| program(1.0) | |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "5.33.5"}, {"coremlc-version", "1877.40.3"}, {"coremltools-component-torch", "1.13.1"}, {"coremltools-version", "7.0"}})] | |
| { | |
| func main<ios16>(tensor<int32, [1, 77]> input_tokens) { | |
| tensor<int32, []> var_19 = const()[name = tensor<string, []>("op_19"), val = tensor<int32, []>(-1)]; | |
| tensor<bool, []> var_20 = const()[name = tensor<string, []>("op_20"), val = tensor<bool, []>(false)]; | |
| 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<fp16, [49408, 512]> original_model_text_encoder_embedding_layer_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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 = gather(axis = token_emb_1_axis_0, batch_dims = token_emb_1_batch_dims_0, indices = input_tokens, x = original_model_text_encoder_embedding_layer_weight_to_fp16)[name = tensor<string, []>("token_emb_1_cast")]; | |
| tensor<fp16, [1, 77, 512]> var_56_to_fp16 = const()[name = tensor<string, []>("op_56_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 = add(x = token_emb_1_cast, y = var_56_to_fp16)[name = tensor<string, []>("input_1_cast")]; | |
| tensor<int32, [1]> var_68_axes_0 = const()[name = tensor<string, []>("op_68_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_6_to_fp16 = const()[name = tensor<string, []>("op_6_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)]; | |
| tensor<fp16, [1, 77, 512]> var_68_cast = layer_norm(axes = var_68_axes_0, beta = original_model_text_encoder_transformer_0_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_0_pre_norm_mha_0_weight_to_fp16, x = input_1_cast)[name = tensor<string, []>("op_68_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_68_cast)[name = tensor<string, []>("linear_0_cast")]; | |
| tensor<int32, [5]> var_80 = const()[name = tensor<string, []>("op_80"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_1_cast = reshape(shape = var_80, x = linear_0_cast)[name = tensor<string, []>("qkv_1_cast")]; | |
| tensor<int32, [5]> var_82_perm_0 = const()[name = tensor<string, []>("op_82_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]> transpose_36 = transpose(perm = var_82_perm_0, x = qkv_1_cast)[name = tensor<string, []>("transpose_36")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_1_cast = 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 = transpose_36)[name = tensor<string, []>("query_1_cast")]; | |
| 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 = 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 = transpose_36)[name = tensor<string, []>("key_1_cast")]; | |
| 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 = 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 = transpose_36)[name = tensor<string, []>("value_1_cast")]; | |
| tensor<fp16, []> var_93_to_fp16 = const()[name = tensor<string, []>("op_93_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_3_cast = mul(x = query_1_cast, y = var_93_to_fp16)[name = tensor<string, []>("query_3_cast")]; | |
| tensor<int32, [4]> key_3_perm_0 = const()[name = tensor<string, []>("key_3_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_1_transpose_x_0 = const()[name = tensor<string, []>("attn_1_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_1_transpose_y_0 = const()[name = tensor<string, []>("attn_1_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_35 = transpose(perm = key_3_perm_0, x = key_1_cast)[name = tensor<string, []>("transpose_35")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_1_cast = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = query_3_cast, y = transpose_35)[name = tensor<string, []>("attn_1_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_1_cast = softmax(axis = var_19, x = attn_1_cast)[name = tensor<string, []>("attn_as_float_1_cast")]; | |
| 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 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_as_float_1_cast, y = value_1_cast)[name = tensor<string, []>("out_1_cast")]; | |
| tensor<int32, [4]> var_102_perm_0 = const()[name = tensor<string, []>("op_102_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_103 = const()[name = tensor<string, []>("op_103"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_34 = transpose(perm = var_102_perm_0, x = out_1_cast)[name = tensor<string, []>("transpose_34")]; | |
| tensor<fp16, [1, 77, 512]> input_9_cast = reshape(shape = var_103, x = transpose_34)[name = tensor<string, []>("input_9_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_9_cast)[name = tensor<string, []>("linear_1_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_5_cast = add(x = linear_1_cast, y = input_1_cast)[name = tensor<string, []>("x_5_cast")]; | |
| tensor<int32, [1]> var_117_axes_0 = const()[name = tensor<string, []>("op_117_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_117_cast = layer_norm(axes = var_117_axes_0, beta = original_model_text_encoder_transformer_0_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_0_pre_norm_ffn_0_weight_to_fp16, x = x_5_cast)[name = tensor<string, []>("op_117_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_0_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_ffn_1_weight_to_fp16, x = var_117_cast)[name = tensor<string, []>("linear_2_cast")]; | |
| 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 = gelu(mode = input_19_mode_0, x = linear_2_cast)[name = tensor<string, []>("input_19_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_0_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_0_pre_norm_ffn_4_weight_to_fp16, x = input_19_cast)[name = tensor<string, []>("linear_3_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_7_cast = add(x = x_5_cast, y = linear_3_cast)[name = tensor<string, []>("x_7_cast")]; | |
| tensor<int32, [1]> var_144_axes_0 = const()[name = tensor<string, []>("op_144_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_144_cast = layer_norm(axes = var_144_axes_0, beta = original_model_text_encoder_transformer_1_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_1_pre_norm_mha_0_weight_to_fp16, x = x_7_cast)[name = tensor<string, []>("op_144_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_144_cast)[name = tensor<string, []>("linear_4_cast")]; | |
| tensor<int32, [5]> var_156 = const()[name = tensor<string, []>("op_156"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_5_cast = reshape(shape = var_156, x = linear_4_cast)[name = tensor<string, []>("qkv_5_cast")]; | |
| tensor<int32, [5]> var_158_perm_0 = const()[name = tensor<string, []>("op_158_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]> transpose_33 = transpose(perm = var_158_perm_0, x = qkv_5_cast)[name = tensor<string, []>("transpose_33")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_5_cast = 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 = transpose_33)[name = tensor<string, []>("query_5_cast")]; | |
| 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 = 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 = transpose_33)[name = tensor<string, []>("key_5_cast")]; | |
| 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 = 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 = transpose_33)[name = tensor<string, []>("value_3_cast")]; | |
| tensor<fp16, []> var_169_to_fp16 = const()[name = tensor<string, []>("op_169_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_7_cast = mul(x = query_5_cast, y = var_169_to_fp16)[name = tensor<string, []>("query_7_cast")]; | |
| tensor<int32, [4]> key_7_perm_0 = const()[name = tensor<string, []>("key_7_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_5_transpose_x_0 = const()[name = tensor<string, []>("attn_5_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_5_transpose_y_0 = const()[name = tensor<string, []>("attn_5_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_32 = transpose(perm = key_7_perm_0, x = key_5_cast)[name = tensor<string, []>("transpose_32")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_5_cast = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = query_7_cast, y = transpose_32)[name = tensor<string, []>("attn_5_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_3_cast = softmax(axis = var_19, x = attn_5_cast)[name = tensor<string, []>("attn_as_float_3_cast")]; | |
| 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 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_as_float_3_cast, y = value_3_cast)[name = tensor<string, []>("out_3_cast")]; | |
| tensor<int32, [4]> var_178_perm_0 = const()[name = tensor<string, []>("op_178_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_179 = const()[name = tensor<string, []>("op_179"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_31 = transpose(perm = var_178_perm_0, x = out_3_cast)[name = tensor<string, []>("transpose_31")]; | |
| tensor<fp16, [1, 77, 512]> input_31_cast = reshape(shape = var_179, x = transpose_31)[name = tensor<string, []>("input_31_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_31_cast)[name = tensor<string, []>("linear_5_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_11_cast = add(x = linear_5_cast, y = x_7_cast)[name = tensor<string, []>("x_11_cast")]; | |
| tensor<int32, [1]> var_193_axes_0 = const()[name = tensor<string, []>("op_193_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_193_cast = layer_norm(axes = var_193_axes_0, beta = original_model_text_encoder_transformer_1_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_1_pre_norm_ffn_0_weight_to_fp16, x = x_11_cast)[name = tensor<string, []>("op_193_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_1_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_ffn_1_weight_to_fp16, x = var_193_cast)[name = tensor<string, []>("linear_6_cast")]; | |
| 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 = gelu(mode = input_41_mode_0, x = linear_6_cast)[name = tensor<string, []>("input_41_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_1_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_1_pre_norm_ffn_4_weight_to_fp16, x = input_41_cast)[name = tensor<string, []>("linear_7_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_13_cast = add(x = x_11_cast, y = linear_7_cast)[name = tensor<string, []>("x_13_cast")]; | |
| tensor<int32, [1]> var_220_axes_0 = const()[name = tensor<string, []>("op_220_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_220_cast = layer_norm(axes = var_220_axes_0, beta = original_model_text_encoder_transformer_2_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_2_pre_norm_mha_0_weight_to_fp16, x = x_13_cast)[name = tensor<string, []>("op_220_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_220_cast)[name = tensor<string, []>("linear_8_cast")]; | |
| tensor<int32, [5]> var_232 = const()[name = tensor<string, []>("op_232"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_9_cast = reshape(shape = var_232, x = linear_8_cast)[name = tensor<string, []>("qkv_9_cast")]; | |
| tensor<int32, [5]> var_234_perm_0 = const()[name = tensor<string, []>("op_234_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]> transpose_30 = transpose(perm = var_234_perm_0, x = qkv_9_cast)[name = tensor<string, []>("transpose_30")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_9_cast = 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 = transpose_30)[name = tensor<string, []>("query_9_cast")]; | |
| 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 = 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 = transpose_30)[name = tensor<string, []>("key_9_cast")]; | |
| 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 = 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 = transpose_30)[name = tensor<string, []>("value_5_cast")]; | |
| tensor<fp16, []> var_245_to_fp16 = const()[name = tensor<string, []>("op_245_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_11_cast = mul(x = query_9_cast, y = var_245_to_fp16)[name = tensor<string, []>("query_11_cast")]; | |
| tensor<int32, [4]> key_11_perm_0 = const()[name = tensor<string, []>("key_11_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_9_transpose_x_0 = const()[name = tensor<string, []>("attn_9_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_9_transpose_y_0 = const()[name = tensor<string, []>("attn_9_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_29 = transpose(perm = key_11_perm_0, x = key_9_cast)[name = tensor<string, []>("transpose_29")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_9_cast = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = query_11_cast, y = transpose_29)[name = tensor<string, []>("attn_9_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_5_cast = softmax(axis = var_19, x = attn_9_cast)[name = tensor<string, []>("attn_as_float_5_cast")]; | |
| 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 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_as_float_5_cast, y = value_5_cast)[name = tensor<string, []>("out_5_cast")]; | |
| tensor<int32, [4]> var_254_perm_0 = const()[name = tensor<string, []>("op_254_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_255 = const()[name = tensor<string, []>("op_255"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_28 = transpose(perm = var_254_perm_0, x = out_5_cast)[name = tensor<string, []>("transpose_28")]; | |
| tensor<fp16, [1, 77, 512]> input_53_cast = reshape(shape = var_255, x = transpose_28)[name = tensor<string, []>("input_53_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_53_cast)[name = tensor<string, []>("linear_9_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_17_cast = add(x = linear_9_cast, y = x_13_cast)[name = tensor<string, []>("x_17_cast")]; | |
| tensor<int32, [1]> var_269_axes_0 = const()[name = tensor<string, []>("op_269_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_269_cast = layer_norm(axes = var_269_axes_0, beta = original_model_text_encoder_transformer_2_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_2_pre_norm_ffn_0_weight_to_fp16, x = x_17_cast)[name = tensor<string, []>("op_269_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_2_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_ffn_1_weight_to_fp16, x = var_269_cast)[name = tensor<string, []>("linear_10_cast")]; | |
| 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 = gelu(mode = input_63_mode_0, x = linear_10_cast)[name = tensor<string, []>("input_63_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_2_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_2_pre_norm_ffn_4_weight_to_fp16, x = input_63_cast)[name = tensor<string, []>("linear_11_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_19_cast = add(x = x_17_cast, y = linear_11_cast)[name = tensor<string, []>("x_19_cast")]; | |
| tensor<int32, [1]> var_296_axes_0 = const()[name = tensor<string, []>("op_296_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_296_cast = layer_norm(axes = var_296_axes_0, beta = original_model_text_encoder_transformer_3_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_3_pre_norm_mha_0_weight_to_fp16, x = x_19_cast)[name = tensor<string, []>("op_296_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_296_cast)[name = tensor<string, []>("linear_12_cast")]; | |
| tensor<int32, [5]> var_308 = const()[name = tensor<string, []>("op_308"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_13_cast = reshape(shape = var_308, x = linear_12_cast)[name = tensor<string, []>("qkv_13_cast")]; | |
| tensor<int32, [5]> var_310_perm_0 = const()[name = tensor<string, []>("op_310_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]> transpose_27 = transpose(perm = var_310_perm_0, x = qkv_13_cast)[name = tensor<string, []>("transpose_27")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_13_cast = 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 = transpose_27)[name = tensor<string, []>("query_13_cast")]; | |
| 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 = 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 = transpose_27)[name = tensor<string, []>("key_13_cast")]; | |
| 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 = 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 = transpose_27)[name = tensor<string, []>("value_7_cast")]; | |
| tensor<fp16, []> var_321_to_fp16 = const()[name = tensor<string, []>("op_321_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_15_cast = mul(x = query_13_cast, y = var_321_to_fp16)[name = tensor<string, []>("query_15_cast")]; | |
| tensor<int32, [4]> key_15_perm_0 = const()[name = tensor<string, []>("key_15_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_13_transpose_x_0 = const()[name = tensor<string, []>("attn_13_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_13_transpose_y_0 = const()[name = tensor<string, []>("attn_13_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_26 = transpose(perm = key_15_perm_0, x = key_13_cast)[name = tensor<string, []>("transpose_26")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_13_cast = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = query_15_cast, y = transpose_26)[name = tensor<string, []>("attn_13_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_7_cast = softmax(axis = var_19, x = attn_13_cast)[name = tensor<string, []>("attn_as_float_7_cast")]; | |
| 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 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_as_float_7_cast, y = value_7_cast)[name = tensor<string, []>("out_7_cast")]; | |
| tensor<int32, [4]> var_330_perm_0 = const()[name = tensor<string, []>("op_330_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_331 = const()[name = tensor<string, []>("op_331"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_25 = transpose(perm = var_330_perm_0, x = out_7_cast)[name = tensor<string, []>("transpose_25")]; | |
| tensor<fp16, [1, 77, 512]> input_75_cast = reshape(shape = var_331, x = transpose_25)[name = tensor<string, []>("input_75_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_75_cast)[name = tensor<string, []>("linear_13_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_23_cast = add(x = linear_13_cast, y = x_19_cast)[name = tensor<string, []>("x_23_cast")]; | |
| tensor<int32, [1]> var_345_axes_0 = const()[name = tensor<string, []>("op_345_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_345_cast = layer_norm(axes = var_345_axes_0, beta = original_model_text_encoder_transformer_3_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_3_pre_norm_ffn_0_weight_to_fp16, x = x_23_cast)[name = tensor<string, []>("op_345_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_3_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_ffn_1_weight_to_fp16, x = var_345_cast)[name = tensor<string, []>("linear_14_cast")]; | |
| 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 = gelu(mode = input_85_mode_0, x = linear_14_cast)[name = tensor<string, []>("input_85_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_3_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_3_pre_norm_ffn_4_weight_to_fp16, x = input_85_cast)[name = tensor<string, []>("linear_15_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_25_cast = add(x = x_23_cast, y = linear_15_cast)[name = tensor<string, []>("x_25_cast")]; | |
| tensor<int32, [1]> var_372_axes_0 = const()[name = tensor<string, []>("op_372_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_372_cast = layer_norm(axes = var_372_axes_0, beta = original_model_text_encoder_transformer_4_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_4_pre_norm_mha_0_weight_to_fp16, x = x_25_cast)[name = tensor<string, []>("op_372_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_372_cast)[name = tensor<string, []>("linear_16_cast")]; | |
| tensor<int32, [5]> var_384 = const()[name = tensor<string, []>("op_384"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_17_cast = reshape(shape = var_384, x = linear_16_cast)[name = tensor<string, []>("qkv_17_cast")]; | |
| tensor<int32, [5]> var_386_perm_0 = const()[name = tensor<string, []>("op_386_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]> transpose_24 = transpose(perm = var_386_perm_0, x = qkv_17_cast)[name = tensor<string, []>("transpose_24")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_17_cast = 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 = transpose_24)[name = tensor<string, []>("query_17_cast")]; | |
| 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 = 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 = transpose_24)[name = tensor<string, []>("key_17_cast")]; | |
| 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 = 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 = transpose_24)[name = tensor<string, []>("value_9_cast")]; | |
| tensor<fp16, []> var_397_to_fp16 = const()[name = tensor<string, []>("op_397_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_19_cast = mul(x = query_17_cast, y = var_397_to_fp16)[name = tensor<string, []>("query_19_cast")]; | |
| tensor<int32, [4]> key_19_perm_0 = const()[name = tensor<string, []>("key_19_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_17_transpose_x_0 = const()[name = tensor<string, []>("attn_17_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_17_transpose_y_0 = const()[name = tensor<string, []>("attn_17_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_23 = transpose(perm = key_19_perm_0, x = key_17_cast)[name = tensor<string, []>("transpose_23")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_17_cast = matmul(transpose_x = attn_17_transpose_x_0, transpose_y = attn_17_transpose_y_0, x = query_19_cast, y = transpose_23)[name = tensor<string, []>("attn_17_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_9_cast = softmax(axis = var_19, x = attn_17_cast)[name = tensor<string, []>("attn_as_float_9_cast")]; | |
| 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 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_as_float_9_cast, y = value_9_cast)[name = tensor<string, []>("out_9_cast")]; | |
| tensor<int32, [4]> var_406_perm_0 = const()[name = tensor<string, []>("op_406_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_407 = const()[name = tensor<string, []>("op_407"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_22 = transpose(perm = var_406_perm_0, x = out_9_cast)[name = tensor<string, []>("transpose_22")]; | |
| tensor<fp16, [1, 77, 512]> input_97_cast = reshape(shape = var_407, x = transpose_22)[name = tensor<string, []>("input_97_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_97_cast)[name = tensor<string, []>("linear_17_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_29_cast = add(x = linear_17_cast, y = x_25_cast)[name = tensor<string, []>("x_29_cast")]; | |
| tensor<int32, [1]> var_421_axes_0 = const()[name = tensor<string, []>("op_421_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_421_cast = layer_norm(axes = var_421_axes_0, beta = original_model_text_encoder_transformer_4_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_4_pre_norm_ffn_0_weight_to_fp16, x = x_29_cast)[name = tensor<string, []>("op_421_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_4_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_ffn_1_weight_to_fp16, x = var_421_cast)[name = tensor<string, []>("linear_18_cast")]; | |
| 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 = gelu(mode = input_107_mode_0, x = linear_18_cast)[name = tensor<string, []>("input_107_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_4_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_4_pre_norm_ffn_4_weight_to_fp16, x = input_107_cast)[name = tensor<string, []>("linear_19_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_31_cast = add(x = x_29_cast, y = linear_19_cast)[name = tensor<string, []>("x_31_cast")]; | |
| tensor<int32, [1]> var_448_axes_0 = const()[name = tensor<string, []>("op_448_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_448_cast = layer_norm(axes = var_448_axes_0, beta = original_model_text_encoder_transformer_5_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_5_pre_norm_mha_0_weight_to_fp16, x = x_31_cast)[name = tensor<string, []>("op_448_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_448_cast)[name = tensor<string, []>("linear_20_cast")]; | |
| tensor<int32, [5]> var_460 = const()[name = tensor<string, []>("op_460"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_21_cast = reshape(shape = var_460, x = linear_20_cast)[name = tensor<string, []>("qkv_21_cast")]; | |
| tensor<int32, [5]> var_462_perm_0 = const()[name = tensor<string, []>("op_462_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]> transpose_21 = transpose(perm = var_462_perm_0, x = qkv_21_cast)[name = tensor<string, []>("transpose_21")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_21_cast = 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 = transpose_21)[name = tensor<string, []>("query_21_cast")]; | |
| 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 = 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 = transpose_21)[name = tensor<string, []>("key_21_cast")]; | |
| 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 = 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 = transpose_21)[name = tensor<string, []>("value_11_cast")]; | |
| tensor<fp16, []> var_473_to_fp16 = const()[name = tensor<string, []>("op_473_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_23_cast = mul(x = query_21_cast, y = var_473_to_fp16)[name = tensor<string, []>("query_23_cast")]; | |
| tensor<int32, [4]> key_23_perm_0 = const()[name = tensor<string, []>("key_23_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_21_transpose_x_0 = const()[name = tensor<string, []>("attn_21_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_21_transpose_y_0 = const()[name = tensor<string, []>("attn_21_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_20 = transpose(perm = key_23_perm_0, x = key_21_cast)[name = tensor<string, []>("transpose_20")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_21_cast = matmul(transpose_x = attn_21_transpose_x_0, transpose_y = attn_21_transpose_y_0, x = query_23_cast, y = transpose_20)[name = tensor<string, []>("attn_21_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_11_cast = softmax(axis = var_19, x = attn_21_cast)[name = tensor<string, []>("attn_as_float_11_cast")]; | |
| 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 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_as_float_11_cast, y = value_11_cast)[name = tensor<string, []>("out_11_cast")]; | |
| tensor<int32, [4]> var_482_perm_0 = const()[name = tensor<string, []>("op_482_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_483 = const()[name = tensor<string, []>("op_483"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_19 = transpose(perm = var_482_perm_0, x = out_11_cast)[name = tensor<string, []>("transpose_19")]; | |
| tensor<fp16, [1, 77, 512]> input_119_cast = reshape(shape = var_483, x = transpose_19)[name = tensor<string, []>("input_119_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_119_cast)[name = tensor<string, []>("linear_21_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_35_cast = add(x = linear_21_cast, y = x_31_cast)[name = tensor<string, []>("x_35_cast")]; | |
| tensor<int32, [1]> var_497_axes_0 = const()[name = tensor<string, []>("op_497_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_497_cast = layer_norm(axes = var_497_axes_0, beta = original_model_text_encoder_transformer_5_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_5_pre_norm_ffn_0_weight_to_fp16, x = x_35_cast)[name = tensor<string, []>("op_497_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_5_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_ffn_1_weight_to_fp16, x = var_497_cast)[name = tensor<string, []>("linear_22_cast")]; | |
| 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 = gelu(mode = input_129_mode_0, x = linear_22_cast)[name = tensor<string, []>("input_129_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_5_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_5_pre_norm_ffn_4_weight_to_fp16, x = input_129_cast)[name = tensor<string, []>("linear_23_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_37_cast = add(x = x_35_cast, y = linear_23_cast)[name = tensor<string, []>("x_37_cast")]; | |
| tensor<int32, [1]> var_524_axes_0 = const()[name = tensor<string, []>("op_524_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_524_cast = layer_norm(axes = var_524_axes_0, beta = original_model_text_encoder_transformer_6_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_6_pre_norm_mha_0_weight_to_fp16, x = x_37_cast)[name = tensor<string, []>("op_524_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_524_cast)[name = tensor<string, []>("linear_24_cast")]; | |
| tensor<int32, [5]> var_536 = const()[name = tensor<string, []>("op_536"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_25_cast = reshape(shape = var_536, x = linear_24_cast)[name = tensor<string, []>("qkv_25_cast")]; | |
| tensor<int32, [5]> var_538_perm_0 = const()[name = tensor<string, []>("op_538_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]> transpose_18 = transpose(perm = var_538_perm_0, x = qkv_25_cast)[name = tensor<string, []>("transpose_18")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_25_cast = 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 = transpose_18)[name = tensor<string, []>("query_25_cast")]; | |
| 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 = 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 = transpose_18)[name = tensor<string, []>("key_25_cast")]; | |
| 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 = 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 = transpose_18)[name = tensor<string, []>("value_13_cast")]; | |
| tensor<fp16, []> var_549_to_fp16 = const()[name = tensor<string, []>("op_549_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_27_cast = mul(x = query_25_cast, y = var_549_to_fp16)[name = tensor<string, []>("query_27_cast")]; | |
| tensor<int32, [4]> key_27_perm_0 = const()[name = tensor<string, []>("key_27_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_25_transpose_x_0 = const()[name = tensor<string, []>("attn_25_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_25_transpose_y_0 = const()[name = tensor<string, []>("attn_25_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_17 = transpose(perm = key_27_perm_0, x = key_25_cast)[name = tensor<string, []>("transpose_17")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_25_cast = matmul(transpose_x = attn_25_transpose_x_0, transpose_y = attn_25_transpose_y_0, x = query_27_cast, y = transpose_17)[name = tensor<string, []>("attn_25_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_13_cast = softmax(axis = var_19, x = attn_25_cast)[name = tensor<string, []>("attn_as_float_13_cast")]; | |
| 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 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = attn_as_float_13_cast, y = value_13_cast)[name = tensor<string, []>("out_13_cast")]; | |
| tensor<int32, [4]> var_558_perm_0 = const()[name = tensor<string, []>("op_558_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_16 = transpose(perm = var_558_perm_0, x = out_13_cast)[name = tensor<string, []>("transpose_16")]; | |
| tensor<fp16, [1, 77, 512]> input_141_cast = reshape(shape = var_559, x = transpose_16)[name = tensor<string, []>("input_141_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_141_cast)[name = tensor<string, []>("linear_25_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_41_cast = add(x = linear_25_cast, y = x_37_cast)[name = tensor<string, []>("x_41_cast")]; | |
| tensor<int32, [1]> var_573_axes_0 = const()[name = tensor<string, []>("op_573_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_573_cast = layer_norm(axes = var_573_axes_0, beta = original_model_text_encoder_transformer_6_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_6_pre_norm_ffn_0_weight_to_fp16, x = x_41_cast)[name = tensor<string, []>("op_573_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_6_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_ffn_1_weight_to_fp16, x = var_573_cast)[name = tensor<string, []>("linear_26_cast")]; | |
| 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 = gelu(mode = input_151_mode_0, x = linear_26_cast)[name = tensor<string, []>("input_151_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_6_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_6_pre_norm_ffn_4_weight_to_fp16, x = input_151_cast)[name = tensor<string, []>("linear_27_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_43_cast = add(x = x_41_cast, y = linear_27_cast)[name = tensor<string, []>("x_43_cast")]; | |
| tensor<int32, [1]> var_600_axes_0 = const()[name = tensor<string, []>("op_600_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_600_cast = layer_norm(axes = var_600_axes_0, beta = original_model_text_encoder_transformer_7_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_7_pre_norm_mha_0_weight_to_fp16, x = x_43_cast)[name = tensor<string, []>("op_600_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_600_cast)[name = tensor<string, []>("linear_28_cast")]; | |
| tensor<int32, [5]> var_612 = const()[name = tensor<string, []>("op_612"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_29_cast = reshape(shape = var_612, x = linear_28_cast)[name = tensor<string, []>("qkv_29_cast")]; | |
| tensor<int32, [5]> var_614_perm_0 = const()[name = tensor<string, []>("op_614_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]> transpose_15 = transpose(perm = var_614_perm_0, x = qkv_29_cast)[name = tensor<string, []>("transpose_15")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_29_cast = 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 = transpose_15)[name = tensor<string, []>("query_29_cast")]; | |
| 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 = 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 = transpose_15)[name = tensor<string, []>("key_29_cast")]; | |
| 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 = 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 = transpose_15)[name = tensor<string, []>("value_15_cast")]; | |
| tensor<fp16, []> var_625_to_fp16 = const()[name = tensor<string, []>("op_625_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_31_cast = mul(x = query_29_cast, y = var_625_to_fp16)[name = tensor<string, []>("query_31_cast")]; | |
| tensor<int32, [4]> key_31_perm_0 = const()[name = tensor<string, []>("key_31_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_29_transpose_x_0 = const()[name = tensor<string, []>("attn_29_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_29_transpose_y_0 = const()[name = tensor<string, []>("attn_29_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_14 = transpose(perm = key_31_perm_0, x = key_29_cast)[name = tensor<string, []>("transpose_14")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_29_cast = matmul(transpose_x = attn_29_transpose_x_0, transpose_y = attn_29_transpose_y_0, x = query_31_cast, y = transpose_14)[name = tensor<string, []>("attn_29_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_15_cast = softmax(axis = var_19, x = attn_29_cast)[name = tensor<string, []>("attn_as_float_15_cast")]; | |
| 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 = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = attn_as_float_15_cast, y = value_15_cast)[name = tensor<string, []>("out_15_cast")]; | |
| tensor<int32, [4]> var_634_perm_0 = const()[name = tensor<string, []>("op_634_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_635 = const()[name = tensor<string, []>("op_635"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_13 = transpose(perm = var_634_perm_0, x = out_15_cast)[name = tensor<string, []>("transpose_13")]; | |
| tensor<fp16, [1, 77, 512]> input_163_cast = reshape(shape = var_635, x = transpose_13)[name = tensor<string, []>("input_163_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_163_cast)[name = tensor<string, []>("linear_29_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_47_cast = add(x = linear_29_cast, y = x_43_cast)[name = tensor<string, []>("x_47_cast")]; | |
| tensor<int32, [1]> var_649_axes_0 = const()[name = tensor<string, []>("op_649_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_649_cast = layer_norm(axes = var_649_axes_0, beta = original_model_text_encoder_transformer_7_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_7_pre_norm_ffn_0_weight_to_fp16, x = x_47_cast)[name = tensor<string, []>("op_649_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_7_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_ffn_1_weight_to_fp16, x = var_649_cast)[name = tensor<string, []>("linear_30_cast")]; | |
| 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 = gelu(mode = input_173_mode_0, x = linear_30_cast)[name = tensor<string, []>("input_173_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_7_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_7_pre_norm_ffn_4_weight_to_fp16, x = input_173_cast)[name = tensor<string, []>("linear_31_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_49_cast = add(x = x_47_cast, y = linear_31_cast)[name = tensor<string, []>("x_49_cast")]; | |
| tensor<int32, [1]> var_676_axes_0 = const()[name = tensor<string, []>("op_676_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_676_cast = layer_norm(axes = var_676_axes_0, beta = original_model_text_encoder_transformer_8_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_8_pre_norm_mha_0_weight_to_fp16, x = x_49_cast)[name = tensor<string, []>("op_676_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_676_cast)[name = tensor<string, []>("linear_32_cast")]; | |
| tensor<int32, [5]> var_688 = const()[name = tensor<string, []>("op_688"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_33_cast = reshape(shape = var_688, x = linear_32_cast)[name = tensor<string, []>("qkv_33_cast")]; | |
| tensor<int32, [5]> var_690_perm_0 = const()[name = tensor<string, []>("op_690_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]> transpose_12 = transpose(perm = var_690_perm_0, x = qkv_33_cast)[name = tensor<string, []>("transpose_12")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_33_cast = 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 = transpose_12)[name = tensor<string, []>("query_33_cast")]; | |
| 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 = 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 = transpose_12)[name = tensor<string, []>("key_33_cast")]; | |
| 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 = 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 = transpose_12)[name = tensor<string, []>("value_17_cast")]; | |
| tensor<fp16, []> var_701_to_fp16 = const()[name = tensor<string, []>("op_701_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_35_cast = mul(x = query_33_cast, y = var_701_to_fp16)[name = tensor<string, []>("query_35_cast")]; | |
| tensor<int32, [4]> key_35_perm_0 = const()[name = tensor<string, []>("key_35_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_33_transpose_x_0 = const()[name = tensor<string, []>("attn_33_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_33_transpose_y_0 = const()[name = tensor<string, []>("attn_33_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_11 = transpose(perm = key_35_perm_0, x = key_33_cast)[name = tensor<string, []>("transpose_11")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_33_cast = matmul(transpose_x = attn_33_transpose_x_0, transpose_y = attn_33_transpose_y_0, x = query_35_cast, y = transpose_11)[name = tensor<string, []>("attn_33_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_17_cast = softmax(axis = var_19, x = attn_33_cast)[name = tensor<string, []>("attn_as_float_17_cast")]; | |
| 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 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = attn_as_float_17_cast, y = value_17_cast)[name = tensor<string, []>("out_17_cast")]; | |
| tensor<int32, [4]> var_710_perm_0 = const()[name = tensor<string, []>("op_710_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_711 = const()[name = tensor<string, []>("op_711"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_10 = transpose(perm = var_710_perm_0, x = out_17_cast)[name = tensor<string, []>("transpose_10")]; | |
| tensor<fp16, [1, 77, 512]> input_185_cast = reshape(shape = var_711, x = transpose_10)[name = tensor<string, []>("input_185_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_185_cast)[name = tensor<string, []>("linear_33_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_53_cast = add(x = linear_33_cast, y = x_49_cast)[name = tensor<string, []>("x_53_cast")]; | |
| tensor<int32, [1]> var_725_axes_0 = const()[name = tensor<string, []>("op_725_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_725_cast = layer_norm(axes = var_725_axes_0, beta = original_model_text_encoder_transformer_8_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_8_pre_norm_ffn_0_weight_to_fp16, x = x_53_cast)[name = tensor<string, []>("op_725_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_8_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_ffn_1_weight_to_fp16, x = var_725_cast)[name = tensor<string, []>("linear_34_cast")]; | |
| 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 = gelu(mode = input_195_mode_0, x = linear_34_cast)[name = tensor<string, []>("input_195_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_8_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_8_pre_norm_ffn_4_weight_to_fp16, x = input_195_cast)[name = tensor<string, []>("linear_35_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_55_cast = add(x = x_53_cast, y = linear_35_cast)[name = tensor<string, []>("x_55_cast")]; | |
| tensor<int32, [1]> var_752_axes_0 = const()[name = tensor<string, []>("op_752_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_752_cast = layer_norm(axes = var_752_axes_0, beta = original_model_text_encoder_transformer_9_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_9_pre_norm_mha_0_weight_to_fp16, x = x_55_cast)[name = tensor<string, []>("op_752_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_752_cast)[name = tensor<string, []>("linear_36_cast")]; | |
| tensor<int32, [5]> var_764 = const()[name = tensor<string, []>("op_764"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_37_cast = reshape(shape = var_764, x = linear_36_cast)[name = tensor<string, []>("qkv_37_cast")]; | |
| tensor<int32, [5]> var_766_perm_0 = const()[name = tensor<string, []>("op_766_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]> transpose_9 = transpose(perm = var_766_perm_0, x = qkv_37_cast)[name = tensor<string, []>("transpose_9")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_37_cast = 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 = transpose_9)[name = tensor<string, []>("query_37_cast")]; | |
| 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 = 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 = transpose_9)[name = tensor<string, []>("key_37_cast")]; | |
| 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 = 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 = transpose_9)[name = tensor<string, []>("value_19_cast")]; | |
| tensor<fp16, []> var_777_to_fp16 = const()[name = tensor<string, []>("op_777_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_39_cast = mul(x = query_37_cast, y = var_777_to_fp16)[name = tensor<string, []>("query_39_cast")]; | |
| tensor<int32, [4]> key_39_perm_0 = const()[name = tensor<string, []>("key_39_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_37_transpose_x_0 = const()[name = tensor<string, []>("attn_37_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_37_transpose_y_0 = const()[name = tensor<string, []>("attn_37_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_8 = transpose(perm = key_39_perm_0, x = key_37_cast)[name = tensor<string, []>("transpose_8")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_37_cast = matmul(transpose_x = attn_37_transpose_x_0, transpose_y = attn_37_transpose_y_0, x = query_39_cast, y = transpose_8)[name = tensor<string, []>("attn_37_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_19_cast = softmax(axis = var_19, x = attn_37_cast)[name = tensor<string, []>("attn_as_float_19_cast")]; | |
| 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 = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = attn_as_float_19_cast, y = value_19_cast)[name = tensor<string, []>("out_19_cast")]; | |
| tensor<int32, [4]> var_786_perm_0 = const()[name = tensor<string, []>("op_786_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_787 = const()[name = tensor<string, []>("op_787"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_7 = transpose(perm = var_786_perm_0, x = out_19_cast)[name = tensor<string, []>("transpose_7")]; | |
| tensor<fp16, [1, 77, 512]> input_207_cast = reshape(shape = var_787, x = transpose_7)[name = tensor<string, []>("input_207_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_207_cast)[name = tensor<string, []>("linear_37_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_59_cast = add(x = linear_37_cast, y = x_55_cast)[name = tensor<string, []>("x_59_cast")]; | |
| tensor<int32, [1]> var_801_axes_0 = const()[name = tensor<string, []>("op_801_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_801_cast = layer_norm(axes = var_801_axes_0, beta = original_model_text_encoder_transformer_9_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_9_pre_norm_ffn_0_weight_to_fp16, x = x_59_cast)[name = tensor<string, []>("op_801_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_9_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_ffn_1_weight_to_fp16, x = var_801_cast)[name = tensor<string, []>("linear_38_cast")]; | |
| 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 = gelu(mode = input_217_mode_0, x = linear_38_cast)[name = tensor<string, []>("input_217_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_9_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_9_pre_norm_ffn_4_weight_to_fp16, x = input_217_cast)[name = tensor<string, []>("linear_39_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_61_cast = add(x = x_59_cast, y = linear_39_cast)[name = tensor<string, []>("x_61_cast")]; | |
| tensor<int32, [1]> var_828_axes_0 = const()[name = tensor<string, []>("op_828_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_828_cast = layer_norm(axes = var_828_axes_0, beta = original_model_text_encoder_transformer_10_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_10_pre_norm_mha_0_weight_to_fp16, x = x_61_cast)[name = tensor<string, []>("op_828_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_828_cast)[name = tensor<string, []>("linear_40_cast")]; | |
| tensor<int32, [5]> var_840 = const()[name = tensor<string, []>("op_840"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_41_cast = reshape(shape = var_840, x = linear_40_cast)[name = tensor<string, []>("qkv_41_cast")]; | |
| tensor<int32, [5]> var_842_perm_0 = const()[name = tensor<string, []>("op_842_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]> transpose_6 = transpose(perm = var_842_perm_0, x = qkv_41_cast)[name = tensor<string, []>("transpose_6")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_41_cast = 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 = transpose_6)[name = tensor<string, []>("query_41_cast")]; | |
| 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 = 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 = transpose_6)[name = tensor<string, []>("key_41_cast")]; | |
| 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 = 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 = transpose_6)[name = tensor<string, []>("value_21_cast")]; | |
| tensor<fp16, []> var_853_to_fp16 = const()[name = tensor<string, []>("op_853_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_43_cast = mul(x = query_41_cast, y = var_853_to_fp16)[name = tensor<string, []>("query_43_cast")]; | |
| tensor<int32, [4]> key_43_perm_0 = const()[name = tensor<string, []>("key_43_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_41_transpose_x_0 = const()[name = tensor<string, []>("attn_41_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_41_transpose_y_0 = const()[name = tensor<string, []>("attn_41_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_5 = transpose(perm = key_43_perm_0, x = key_41_cast)[name = tensor<string, []>("transpose_5")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_41_cast = matmul(transpose_x = attn_41_transpose_x_0, transpose_y = attn_41_transpose_y_0, x = query_43_cast, y = transpose_5)[name = tensor<string, []>("attn_41_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_21_cast = softmax(axis = var_19, x = attn_41_cast)[name = tensor<string, []>("attn_as_float_21_cast")]; | |
| 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 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = attn_as_float_21_cast, y = value_21_cast)[name = tensor<string, []>("out_21_cast")]; | |
| tensor<int32, [4]> var_862_perm_0 = const()[name = tensor<string, []>("op_862_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_863 = const()[name = tensor<string, []>("op_863"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_4 = transpose(perm = var_862_perm_0, x = out_21_cast)[name = tensor<string, []>("transpose_4")]; | |
| tensor<fp16, [1, 77, 512]> input_229_cast = reshape(shape = var_863, x = transpose_4)[name = tensor<string, []>("input_229_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_229_cast)[name = tensor<string, []>("linear_41_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_65_cast = add(x = linear_41_cast, y = x_61_cast)[name = tensor<string, []>("x_65_cast")]; | |
| tensor<int32, [1]> var_877_axes_0 = const()[name = tensor<string, []>("op_877_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_877_cast = layer_norm(axes = var_877_axes_0, beta = original_model_text_encoder_transformer_10_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_10_pre_norm_ffn_0_weight_to_fp16, x = x_65_cast)[name = tensor<string, []>("op_877_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_10_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_ffn_1_weight_to_fp16, x = var_877_cast)[name = tensor<string, []>("linear_42_cast")]; | |
| 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 = gelu(mode = input_239_mode_0, x = linear_42_cast)[name = tensor<string, []>("input_239_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_10_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_10_pre_norm_ffn_4_weight_to_fp16, x = input_239_cast)[name = tensor<string, []>("linear_43_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_67_cast = add(x = x_65_cast, y = linear_43_cast)[name = tensor<string, []>("x_67_cast")]; | |
| tensor<int32, [1]> var_904_axes_0 = const()[name = tensor<string, []>("op_904_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_904_cast = layer_norm(axes = var_904_axes_0, beta = original_model_text_encoder_transformer_11_pre_norm_mha_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_11_pre_norm_mha_0_weight_to_fp16, x = x_67_cast)[name = tensor<string, []>("op_904_cast")]; | |
| tensor<fp16, [1536, 512]> original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_mha_1_qkv_proj_weight_to_fp16, x = var_904_cast)[name = tensor<string, []>("linear_44_cast")]; | |
| tensor<int32, [5]> var_916 = const()[name = tensor<string, []>("op_916"), val = tensor<int32, [5]>([1, 77, 3, 8, -1])]; | |
| tensor<fp16, [1, 77, 3, 8, 64]> qkv_45_cast = reshape(shape = var_916, x = linear_44_cast)[name = tensor<string, []>("qkv_45_cast")]; | |
| tensor<int32, [5]> var_918_perm_0 = const()[name = tensor<string, []>("op_918_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]> transpose_3 = transpose(perm = var_918_perm_0, x = qkv_45_cast)[name = tensor<string, []>("transpose_3")]; | |
| tensor<fp16, [1, 8, 77, 64]> query_45_cast = 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 = transpose_3)[name = tensor<string, []>("query_45_cast")]; | |
| 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 = 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 = transpose_3)[name = tensor<string, []>("key_45_cast")]; | |
| 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 = slice_by_index(begin = value_begin_0, end = value_end_0, end_mask = value_end_mask_0, squeeze_mask = value_squeeze_mask_0, x = transpose_3)[name = tensor<string, []>("value_cast")]; | |
| tensor<fp16, []> var_929_to_fp16 = const()[name = tensor<string, []>("op_929_to_fp16"), val = tensor<fp16, []>(0x1p-3)]; | |
| tensor<fp16, [1, 8, 77, 64]> query_cast = mul(x = query_45_cast, y = var_929_to_fp16)[name = tensor<string, []>("query_cast")]; | |
| tensor<int32, [4]> key_perm_0 = const()[name = tensor<string, []>("key_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])]; | |
| tensor<bool, []> attn_45_transpose_x_0 = const()[name = tensor<string, []>("attn_45_transpose_x_0"), val = tensor<bool, []>(false)]; | |
| tensor<bool, []> attn_45_transpose_y_0 = const()[name = tensor<string, []>("attn_45_transpose_y_0"), val = tensor<bool, []>(false)]; | |
| tensor<fp16, [1, 8, 64, 77]> transpose_2 = transpose(perm = key_perm_0, x = key_45_cast)[name = tensor<string, []>("transpose_2")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_45_cast = matmul(transpose_x = attn_45_transpose_x_0, transpose_y = attn_45_transpose_y_0, x = query_cast, y = transpose_2)[name = tensor<string, []>("attn_45_cast")]; | |
| tensor<fp16, [1, 8, 77, 77]> attn_as_float_cast = softmax(axis = var_19, x = attn_45_cast)[name = tensor<string, []>("attn_as_float_cast")]; | |
| 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 = matmul(transpose_x = out_transpose_x_0, transpose_y = out_transpose_y_0, x = attn_as_float_cast, y = value_cast)[name = tensor<string, []>("out_cast")]; | |
| tensor<int32, [4]> var_938_perm_0 = const()[name = tensor<string, []>("op_938_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; | |
| tensor<int32, [3]> var_939 = const()[name = tensor<string, []>("op_939"), val = tensor<int32, [3]>([1, 77, -1])]; | |
| tensor<fp16, [1, 77, 8, 64]> transpose_1 = transpose(perm = var_938_perm_0, x = out_cast)[name = tensor<string, []>("transpose_1")]; | |
| tensor<fp16, [1, 77, 512]> input_251_cast = reshape(shape = var_939, x = transpose_1)[name = tensor<string, []>("input_251_cast")]; | |
| tensor<fp16, [512, 512]> original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_mha_1_out_proj_weight_to_fp16, x = input_251_cast)[name = tensor<string, []>("linear_45_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_71_cast = add(x = linear_45_cast, y = x_67_cast)[name = tensor<string, []>("x_71_cast")]; | |
| tensor<int32, [1]> var_953_axes_0 = const()[name = tensor<string, []>("op_953_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_953_cast = layer_norm(axes = var_953_axes_0, beta = original_model_text_encoder_transformer_11_pre_norm_ffn_0_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_transformer_11_pre_norm_ffn_0_weight_to_fp16, x = x_71_cast)[name = tensor<string, []>("op_953_cast")]; | |
| tensor<fp16, [2048, 512]> original_model_text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_11_pre_norm_ffn_1_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_ffn_1_weight_to_fp16, x = var_953_cast)[name = tensor<string, []>("linear_46_cast")]; | |
| 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 = gelu(mode = input_261_mode_0, x = linear_46_cast)[name = tensor<string, []>("input_261_cast")]; | |
| tensor<fp16, [512, 2048]> original_model_text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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 = linear(bias = original_model_text_encoder_transformer_11_pre_norm_ffn_4_bias_to_fp16, weight = original_model_text_encoder_transformer_11_pre_norm_ffn_4_weight_to_fp16, x = input_261_cast)[name = tensor<string, []>("linear_47_cast")]; | |
| tensor<fp16, [1, 77, 512]> x_cast = add(x = x_71_cast, y = linear_47_cast)[name = tensor<string, []>("x_cast")]; | |
| tensor<int32, [1]> var_975_axes_0 = const()[name = tensor<string, []>("op_975_axes_0"), val = tensor<int32, [1]>([-1])]; | |
| tensor<fp16, [512]> original_model_text_encoder_final_layer_norm_weight_to_fp16 = const()[name = tensor<string, []>("original_model_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]> original_model_text_encoder_final_layer_norm_bias_to_fp16 = const()[name = tensor<string, []>("original_model_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_975_cast = layer_norm(axes = var_975_axes_0, beta = original_model_text_encoder_final_layer_norm_bias_to_fp16, epsilon = var_6_to_fp16, gamma = original_model_text_encoder_final_layer_norm_weight_to_fp16, x = x_cast)[name = tensor<string, []>("op_975_cast")]; | |
| tensor<int32, [1]> var_978 = const()[name = tensor<string, []>("op_978"), val = tensor<int32, [1]>([0])]; | |
| tensor<int32, [1]> var_979 = reduce_argmax(axis = var_19, keep_dims = var_20, x = input_tokens)[name = tensor<string, []>("op_979")]; | |
| 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_978, var_979))[name = tensor<string, []>("stack_0")]; | |
| tensor<int32, []> token_emb_transpose_batch_dims_0 = const()[name = tensor<string, []>("token_emb_transpose_batch_dims_0"), val = tensor<int32, []>(0)]; | |
| tensor<fp16, [1, 512]> token_emb_transpose_cast = gather_nd(batch_dims = token_emb_transpose_batch_dims_0, indices = stack_0, x = var_975_cast)[name = tensor<string, []>("token_emb_transpose_cast")]; | |
| tensor<fp16, [512, 512]> var_982_weight_0_to_fp16 = const()[name = tensor<string, []>("op_982_weight_0_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126341440)))]; | |
| tensor<fp16, [512]> var_982_bias_0_to_fp16 = const()[name = tensor<string, []>("op_982_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]> var_982_cast = linear(bias = var_982_bias_0_to_fp16, weight = var_982_weight_0_to_fp16, x = token_emb_transpose_cast)[name = tensor<string, []>("op_982_cast")]; | |
| tensor<string, []> var_982_cast_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_982_cast_to_fp32_dtype_0"), val = tensor<string, []>("fp32")]; | |
| tensor<fp32, [1, 512]> text_embeddings = cast(dtype = var_982_cast_to_fp32_dtype_0, x = var_982_cast)[name = tensor<string, []>("cast_28")]; | |
| } -> (text_embeddings); | |
| } |