = tensor<string, []>("key_37_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_37_squeeze_mask_0 = const()[name = tensor<string, []>("key_37_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_37_cast_fp16 = slice_by_index(begin = key_37_begin_0, end = key_37_end_0, end_mask = key_37_end_mask_0, squeeze_mask = key_37_squeeze_mask_0, x = var_765_cast_fp16)[name = tensor<string, []>("key_37_cast_fp16")]; | |
| tensor<int32, [5]> value_19_begin_0 = const()[name = tensor<string, []>("value_19_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_19_end_0 = const()[name = tensor<string, []>("value_19_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_19_end_mask_0 = const()[name = tensor<string, []>("value_19_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_19_squeeze_mask_0 = const()[name = tensor<string, []>("value_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_19_cast_fp16 = slice_by_index(begin = value_19_begin_0, end = value_19_end_0, end_mask = value_19_end_mask_0, squeeze_mask = value_19_squeeze_mask_0, x = var_765_cast_fp16)[name = tensor<string, []>("value_19_cast_fp16")]; | |
| tensor<fp16, []> var_776_to_fp16 = const()[name = tensor<string, []>("op_776_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_39_cast_fp16 = mul(x = query_37_cast_fp16, y = var_776_to_fp16)[name = tensor<string, []>("query_39_cast_fp16")]; | |
| tensor<bool, []> attn_37_transpose_x_1 = const()[name = tensor<string, []>("attn_37_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_37_transpose_y_1 = const()[name = tensor<string, []>("attn_37_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_37_cast_fp16 = matmul(transpose_x = attn_37_transpose_x_1, transpose_y = attn_37_transpose_y_1, x = query_39_cast_fp16, y = key_37_cast_fp16)[name = tensor<string, []>("attn_37_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_19_cast_fp16 = softmax(axis = var_18, x = attn_37_cast_fp16)[name = tensor<string, []>("attn_as_float_19_cast_fp16")]; | |
| tensor<bool, []> out_19_transpose_x_0 = const()[name = tensor<string, []>("out_19_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_19_transpose_y_0 = const()[name = tensor<string, []>("out_19_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_19_cast_fp16 = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = attn_as_float_19_cast_fp16, y = value_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")]; | |
| tensor<int32, [4]> var_785_perm_0 = const()[name = tensor<string, []>("op_785_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_786 = const()[name = tensor<string, []>("op_786"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_785_cast_fp16 = transpose(perm = var_785_perm_0, x = out_19_cast_fp16)[name = tensor<string, []>("transpose_6")]; | |
| tensor<fp16, [1, 77, 512]> input_207_cast_fp16 = reshape(shape = var_786, x = var_785_cast_fp16)[name = tensor<string, []>("input_207_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109000896)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109525248)))]; | |
| tensor<fp16, [1, 77, 512]> linear_37_cast_fp16 = linear(bias = text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_59_cast_fp16 = add(x = linear_37_cast_fp16, y = x_55_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")]; | |
| tensor<int32, [1]> var_800_axes_0 = const()[name = tensor<string, []>("op_800_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109526336)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109527424)))]; | |
| tensor<fp16, [1, 77, 512]> var_800_cast_fp16 = layer_norm(axes = var_800_axes_0, beta = text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16, x = x_59_cast_fp16)[name = tensor<string, []>("op_800_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(109528512)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111625728)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_38_cast_fp16 = linear(bias = text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16, x = var_800_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")]; | |
| tensor<string, []> input_217_mode_0 = const()[name = tensor<string, []>("input_217_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_38_cast_fp16)[name = tensor<string, []>("input_217_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111629888)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113727104)))]; | |
| tensor<fp16, [1, 77, 512]> linear_39_cast_fp16 = linear(bias = text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16, x = input_217_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_39_cast_fp16)[name = tensor<string, []>("x_61_cast_fp16")]; | |
| tensor<int32, [1]> var_827_axes_0 = const()[name = tensor<string, []>("op_827_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113728192)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113729280)))]; | |
| tensor<fp16, [1, 77, 512]> var_827_cast_fp16 = layer_norm(axes = var_827_axes_0, beta = text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16, x = x_61_cast_fp16)[name = tensor<string, []>("op_827_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113730368)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115303296)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_40_cast_fp16 = linear(bias = text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_827_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")]; | |
| tensor<int32, [5]> var_839 = const()[name = tensor<string, []>("op_839"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_41_cast_fp16 = reshape(shape = var_839, x = linear_40_cast_fp16)[name = tensor<string, []>("qkv_41_cast_fp16")]; | |
| tensor<int32, [5]> var_841_perm_0 = const()[name = tensor<string, []>("op_841_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_41_begin_0 = const()[name = tensor<string, []>("query_41_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_41_end_0 = const()[name = tensor<string, []>("query_41_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_41_end_mask_0 = const()[name = tensor<string, []>("query_41_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_41_squeeze_mask_0 = const()[name = tensor<string, []>("query_41_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_841_cast_fp16 = transpose(perm = var_841_perm_0, x = qkv_41_cast_fp16)[name = tensor<string, []>("transpose_5")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_41_cast_fp16 = slice_by_index(begin = query_41_begin_0, end = query_41_end_0, end_mask = query_41_end_mask_0, squeeze_mask = query_41_squeeze_mask_0, x = var_841_cast_fp16)[name = tensor<string, []>("query_41_cast_fp16")]; | |
| tensor<int32, [5]> key_41_begin_0 = const()[name = tensor<string, []>("key_41_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_41_end_0 = const()[name = tensor<string, []>("key_41_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_41_end_mask_0 = const()[name = tensor<string, []>("key_41_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_41_squeeze_mask_0 = const()[name = tensor<string, []>("key_41_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_41_cast_fp16 = slice_by_index(begin = key_41_begin_0, end = key_41_end_0, end_mask = key_41_end_mask_0, squeeze_mask = key_41_squeeze_mask_0, x = var_841_cast_fp16)[name = tensor<string, []>("key_41_cast_fp16")]; | |
| tensor<int32, [5]> value_21_begin_0 = const()[name = tensor<string, []>("value_21_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_21_end_0 = const()[name = tensor<string, []>("value_21_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_21_end_mask_0 = const()[name = tensor<string, []>("value_21_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_21_squeeze_mask_0 = const()[name = tensor<string, []>("value_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_21_cast_fp16 = slice_by_index(begin = value_21_begin_0, end = value_21_end_0, end_mask = value_21_end_mask_0, squeeze_mask = value_21_squeeze_mask_0, x = var_841_cast_fp16)[name = tensor<string, []>("value_21_cast_fp16")]; | |
| tensor<fp16, []> var_852_to_fp16 = const()[name = tensor<string, []>("op_852_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_43_cast_fp16 = mul(x = query_41_cast_fp16, y = var_852_to_fp16)[name = tensor<string, []>("query_43_cast_fp16")]; | |
| tensor<bool, []> attn_41_transpose_x_1 = const()[name = tensor<string, []>("attn_41_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_41_transpose_y_1 = const()[name = tensor<string, []>("attn_41_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_41_cast_fp16 = matmul(transpose_x = attn_41_transpose_x_1, transpose_y = attn_41_transpose_y_1, x = query_43_cast_fp16, y = key_41_cast_fp16)[name = tensor<string, []>("attn_41_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_21_cast_fp16 = softmax(axis = var_18, x = attn_41_cast_fp16)[name = tensor<string, []>("attn_as_float_21_cast_fp16")]; | |
| tensor<bool, []> out_21_transpose_x_0 = const()[name = tensor<string, []>("out_21_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_21_transpose_y_0 = const()[name = tensor<string, []>("out_21_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = attn_as_float_21_cast_fp16, y = value_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")]; | |
| tensor<int32, [4]> var_861_perm_0 = const()[name = tensor<string, []>("op_861_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_862 = const()[name = tensor<string, []>("op_862"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_861_cast_fp16 = transpose(perm = var_861_perm_0, x = out_21_cast_fp16)[name = tensor<string, []>("transpose_4")]; | |
| tensor<fp16, [1, 77, 512]> input_229_cast_fp16 = reshape(shape = var_862, x = var_861_cast_fp16)[name = tensor<string, []>("input_229_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115306432)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115830784)))]; | |
| tensor<fp16, [1, 77, 512]> linear_41_cast_fp16 = linear(bias = text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_229_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_65_cast_fp16 = add(x = linear_41_cast_fp16, y = x_61_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")]; | |
| tensor<int32, [1]> var_876_axes_0 = const()[name = tensor<string, []>("op_876_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115831872)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115832960)))]; | |
| tensor<fp16, [1, 77, 512]> var_876_cast_fp16 = layer_norm(axes = var_876_axes_0, beta = text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16, x = x_65_cast_fp16)[name = tensor<string, []>("op_876_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(115834048)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117931264)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_42_cast_fp16 = linear(bias = text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16, x = var_876_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")]; | |
| tensor<string, []> input_239_mode_0 = const()[name = tensor<string, []>("input_239_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_42_cast_fp16)[name = tensor<string, []>("input_239_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117935424)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120032640)))]; | |
| tensor<fp16, [1, 77, 512]> linear_43_cast_fp16 = linear(bias = text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16, x = input_239_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_67_cast_fp16 = add(x = x_65_cast_fp16, y = linear_43_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")]; | |
| tensor<int32, [1]> var_903_axes_0 = const()[name = tensor<string, []>("op_903_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120033728)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120034816)))]; | |
| tensor<fp16, [1, 77, 512]> var_903_cast_fp16 = layer_norm(axes = var_903_axes_0, beta = text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16, x = x_67_cast_fp16)[name = tensor<string, []>("op_903_cast_fp16")]; | |
| tensor<fp16, [1536, 512]> text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(120035904)))]; | |
| tensor<fp16, [1536]> text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121608832)))]; | |
| tensor<fp16, [1, 77, 1536]> linear_44_cast_fp16 = linear(bias = text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_903_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")]; | |
| tensor<int32, [5]> var_915 = const()[name = tensor<string, []>("op_915"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_45_cast_fp16 = reshape(shape = var_915, x = linear_44_cast_fp16)[name = tensor<string, []>("qkv_45_cast_fp16")]; | |
| tensor<int32, [5]> var_917_perm_0 = const()[name = tensor<string, []>("op_917_perm_0"), val = tensor<int32, [5]>([0, 3, 2, 1, 4])]; | |
| tensor<int32, [5]> query_45_begin_0 = const()[name = tensor<string, []>("query_45_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; | |
| tensor<int32, [5]> query_45_end_0 = const()[name = tensor<string, []>("query_45_end_0"), val = tensor<int32, [5]>([1, 8, 1, 77, 64])]; | |
| tensor<bool, [5]> query_45_end_mask_0 = const()[name = tensor<string, []>("query_45_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> query_45_squeeze_mask_0 = const()[name = tensor<string, []>("query_45_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 3, 77, 64]> var_917_cast_fp16 = transpose(perm = var_917_perm_0, x = qkv_45_cast_fp16)[name = tensor<string, []>("transpose_3")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_45_cast_fp16 = slice_by_index(begin = query_45_begin_0, end = query_45_end_0, end_mask = query_45_end_mask_0, squeeze_mask = query_45_squeeze_mask_0, x = var_917_cast_fp16)[name = tensor<string, []>("query_45_cast_fp16")]; | |
| tensor<int32, [5]> key_45_begin_0 = const()[name = tensor<string, []>("key_45_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])]; | |
| tensor<int32, [5]> key_45_end_0 = const()[name = tensor<string, []>("key_45_end_0"), val = tensor<int32, [5]>([1, 8, 2, 77, 64])]; | |
| tensor<bool, [5]> key_45_end_mask_0 = const()[name = tensor<string, []>("key_45_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> key_45_squeeze_mask_0 = const()[name = tensor<string, []>("key_45_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> key_45_cast_fp16 = slice_by_index(begin = key_45_begin_0, end = key_45_end_0, end_mask = key_45_end_mask_0, squeeze_mask = key_45_squeeze_mask_0, x = var_917_cast_fp16)[name = tensor<string, []>("key_45_cast_fp16")]; | |
| tensor<int32, [5]> value_begin_0 = const()[name = tensor<string, []>("value_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])]; | |
| tensor<int32, [5]> value_end_0 = const()[name = tensor<string, []>("value_end_0"), val = tensor<int32, [5]>([1, 8, 3, 77, 64])]; | |
| tensor<bool, [5]> value_end_mask_0 = const()[name = tensor<string, []>("value_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])]; | |
| tensor<bool, [5]> value_squeeze_mask_0 = const()[name = tensor<string, []>("value_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])]; | |
| tensor<fp16, [1, 8, 77, 64]> value_cast_fp16 = slice_by_index(begin = value_begin_0, end = value_end_0, end_mask = value_end_mask_0, squeeze_mask = value_squeeze_mask_0, x = var_917_cast_fp16)[name = tensor<string, []>("value_cast_fp16")]; | |
| tensor<fp16, []> var_928_to_fp16 = const()[name = tensor<string, []>("op_928_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_cast_fp16 = mul(x = query_45_cast_fp16, y = var_928_to_fp16)[name = tensor<string, []>("query_cast_fp16")]; | |
| tensor<bool, []> attn_45_transpose_x_1 = const()[name = tensor<string, []>("attn_45_transpose_x_1"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_45_transpose_y_1 = const()[name = tensor<string, []>("attn_45_transpose_y_1"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_45_cast_fp16 = matmul(transpose_x = attn_45_transpose_x_1, transpose_y = attn_45_transpose_y_1, x = query_cast_fp16, y = key_45_cast_fp16)[name = tensor<string, []>("attn_45_cast_fp16")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_cast_fp16 = softmax(axis = var_18, x = attn_45_cast_fp16)[name = tensor<string, []>("attn_as_float_cast_fp16")]; | |
| tensor<bool, []> out_transpose_x_0 = const()[name = tensor<string, []>("out_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> out_transpose_y_0 = const()[name = tensor<string, []>("out_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 77, 64]> out_cast_fp16 = matmul(transpose_x = out_transpose_x_0, transpose_y = out_transpose_y_0, x = attn_as_float_cast_fp16, y = value_cast_fp16)[name = tensor<string, []>("out_cast_fp16")]; | |
| tensor<int32, [4]> var_937_perm_0 = const()[name = tensor<string, []>("op_937_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_938 = const()[name = tensor<string, []>("op_938"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> var_937_cast_fp16 = transpose(perm = var_937_perm_0, x = out_cast_fp16)[name = tensor<string, []>("transpose_2")]; | |
| tensor<fp16, [1, 77, 512]> input_251_cast_fp16 = reshape(shape = var_938, x = var_937_cast_fp16)[name = tensor<string, []>("input_251_cast_fp16")]; | |
| tensor<fp16, [512, 512]> text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121611968)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122136320)))]; | |
| tensor<fp16, [1, 77, 512]> linear_45_cast_fp16 = linear(bias = text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16, weight = text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_251_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_71_cast_fp16 = add(x = linear_45_cast_fp16, y = x_67_cast_fp16)[name = tensor<string, []>("x_71_cast_fp16")]; | |
| tensor<int32, [1]> var_952_axes_0 = const()[name = tensor<string, []>("op_952_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122137408)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122138496)))]; | |
| tensor<fp16, [1, 77, 512]> var_952_cast_fp16 = layer_norm(axes = var_952_axes_0, beta = text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16, x = x_71_cast_fp16)[name = tensor<string, []>("op_952_cast_fp16")]; | |
| tensor<fp16, [2048, 512]> text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16"), val = tensor<fp16, [2048, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(122139584)))]; | |
| tensor<fp16, [2048]> text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124236800)))]; | |
| tensor<fp16, [1, 77, 2048]> linear_46_cast_fp16 = linear(bias = text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16, weight = text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16, x = var_952_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")]; | |
| tensor<string, []> input_261_mode_0 = const()[name = tensor<string, []>("input_261_mode_0"), val = tensor<string, []>("EXACT")]; | |
| tensor<fp16, [1, 77, 2048]> input_261_cast_fp16 = gelu(mode = input_261_mode_0, x = linear_46_cast_fp16)[name = tensor<string, []>("input_261_cast_fp16")]; | |
| tensor<fp16, [512, 2048]> text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16"), val = tensor<fp16, [512, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(124240960)))]; | |
| tensor<fp16, [512]> text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126338176)))]; | |
| tensor<fp16, [1, 77, 512]> linear_47_cast_fp16 = linear(bias = text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16, weight = text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16, x = input_261_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")]; | |
| tensor<fp16, [1, 77, 512]> x_cast_fp16 = add(x = x_71_cast_fp16, y = linear_47_cast_fp16)[name = tensor<string, []>("x_cast_fp16")]; | |
| tensor<int32, [1]> var_974_axes_0 = const()[name = tensor<string, []>("op_974_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> text_encoder_final_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("text_encoder_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126339264)))]; | |
| tensor<fp16, [512]> text_encoder_final_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("text_encoder_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126340352)))]; | |
| tensor<fp16, [1, 77, 512]> var_974_cast_fp16 = layer_norm(axes = var_974_axes_0, beta = text_encoder_final_layer_norm_bias_to_fp16, epsilon = var_5_to_fp16, gamma = text_encoder_final_layer_norm_weight_to_fp16, x = x_cast_fp16)[name = tensor<string, []>("op_974_cast_fp16")]; | |
| tensor<int32, [1]> var_977 = const()[name = tensor<string, []>("op_977"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, []> var_978_axis_0 = const()[name = tensor<string, []>("op_978_axis_0"), val = tensor<int32, []>(-1)]; | |
| tensor<bool, []> var_978_keep_dims_0 = const()[name = tensor<string, []>("op_978_keep_dims_0"), val = tensor<bool, []>(false)]; | |
| tensor<string, []> var_978_output_dtype_0 = const()[name = tensor<string, []>("op_978_output_dtype_0"), val = tensor<string, []>("int32")]; | |
| tensor<int32, [1]> var_978 = reduce_argmax(axis = var_978_axis_0, keep_dims = var_978_keep_dims_0, output_dtype = var_978_output_dtype_0, x = input_text)[name = tensor<string, []>("op_978")]; | |
| tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(1)]; | |
| tensor<int32, [1, 2]> stack_0 = stack(axis = stack_0_axis_0, values = (var_977, var_978))[name = tensor<string, []>("stack_0")]; | |
| tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)]; | |
| tensor<bool, [1, 2]> greater_equal_0 = greater_equal(x = stack_0, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")]; | |
| tensor<int32, [2]> slice_by_size_0 = const()[name = tensor<string, []>("slice_by_size_0"), val = tensor<int32, [2]>([1, 77])]; | |
| tensor<int32, [1, 2]> add_0 = add(x = stack_0, y = slice_by_size_0)[name = tensor<string, []>("add_0")]; | |
| tensor<int32, [1, 2]> select_0 = select(a = stack_0, b = add_0, cond = greater_equal_0)[name = tensor<string, []>("select_0")]; | |
| tensor<int32, []> token_emb_transpose_batch_dims_0 = const()[name = tensor<string, []>("token_emb_transpose_batch_dims_0"), val = tensor<int32, []>(0)]; | |
| tensor<bool, []> token_emb_transpose_validate_indices_0 = const()[name = tensor<string, []>("token_emb_transpose_validate_indices_0"), val = tensor<bool, []>(false)]; | |
| tensor<string, []> select_0_to_uint16_dtype_0 = const()[name = tensor<string, []>("select_0_to_uint16_dtype_0"), val = tensor<string, []>("uint16")]; | |
| tensor<uint16, [1, 2]> select_0_to_uint16 = cast(dtype = select_0_to_uint16_dtype_0, x = select_0)[name = tensor<string, []>("cast_103")]; | |
| tensor<fp16, [1, 512]> token_emb_transpose_cast_fp16_cast_uint16 = gather_nd(batch_dims = token_emb_transpose_batch_dims_0, indices = select_0_to_uint16, validate_indices = token_emb_transpose_validate_indices_0, x = var_974_cast_fp16)[name = tensor<string, []>("token_emb_transpose_cast_fp16_cast_uint16")]; | |
| tensor<fp16, [512, 512]> transpose_1_to_fp16 = const()[name = tensor<string, []>("transpose_1_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126341440)))]; | |
| tensor<fp16, [512]> input_bias_0_to_fp16 = const()[name = tensor<string, []>("input_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126865792)))]; | |
| tensor<fp16, [1, 512]> input_cast_fp16 = linear(bias = input_bias_0_to_fp16, weight = transpose_1_to_fp16, x = token_emb_transpose_cast_fp16_cast_uint16)[name = tensor<string, []>("input_cast_fp16")]; | |
| tensor<int32, [1]> var_984 = const()[name = tensor<string, []>("op_984"), val = tensor<int32, [1]>([-1])]; | |
| tensor<bool, []> var_985 = const()[name = tensor<string, []>("op_985"), val = tensor<bool, []>(true)]; | |
| tensor<fp16, [1, 1]> var_987_cast_fp16 = reduce_l2_norm(axes = var_984, keep_dims = var_985, x = input_cast_fp16)[name = tensor<string, []>("op_987_cast_fp16")]; | |
| tensor<fp16, []> var_988_to_fp16 = const()[name = tensor<string, []>("op_988_to_fp16"), val = tensor<fp16, []>(0x1p-24)]; | |
| tensor<fp16, [1, 1]> var_989_cast_fp16 = maximum(x = var_987_cast_fp16, y = var_988_to_fp16)[name = tensor<string, []>("op_989_cast_fp16")]; | |
| tensor<int32, [2]> denom_reps_0 = const()[name = tensor<string, []>("denom_reps_0"), val = tensor<int32, [2]>([1, 512])]; | |
| tensor<fp16, [1, 512]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_989_cast_fp16)[name = tensor<string, []>("denom_cast_fp16")]; | |
| tensor<fp16, [1, 512]> output_embeddings = real_div(x = input_cast_fp16, y = denom_cast_fp16)[name = tensor<string, []>("op_991_cast_fp16")]; | |
| } -> (output_embeddings); | |
| } |