| program(1.0) |
| [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1.2"}, {"coremltools-component-torch", "2.7.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] |
| { |
| func main<ios17>(tensor<fp32, [1, 2, ?]> audio) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 2, 380160]}}), ("RangeDims", {{"audio", [[1, 1], [2, 2], [3840, 721920]]}})))] { |
| tensor<fp32, [1024, 8]> quantizers_14_codebook_weight = const()[name = tensor<string, []>("quantizers_14_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))]; |
| tensor<fp32, [1024, 8]> quantizers_13_codebook_weight = const()[name = tensor<string, []>("quantizers_13_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32896)))]; |
| tensor<fp32, [1024, 8]> quantizers_12_codebook_weight = const()[name = tensor<string, []>("quantizers_12_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65728)))]; |
| tensor<fp32, [1024, 8]> quantizers_11_codebook_weight = const()[name = tensor<string, []>("quantizers_11_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98560)))]; |
| tensor<fp32, [1024, 8]> quantizers_10_codebook_weight = const()[name = tensor<string, []>("quantizers_10_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131392)))]; |
| tensor<fp32, [1024, 8]> quantizers_9_codebook_weight = const()[name = tensor<string, []>("quantizers_9_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(164224)))]; |
| tensor<fp32, [1024, 8]> quantizers_8_codebook_weight = const()[name = tensor<string, []>("quantizers_8_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197056)))]; |
| tensor<fp32, [1024, 8]> quantizers_7_codebook_weight = const()[name = tensor<string, []>("quantizers_7_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229888)))]; |
| tensor<fp32, [1024, 8]> quantizers_6_codebook_weight = const()[name = tensor<string, []>("quantizers_6_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262720)))]; |
| tensor<fp32, [1024, 8]> quantizers_5_codebook_weight = const()[name = tensor<string, []>("quantizers_5_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295552)))]; |
| tensor<fp32, [1024, 8]> quantizers_4_codebook_weight = const()[name = tensor<string, []>("quantizers_4_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328384)))]; |
| tensor<fp32, [1024, 8]> quantizers_3_codebook_weight = const()[name = tensor<string, []>("quantizers_3_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361216)))]; |
| tensor<fp32, [1024, 8]> quantizers_2_codebook_weight = const()[name = tensor<string, []>("quantizers_2_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394048)))]; |
| tensor<fp32, [1024, 8]> quantizers_1_codebook_weight = const()[name = tensor<string, []>("quantizers_1_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426880)))]; |
| tensor<fp32, [1024, 8]> quantizers_0_codebook_weight = const()[name = tensor<string, []>("quantizers_0_codebook_weight"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(459712)))]; |
| tensor<fp32, [256, 240]> encoder_1_input_proj_weight = const()[name = tensor<string, []>("encoder_1_input_proj_weight"), val = tensor<fp32, [256, 240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(492544)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_0_norm1_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(738368)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_0_norm1_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(739456)))]; |
| tensor<fp32, [768, 256]> encoder_1_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(740544)))]; |
| tensor<fp32, [256, 256]> encoder_1_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1527040)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_0_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1789248)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_0_norm2_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1790336)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_0_norm2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1791424)))]; |
| tensor<fp32, [1024, 256]> encoder_1_transformer_layers_0_ffn_0_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1792512)))]; |
| tensor<fp32, [256, 1024]> encoder_1_transformer_layers_0_ffn_2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2841152)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_0_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_0_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3889792)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_1_norm1_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3890880)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_1_norm1_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3891968)))]; |
| tensor<fp32, [768, 256]> encoder_1_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3893056)))]; |
| tensor<fp32, [256, 256]> encoder_1_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4679552)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_1_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4941760)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_1_norm2_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4942848)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_1_norm2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4943936)))]; |
| tensor<fp32, [1024, 256]> encoder_1_transformer_layers_1_ffn_0_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4945024)))]; |
| tensor<fp32, [256, 1024]> encoder_1_transformer_layers_1_ffn_2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5993664)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_1_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_1_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7042304)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_2_norm1_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7043392)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_2_norm1_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7044480)))]; |
| tensor<fp32, [768, 256]> encoder_1_transformer_layers_2_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7045568)))]; |
| tensor<fp32, [256, 256]> encoder_1_transformer_layers_2_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(7832064)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_2_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8094272)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_2_norm2_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8095360)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_2_norm2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8096448)))]; |
| tensor<fp32, [1024, 256]> encoder_1_transformer_layers_2_ffn_0_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8097536)))]; |
| tensor<fp32, [256, 1024]> encoder_1_transformer_layers_2_ffn_2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9146176)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_2_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_2_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10194816)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_3_norm1_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10195904)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_3_norm1_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10196992)))]; |
| tensor<fp32, [768, 256]> encoder_1_transformer_layers_3_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10198080)))]; |
| tensor<fp32, [256, 256]> encoder_1_transformer_layers_3_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10984576)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_3_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11246784)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_3_norm2_bias = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11247872)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_3_norm2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11248960)))]; |
| tensor<fp32, [1024, 256]> encoder_1_transformer_layers_3_ffn_0_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(11250048)))]; |
| tensor<fp32, [256, 1024]> encoder_1_transformer_layers_3_ffn_2_weight = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12298688)))]; |
| tensor<fp32, [256]> encoder_1_transformer_layers_3_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_1_transformer_layers_3_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13347328)))]; |
| tensor<fp32, [384, 256]> encoder_1_output_proj_weight = const()[name = tensor<string, []>("encoder_1_output_proj_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13348416)))]; |
| tensor<fp32, [256, 768]> encoder_3_input_proj_weight = const()[name = tensor<string, []>("encoder_3_input_proj_weight"), val = tensor<fp32, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13741696)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_0_norm1_bias = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14528192)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_0_norm1_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14529280)))]; |
| tensor<fp32, [768, 256]> encoder_3_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14530368)))]; |
| tensor<fp32, [256, 256]> encoder_3_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15316864)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_0_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15579072)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_0_norm2_bias = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15580160)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_0_norm2_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15581248)))]; |
| tensor<fp32, [1024, 256]> encoder_3_transformer_layers_0_ffn_0_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15582336)))]; |
| tensor<fp32, [256, 1024]> encoder_3_transformer_layers_0_ffn_2_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16630976)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_0_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_3_transformer_layers_0_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17679616)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_1_norm1_bias = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17680704)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_1_norm1_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17681792)))]; |
| tensor<fp32, [768, 256]> encoder_3_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17682880)))]; |
| tensor<fp32, [256, 256]> encoder_3_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18469376)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_1_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18731584)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_1_norm2_bias = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18732672)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_1_norm2_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18733760)))]; |
| tensor<fp32, [1024, 256]> encoder_3_transformer_layers_1_ffn_0_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(18734848)))]; |
| tensor<fp32, [256, 1024]> encoder_3_transformer_layers_1_ffn_2_weight = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19783488)))]; |
| tensor<fp32, [256]> encoder_3_transformer_layers_1_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_3_transformer_layers_1_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20832128)))]; |
| tensor<fp32, [384, 256]> encoder_3_output_proj_weight = const()[name = tensor<string, []>("encoder_3_output_proj_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20833216)))]; |
| tensor<fp32, [256, 768]> encoder_5_input_proj_weight = const()[name = tensor<string, []>("encoder_5_input_proj_weight"), val = tensor<fp32, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(21226496)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_0_norm1_bias = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22012992)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_0_norm1_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22014080)))]; |
| tensor<fp32, [768, 256]> encoder_5_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22015168)))]; |
| tensor<fp32, [256, 256]> encoder_5_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22801664)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_0_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23063872)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_0_norm2_bias = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23064960)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_0_norm2_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23066048)))]; |
| tensor<fp32, [1024, 256]> encoder_5_transformer_layers_0_ffn_0_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23067136)))]; |
| tensor<fp32, [256, 1024]> encoder_5_transformer_layers_0_ffn_2_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24115776)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_0_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_5_transformer_layers_0_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25164416)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_1_norm1_bias = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25165504)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_1_norm1_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25166592)))]; |
| tensor<fp32, [768, 256]> encoder_5_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25167680)))]; |
| tensor<fp32, [256, 256]> encoder_5_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(25954176)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_1_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26216384)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_1_norm2_bias = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26217472)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_1_norm2_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26218560)))]; |
| tensor<fp32, [1024, 256]> encoder_5_transformer_layers_1_ffn_0_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(26219648)))]; |
| tensor<fp32, [256, 1024]> encoder_5_transformer_layers_1_ffn_2_weight = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(27268288)))]; |
| tensor<fp32, [256]> encoder_5_transformer_layers_1_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_5_transformer_layers_1_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28316928)))]; |
| tensor<fp32, [384, 256]> encoder_5_output_proj_weight = const()[name = tensor<string, []>("encoder_5_output_proj_weight"), val = tensor<fp32, [384, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28318016)))]; |
| tensor<fp32, [256, 768]> encoder_7_input_proj_weight = const()[name = tensor<string, []>("encoder_7_input_proj_weight"), val = tensor<fp32, [256, 768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28711296)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_0_norm1_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29497792)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_0_norm1_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29498880)))]; |
| tensor<fp32, [768, 256]> encoder_7_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(29499968)))]; |
| tensor<fp32, [256, 256]> encoder_7_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30286464)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_0_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30548672)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_0_norm2_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30549760)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_0_norm2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30550848)))]; |
| tensor<fp32, [1024, 256]> encoder_7_transformer_layers_0_ffn_0_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(30551936)))]; |
| tensor<fp32, [256, 1024]> encoder_7_transformer_layers_0_ffn_2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(31600576)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_0_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_0_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32649216)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_1_norm1_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32650304)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_1_norm1_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32651392)))]; |
| tensor<fp32, [768, 256]> encoder_7_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(32652480)))]; |
| tensor<fp32, [256, 256]> encoder_7_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33438976)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_1_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33701184)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_1_norm2_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33702272)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_1_norm2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33703360)))]; |
| tensor<fp32, [1024, 256]> encoder_7_transformer_layers_1_ffn_0_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33704448)))]; |
| tensor<fp32, [256, 1024]> encoder_7_transformer_layers_1_ffn_2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(34753088)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_1_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_1_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35801728)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_2_norm1_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35802816)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_2_norm1_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35803904)))]; |
| tensor<fp32, [768, 256]> encoder_7_transformer_layers_2_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(35804992)))]; |
| tensor<fp32, [256, 256]> encoder_7_transformer_layers_2_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36591488)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_2_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36853696)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_2_norm2_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36854784)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_2_norm2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36855872)))]; |
| tensor<fp32, [1024, 256]> encoder_7_transformer_layers_2_ffn_0_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(36856960)))]; |
| tensor<fp32, [256, 1024]> encoder_7_transformer_layers_2_ffn_2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37905600)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_2_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_2_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38954240)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_3_norm1_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_norm1_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38955328)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_3_norm1_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_norm1_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38956416)))]; |
| tensor<fp32, [768, 256]> encoder_7_transformer_layers_3_self_attn_in_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_self_attn_in_proj_weight"), val = tensor<fp32, [768, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38957504)))]; |
| tensor<fp32, [256, 256]> encoder_7_transformer_layers_3_self_attn_out_proj_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_self_attn_out_proj_weight"), val = tensor<fp32, [256, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39744000)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_3_layer_scale_1_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_layer_scale_1_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40006208)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_3_norm2_bias = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_norm2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40007296)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_3_norm2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_norm2_weight"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40008384)))]; |
| tensor<fp32, [1024, 256]> encoder_7_transformer_layers_3_ffn_0_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_ffn_0_weight"), val = tensor<fp32, [1024, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40009472)))]; |
| tensor<fp32, [256, 1024]> encoder_7_transformer_layers_3_ffn_2_weight = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_ffn_2_weight"), val = tensor<fp32, [256, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41058112)))]; |
| tensor<fp32, [256]> encoder_7_transformer_layers_3_layer_scale_2_scale = const()[name = tensor<string, []>("encoder_7_transformer_layers_3_layer_scale_2_scale"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42106752)))]; |
| tensor<fp32, [192, 256]> encoder_7_output_proj_weight = const()[name = tensor<string, []>("encoder_7_output_proj_weight"), val = tensor<fp32, [192, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42107840)))]; |
| tensor<fp32, [512]> input_proj_bias = const()[name = tensor<string, []>("input_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42304512)))]; |
| tensor<fp32, [512, 768, 1]> input_proj_weight = const()[name = tensor<string, []>("input_proj_weight"), val = tensor<fp32, [512, 768, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(42306624)))]; |
| tensor<fp32, [8]> quantizers_0_in_proj_bias = const()[name = tensor<string, []>("quantizers_0_in_proj_bias"), val = tensor<fp32, [8]>([-0x1.aad71p-5, -0x1.926dfp-2, -0x1.618ad4p-3, 0x1.3657a2p-2, 0x1.7bbce4p-2, -0x1.cc2bf6p-3, 0x1.a6ab9ep-2, 0x1.46015ep-2])]; |
| tensor<fp32, [8, 512, 1]> quantizers_0_in_proj_weight = const()[name = tensor<string, []>("quantizers_0_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43879552)))]; |
| tensor<fp32, [512]> quantizers_0_out_proj_bias = const()[name = tensor<string, []>("quantizers_0_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43896000)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_0_out_proj_weight = const()[name = tensor<string, []>("quantizers_0_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43898112)))]; |
| tensor<fp32, [8]> quantizers_1_in_proj_bias = const()[name = tensor<string, []>("quantizers_1_in_proj_bias"), val = tensor<fp32, [8]>([-0x1.068218p-6, 0x1.0dffa8p-3, 0x1.b9029ap-2, -0x1.4103cap-5, -0x1.67340ep-1, -0x1.1574a6p-5, -0x1.acc404p-2, -0x1.6b86eap-4])]; |
| tensor<fp32, [8, 512, 1]> quantizers_1_in_proj_weight = const()[name = tensor<string, []>("quantizers_1_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43914560)))]; |
| tensor<fp32, [512]> quantizers_1_out_proj_bias = const()[name = tensor<string, []>("quantizers_1_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43931008)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_1_out_proj_weight = const()[name = tensor<string, []>("quantizers_1_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43933120)))]; |
| tensor<fp32, [8]> quantizers_2_in_proj_bias = const()[name = tensor<string, []>("quantizers_2_in_proj_bias"), val = tensor<fp32, [8]>([-0x1.43c7ccp-3, 0x1.d4f8a4p-7, 0x1.6b8472p-3, 0x1.4972d8p-3, 0x1.5e02e4p-1, 0x1.c24a28p-2, -0x1.4aee78p-2, -0x1.817c3cp-3])]; |
| tensor<fp32, [8, 512, 1]> quantizers_2_in_proj_weight = const()[name = tensor<string, []>("quantizers_2_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43949568)))]; |
| tensor<fp32, [512]> quantizers_2_out_proj_bias = const()[name = tensor<string, []>("quantizers_2_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43966016)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_2_out_proj_weight = const()[name = tensor<string, []>("quantizers_2_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43968128)))]; |
| tensor<fp32, [8]> quantizers_3_in_proj_bias = const()[name = tensor<string, []>("quantizers_3_in_proj_bias"), val = tensor<fp32, [8]>([0x1.7e8c2ap-2, 0x1.cefdap-4, -0x1.827cbcp-3, -0x1.6803a2p-6, -0x1.f38168p-2, -0x1.2dac62p-1, 0x1.d17f3ep-6, 0x1.0f510ep-4])]; |
| tensor<fp32, [8, 512, 1]> quantizers_3_in_proj_weight = const()[name = tensor<string, []>("quantizers_3_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43984576)))]; |
| tensor<fp32, [512]> quantizers_3_out_proj_bias = const()[name = tensor<string, []>("quantizers_3_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44001024)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_3_out_proj_weight = const()[name = tensor<string, []>("quantizers_3_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44003136)))]; |
| tensor<fp32, [8]> quantizers_4_in_proj_bias = const()[name = tensor<string, []>("quantizers_4_in_proj_bias"), val = tensor<fp32, [8]>([-0x1.01d078p-3, -0x1.ebb352p-2, 0x1.125e06p-8, -0x1.e1df3p-3, 0x1.3c49b4p-4, 0x1.33094ap-1, -0x1.9c89dep-6, 0x1.00beacp-2])]; |
| tensor<fp32, [8, 512, 1]> quantizers_4_in_proj_weight = const()[name = tensor<string, []>("quantizers_4_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44019584)))]; |
| tensor<fp32, [512]> quantizers_4_out_proj_bias = const()[name = tensor<string, []>("quantizers_4_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44036032)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_4_out_proj_weight = const()[name = tensor<string, []>("quantizers_4_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44038144)))]; |
| tensor<fp32, [8]> quantizers_5_in_proj_bias = const()[name = tensor<string, []>("quantizers_5_in_proj_bias"), val = tensor<fp32, [8]>([0x1.07f572p-2, -0x1.cb3c6cp-1, -0x1.4f4152p-10, -0x1.29f5e8p-5, 0x1.5ad12ap-10, -0x1.35969ep-2, 0x1.daf468p-3, -0x1.5e52f4p-5])]; |
| tensor<fp32, [8, 512, 1]> quantizers_5_in_proj_weight = const()[name = tensor<string, []>("quantizers_5_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44054592)))]; |
| tensor<fp32, [512]> quantizers_5_out_proj_bias = const()[name = tensor<string, []>("quantizers_5_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44071040)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_5_out_proj_weight = const()[name = tensor<string, []>("quantizers_5_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44073152)))]; |
| tensor<fp32, [8]> quantizers_6_in_proj_bias = const()[name = tensor<string, []>("quantizers_6_in_proj_bias"), val = tensor<fp32, [8]>([0x1.577e0ap-3, 0x1.0d17c4p-1, -0x1.027b14p-1, -0x1.11a696p-4, 0x1.81a4bp-5, -0x1.05451cp-6, -0x1.af9232p-4, -0x1.bbe2p-4])]; |
| tensor<fp32, [8, 512, 1]> quantizers_6_in_proj_weight = const()[name = tensor<string, []>("quantizers_6_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44089600)))]; |
| tensor<fp32, [512]> quantizers_6_out_proj_bias = const()[name = tensor<string, []>("quantizers_6_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44106048)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_6_out_proj_weight = const()[name = tensor<string, []>("quantizers_6_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44108160)))]; |
| tensor<fp32, [8]> quantizers_7_in_proj_bias = const()[name = tensor<string, []>("quantizers_7_in_proj_bias"), val = tensor<fp32, [8]>([0x1.c5396cp-6, -0x1.91349ep-5, 0x1.2d837ep-4, 0x1.a8dadep-2, -0x1.4d7122p-3, -0x1.5e248p-2, 0x1.cc1646p-2, 0x1.77c976p-3])]; |
| tensor<fp32, [8, 512, 1]> quantizers_7_in_proj_weight = const()[name = tensor<string, []>("quantizers_7_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44124608)))]; |
| tensor<fp32, [512]> quantizers_7_out_proj_bias = const()[name = tensor<string, []>("quantizers_7_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44141056)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_7_out_proj_weight = const()[name = tensor<string, []>("quantizers_7_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44143168)))]; |
| tensor<fp32, [8]> quantizers_8_in_proj_bias = const()[name = tensor<string, []>("quantizers_8_in_proj_bias"), val = tensor<fp32, [8]>([-0x1.3399a4p-4, 0x1.2c04dep-5, -0x1.71f218p-4, -0x1.d5b79p-3, -0x1.4aa72cp-7, 0x1.2b9226p-1, 0x1.9aa08cp-3, -0x1.1ae422p-4])]; |
| tensor<fp32, [8, 512, 1]> quantizers_8_in_proj_weight = const()[name = tensor<string, []>("quantizers_8_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44159616)))]; |
| tensor<fp32, [512]> quantizers_8_out_proj_bias = const()[name = tensor<string, []>("quantizers_8_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44176064)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_8_out_proj_weight = const()[name = tensor<string, []>("quantizers_8_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44178176)))]; |
| tensor<fp32, [8]> quantizers_9_in_proj_bias = const()[name = tensor<string, []>("quantizers_9_in_proj_bias"), val = tensor<fp32, [8]>([0x1.ef8316p-4, 0x1.10205p-13, 0x1.6761bap-4, 0x1.8ebcd6p-4, 0x1.38b1d8p-2, 0x1.2476d4p-4, -0x1.48ccdcp-6, -0x1.46fe08p-1])]; |
| tensor<fp32, [8, 512, 1]> quantizers_9_in_proj_weight = const()[name = tensor<string, []>("quantizers_9_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44194624)))]; |
| tensor<fp32, [512]> quantizers_9_out_proj_bias = const()[name = tensor<string, []>("quantizers_9_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44211072)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_9_out_proj_weight = const()[name = tensor<string, []>("quantizers_9_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44213184)))]; |
| tensor<fp32, [8]> quantizers_10_in_proj_bias = const()[name = tensor<string, []>("quantizers_10_in_proj_bias"), val = tensor<fp32, [8]>([0x1.603596p-2, -0x1.908814p-2, 0x1.7f909cp-7, -0x1.4772b6p-2, -0x1.6c1b38p-3, 0x1.775ecep-4, -0x1.65faa6p-4, -0x1.954f6ap-5])]; |
| tensor<fp32, [8, 512, 1]> quantizers_10_in_proj_weight = const()[name = tensor<string, []>("quantizers_10_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44229632)))]; |
| tensor<fp32, [512]> quantizers_10_out_proj_bias = const()[name = tensor<string, []>("quantizers_10_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44246080)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_10_out_proj_weight = const()[name = tensor<string, []>("quantizers_10_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44248192)))]; |
| tensor<fp32, [8]> quantizers_11_in_proj_bias = const()[name = tensor<string, []>("quantizers_11_in_proj_bias"), val = tensor<fp32, [8]>([0x1.cc1d98p-3, -0x1.205982p-5, -0x1.5c1a82p-2, 0x1.321cfcp-2, -0x1.05bec8p-3, 0x1.98dd4p-6, -0x1.e5325ep-5, -0x1.5f59c4p-2])]; |
| tensor<fp32, [8, 512, 1]> quantizers_11_in_proj_weight = const()[name = tensor<string, []>("quantizers_11_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44264640)))]; |
| tensor<fp32, [512]> quantizers_11_out_proj_bias = const()[name = tensor<string, []>("quantizers_11_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44281088)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_11_out_proj_weight = const()[name = tensor<string, []>("quantizers_11_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44283200)))]; |
| tensor<fp32, [8]> quantizers_12_in_proj_bias = const()[name = tensor<string, []>("quantizers_12_in_proj_bias"), val = tensor<fp32, [8]>([0x1.4ebddcp-3, -0x1.462922p-3, -0x1.f4f44ap-6, 0x1.b6ef86p-4, -0x1.7cb57ap-3, -0x1.f3eaf4p-6, 0x1.b30e8cp-3, -0x1.d3b4c4p-2])]; |
| tensor<fp32, [8, 512, 1]> quantizers_12_in_proj_weight = const()[name = tensor<string, []>("quantizers_12_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44299648)))]; |
| tensor<fp32, [512]> quantizers_12_out_proj_bias = const()[name = tensor<string, []>("quantizers_12_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44316096)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_12_out_proj_weight = const()[name = tensor<string, []>("quantizers_12_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44318208)))]; |
| tensor<fp32, [8]> quantizers_13_in_proj_bias = const()[name = tensor<string, []>("quantizers_13_in_proj_bias"), val = tensor<fp32, [8]>([0x1.b15388p-4, -0x1.3eb026p-5, 0x1.016644p-1, -0x1.2ef6bep-9, 0x1.3e522cp-6, 0x1.16820ap-3, 0x1.2d9788p-4, 0x1.5a9cc8p-7])]; |
| tensor<fp32, [8, 512, 1]> quantizers_13_in_proj_weight = const()[name = tensor<string, []>("quantizers_13_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44334656)))]; |
| tensor<fp32, [512]> quantizers_13_out_proj_bias = const()[name = tensor<string, []>("quantizers_13_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44351104)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_13_out_proj_weight = const()[name = tensor<string, []>("quantizers_13_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44353216)))]; |
| tensor<fp32, [8]> quantizers_14_in_proj_bias = const()[name = tensor<string, []>("quantizers_14_in_proj_bias"), val = tensor<fp32, [8]>([-0x1.f1d428p-2, -0x1.6a9354p-3, 0x1.4750ccp-3, 0x1.03d804p-5, -0x1.fa3136p-5, 0x1.6cc39cp-8, -0x1.1ff49ep-2, -0x1.7e81c6p-3])]; |
| tensor<fp32, [8, 512, 1]> quantizers_14_in_proj_weight = const()[name = tensor<string, []>("quantizers_14_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44369664)))]; |
| tensor<fp32, [512]> quantizers_14_out_proj_bias = const()[name = tensor<string, []>("quantizers_14_out_proj_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44386112)))]; |
| tensor<fp32, [512, 8, 1]> quantizers_14_out_proj_weight = const()[name = tensor<string, []>("quantizers_14_out_proj_weight"), val = tensor<fp32, [512, 8, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44388224)))]; |
| tensor<fp32, [8]> quantizers_15_in_proj_bias = const()[name = tensor<string, []>("quantizers_15_in_proj_bias"), val = tensor<fp32, [8]>([0x1.662e24p-6, -0x1.f4f232p-6, 0x1.60a78ep-2, 0x1.3a9adp-3, 0x1.8d7758p-7, 0x1.e51944p-6, -0x1.474cd4p-3, -0x1.ffd584p-10])]; |
| tensor<fp32, [8, 512, 1]> quantizers_15_in_proj_weight = const()[name = tensor<string, []>("quantizers_15_in_proj_weight"), val = tensor<fp32, [8, 512, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44404672)))]; |
| tensor<int32, [3]> var_215_perm_0 = const()[name = tensor<string, []>("op_215_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [3]> var_222 = const()[name = tensor<string, []>("op_222"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp32, [1, ?, 2]> var_215 = transpose(perm = var_215_perm_0, x = audio)[name = tensor<string, []>("transpose_84")]; |
| tensor<fp32, [1, 1, ?]> x_1 = reshape(shape = var_222, x = var_215)[name = tensor<string, []>("x_1")]; |
| tensor<int32, [3]> var_231_begin_0 = const()[name = tensor<string, []>("op_231_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> var_231_end_0 = const()[name = tensor<string, []>("op_231_end_0"), val = tensor<int32, [3]>([1, 1, 0])]; |
| tensor<bool, [3]> var_231_end_mask_0 = const()[name = tensor<string, []>("op_231_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; |
| tensor<bool, [3]> var_231_squeeze_mask_0 = const()[name = tensor<string, []>("op_231_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])]; |
| tensor<fp32, [1, ?]> var_231 = slice_by_index(begin = var_231_begin_0, end = var_231_end_0, end_mask = var_231_end_mask_0, squeeze_mask = var_231_squeeze_mask_0, x = x_1)[name = tensor<string, []>("op_231")]; |
| tensor<fp32, []> fill_like_0_value_0 = const()[name = tensor<string, []>("fill_like_0_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?]> fill_like_0 = fill_like(ref_tensor = var_231, value = fill_like_0_value_0)[name = tensor<string, []>("fill_like_0")]; |
| tensor<int32, [1]> var_247_axes_0 = const()[name = tensor<string, []>("op_247_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_247_keep_dims_0 = const()[name = tensor<string, []>("op_247_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1]> var_247 = reduce_sum(axes = var_247_axes_0, keep_dims = var_247_keep_dims_0, x = fill_like_0)[name = tensor<string, []>("op_247")]; |
| tensor<string, []> var_252_dtype_0 = const()[name = tensor<string, []>("op_252_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, []> var_253 = const()[name = tensor<string, []>("op_253"), val = tensor<int32, []>(240)]; |
| tensor<int32, [4]> var_263 = const()[name = tensor<string, []>("op_263"), val = tensor<int32, [4]>([1, 1, -1, 240])]; |
| tensor<fp32, [1, 1, ?, 240]> var_264 = reshape(shape = var_263, x = x_1)[name = tensor<string, []>("op_264")]; |
| tensor<int32, [4]> var_265 = const()[name = tensor<string, []>("op_265"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| tensor<int32, [3]> var_270 = const()[name = tensor<string, []>("op_270"), val = tensor<int32, [3]>([1, 240, -1])]; |
| tensor<fp32, [1, 1, 240, ?]> var_266 = transpose(perm = var_265, x = var_264)[name = tensor<string, []>("transpose_83")]; |
| tensor<fp32, [1, 240, ?]> x_3 = reshape(shape = var_270, x = var_266)[name = tensor<string, []>("x_3")]; |
| tensor<int32, [1]> var_252 = cast(dtype = var_252_dtype_0, x = var_247)[name = tensor<string, []>("cast_478")]; |
| tensor<int32, [1]> floor_div_0 = floor_div(x = var_252, y = var_253)[name = tensor<string, []>("floor_div_0")]; |
| tensor<string, []> input_lengths_1_dtype_0 = const()[name = tensor<string, []>("input_lengths_1_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<int32, []> var_277 = const()[name = tensor<string, []>("op_277"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> var_284 = const()[name = tensor<string, []>("op_284"), val = tensor<int32, []>(1600)]; |
| tensor<fp32, []> var_291 = const()[name = tensor<string, []>("op_291"), val = tensor<fp32, []>(0x1.4f8b58p-17)]; |
| tensor<int32, []> var_295 = const()[name = tensor<string, []>("op_295"), val = tensor<int32, []>(1)]; |
| tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [256]> linear_0_bias_0 = const()[name = tensor<string, []>("linear_0_bias_0"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44421120)))]; |
| tensor<fp32, [1, ?, 240]> input_1 = transpose(perm = input_1_perm_0, x = x_3)[name = tensor<string, []>("transpose_82")]; |
| tensor<fp32, [1, ?, 256]> x_5 = linear(bias = linear_0_bias_0, weight = encoder_1_input_proj_weight, x = input_1)[name = tensor<string, []>("linear_0")]; |
| tensor<int32, [1]> x_7_axes_0 = const()[name = tensor<string, []>("x_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_1_transformer_layers_0_norm1_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_0_norm1_weight, x = x_5)[name = tensor<string, []>("x_7")]; |
| tensor<int32, [3]> var_323_shape = shape(x = x_7)[name = tensor<string, []>("op_323_shape")]; |
| tensor<int32, []> gather_4_batch_dims_0 = const()[name = tensor<string, []>("gather_4_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_4_validate_indices_0 = const()[name = tensor<string, []>("gather_4_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_16 = const()[name = tensor<string, []>("select_16"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_4_axis_1 = const()[name = tensor<string, []>("gather_4_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_4 = gather(axis = gather_4_axis_1, batch_dims = gather_4_batch_dims_0, indices = select_16, validate_indices = gather_4_validate_indices_0, x = var_323_shape)[name = tensor<string, []>("gather_4")]; |
| tensor<fp32, [768]> linear_1_bias_0 = const()[name = tensor<string, []>("linear_1_bias_0"), val = tensor<fp32, [768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44422208)))]; |
| tensor<fp32, [1, ?, 768]> projected_1 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_0_self_attn_in_proj_weight, x = x_7)[name = tensor<string, []>("linear_1")]; |
| tensor<int32, [5]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_330 = reshape(shape = concat_0x, x = projected_1)[name = tensor<string, []>("op_330")]; |
| tensor<int32, [5]> var_331 = const()[name = tensor<string, []>("op_331"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_1_begin_0 = const()[name = tensor<string, []>("q_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_1_end_0 = const()[name = tensor<string, []>("q_1_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_1_end_mask_0 = const()[name = tensor<string, []>("q_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_1_squeeze_mask_0 = const()[name = tensor<string, []>("q_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_3 = transpose(perm = var_331, x = var_330)[name = tensor<string, []>("transpose_81")]; |
| tensor<fp32, [1, 4, ?, 64]> q_1 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = projected_3)[name = tensor<string, []>("q_1")]; |
| tensor<int32, [5]> k_1_begin_0 = const()[name = tensor<string, []>("k_1_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_1_end_0 = const()[name = tensor<string, []>("k_1_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_1_end_mask_0 = const()[name = tensor<string, []>("k_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_1_squeeze_mask_0 = const()[name = tensor<string, []>("k_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_1 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = projected_3)[name = tensor<string, []>("k_1")]; |
| tensor<int32, [5]> v_1_begin_0 = const()[name = tensor<string, []>("v_1_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_1_end_0 = const()[name = tensor<string, []>("v_1_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_1_end_mask_0 = const()[name = tensor<string, []>("v_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_1_squeeze_mask_0 = const()[name = tensor<string, []>("v_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_1 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = projected_3)[name = tensor<string, []>("v_1")]; |
| tensor<fp32, [32]> freqs_1 = const()[name = tensor<string, []>("freqs_1"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44425344)))]; |
| tensor<int32, [4]> var_343_shape = shape(x = q_1)[name = tensor<string, []>("op_343_shape")]; |
| tensor<int32, []> gather_8_batch_dims_0 = const()[name = tensor<string, []>("gather_8_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_8_validate_indices_0 = const()[name = tensor<string, []>("gather_8_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_17 = const()[name = tensor<string, []>("select_17"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_8_axis_1 = const()[name = tensor<string, []>("gather_8_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_8 = gather(axis = gather_8_axis_1, batch_dims = gather_8_batch_dims_0, indices = select_17, validate_indices = gather_8_validate_indices_0, x = var_343_shape)[name = tensor<string, []>("gather_8")]; |
| tensor<fp32, [1, 1]> var_346 = const()[name = tensor<string, []>("op_346"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_4 = const()[name = tensor<string, []>("const_4"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_347 = range_1d(end = gather_8, start = const_3, step = const_4)[name = tensor<string, []>("op_347")]; |
| tensor<string, []> var_347_promoted_dtype_0 = const()[name = tensor<string, []>("op_347_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_347_promoted = cast(dtype = var_347_promoted_dtype_0, x = var_347)[name = tensor<string, []>("cast_476")]; |
| tensor<fp32, [1, ?]> ts_1 = add(x = var_346, y = var_347_promoted)[name = tensor<string, []>("ts_1")]; |
| tensor<int32, [4]> var_350 = const()[name = tensor<string, []>("op_350"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_3 = reshape(shape = var_350, x = ts_1)[name = tensor<string, []>("ts_3")]; |
| tensor<int32, [5]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_1 = reshape(shape = concat_1x, x = q_1)[name = tensor<string, []>("q2_1")]; |
| tensor<int32, [5]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_1 = reshape(shape = concat_2x, x = k_1)[name = tensor<string, []>("k2_1")]; |
| tensor<int32, [5]> qr_1_begin_0 = const()[name = tensor<string, []>("qr_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_1_end_0 = const()[name = tensor<string, []>("qr_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_1_end_mask_0 = const()[name = tensor<string, []>("qr_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_1_squeeze_mask_0 = const()[name = tensor<string, []>("qr_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_1 = slice_by_index(begin = qr_1_begin_0, end = qr_1_end_0, end_mask = qr_1_end_mask_0, squeeze_mask = qr_1_squeeze_mask_0, x = q2_1)[name = tensor<string, []>("qr_1")]; |
| tensor<int32, [5]> qi_1_begin_0 = const()[name = tensor<string, []>("qi_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_1_end_0 = const()[name = tensor<string, []>("qi_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_1_end_mask_0 = const()[name = tensor<string, []>("qi_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_1_squeeze_mask_0 = const()[name = tensor<string, []>("qi_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_1 = slice_by_index(begin = qi_1_begin_0, end = qi_1_end_0, end_mask = qi_1_end_mask_0, squeeze_mask = qi_1_squeeze_mask_0, x = q2_1)[name = tensor<string, []>("qi_1")]; |
| tensor<int32, [5]> kr_1_begin_0 = const()[name = tensor<string, []>("kr_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_1_end_0 = const()[name = tensor<string, []>("kr_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_1_end_mask_0 = const()[name = tensor<string, []>("kr_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_1_squeeze_mask_0 = const()[name = tensor<string, []>("kr_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_1 = slice_by_index(begin = kr_1_begin_0, end = kr_1_end_0, end_mask = kr_1_end_mask_0, squeeze_mask = kr_1_squeeze_mask_0, x = k2_1)[name = tensor<string, []>("kr_1")]; |
| tensor<int32, [5]> ki_1_begin_0 = const()[name = tensor<string, []>("ki_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_1_end_0 = const()[name = tensor<string, []>("ki_1_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_1_end_mask_0 = const()[name = tensor<string, []>("ki_1_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_1_squeeze_mask_0 = const()[name = tensor<string, []>("ki_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_1 = slice_by_index(begin = ki_1_begin_0, end = ki_1_end_0, end_mask = ki_1_end_mask_0, squeeze_mask = ki_1_squeeze_mask_0, x = k2_1)[name = tensor<string, []>("ki_1")]; |
| tensor<fp32, [1, 1, ?, 32]> var_363 = mul(x = freqs_1, y = ts_3)[name = tensor<string, []>("op_363")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_1 = cos(x = var_363)[name = tensor<string, []>("rotr_1")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_1 = sin(x = var_363)[name = tensor<string, []>("roti_1")]; |
| tensor<fp32, [1, 4, ?, 32]> var_367 = mul(x = qr_1, y = rotr_1)[name = tensor<string, []>("op_367")]; |
| tensor<fp32, [1, 4, ?, 32]> var_368 = mul(x = qi_1, y = roti_1)[name = tensor<string, []>("op_368")]; |
| tensor<fp32, [1, 4, ?, 32]> var_369 = sub(x = var_367, y = var_368)[name = tensor<string, []>("op_369")]; |
| tensor<fp32, [1, 4, ?, 32]> var_370 = mul(x = qr_1, y = roti_1)[name = tensor<string, []>("op_370")]; |
| tensor<fp32, [1, 4, ?, 32]> var_371 = mul(x = qi_1, y = rotr_1)[name = tensor<string, []>("op_371")]; |
| tensor<fp32, [1, 4, ?, 32]> var_372 = add(x = var_370, y = var_371)[name = tensor<string, []>("op_372")]; |
| tensor<int32, []> qo_1_axis_0 = const()[name = tensor<string, []>("qo_1_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_1 = stack(axis = qo_1_axis_0, values = (var_369, var_372))[name = tensor<string, []>("qo_1")]; |
| tensor<fp32, [1, 4, ?, 32]> var_375 = mul(x = kr_1, y = rotr_1)[name = tensor<string, []>("op_375")]; |
| tensor<fp32, [1, 4, ?, 32]> var_376 = mul(x = ki_1, y = roti_1)[name = tensor<string, []>("op_376")]; |
| tensor<fp32, [1, 4, ?, 32]> var_377 = sub(x = var_375, y = var_376)[name = tensor<string, []>("op_377")]; |
| tensor<fp32, [1, 4, ?, 32]> var_378 = mul(x = kr_1, y = roti_1)[name = tensor<string, []>("op_378")]; |
| tensor<fp32, [1, 4, ?, 32]> var_379 = mul(x = ki_1, y = rotr_1)[name = tensor<string, []>("op_379")]; |
| tensor<fp32, [1, 4, ?, 32]> var_380 = add(x = var_378, y = var_379)[name = tensor<string, []>("op_380")]; |
| tensor<int32, []> ko_1_axis_0 = const()[name = tensor<string, []>("ko_1_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_1 = stack(axis = ko_1_axis_0, values = (var_377, var_380))[name = tensor<string, []>("ko_1")]; |
| tensor<int32, [4]> concat_3x = const()[name = tensor<string, []>("concat_3x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_3 = reshape(shape = concat_3x, x = qo_1)[name = tensor<string, []>("q_3")]; |
| tensor<int32, [4]> concat_4x = const()[name = tensor<string, []>("concat_4x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_3 = reshape(shape = concat_4x, x = ko_1)[name = tensor<string, []>("k_3")]; |
| tensor<int32, []> const_5 = const()[name = tensor<string, []>("const_5"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_1 = range_1d(end = gather_4, start = const_5, step = const_6)[name = tensor<string, []>("positions_1")]; |
| tensor<int32, [3]> concat_5x = const()[name = tensor<string, []>("concat_5x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_392 = reshape(shape = concat_5x, x = positions_1)[name = tensor<string, []>("op_392")]; |
| tensor<int32, [3]> var_393 = const()[name = tensor<string, []>("op_393"), val = tensor<int32, [3]>([-1, 1, 1])]; |
| tensor<fp32, [1]> input_lengths_1 = cast(dtype = input_lengths_1_dtype_0, x = floor_div_0)[name = tensor<string, []>("cast_477")]; |
| tensor<fp32, [1, 1, 1]> var_394 = reshape(shape = var_393, x = input_lengths_1)[name = tensor<string, []>("op_394")]; |
| tensor<string, []> var_392_promoted_dtype_0 = const()[name = tensor<string, []>("op_392_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_392_promoted = cast(dtype = var_392_promoted_dtype_0, x = var_392)[name = tensor<string, []>("cast_475")]; |
| tensor<bool, [1, 1, ?]> valid_k_1 = less(x = var_392_promoted, y = var_394)[name = tensor<string, []>("valid_k_1")]; |
| tensor<int32, [3]> concat_6x = const()[name = tensor<string, []>("concat_6x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_397 = reshape(shape = concat_6x, x = positions_1)[name = tensor<string, []>("op_397")]; |
| tensor<int32, [1, ?, ?]> delta_1 = sub(x = var_397, y = var_392)[name = tensor<string, []>("delta_1")]; |
| tensor<int32, []> concat_8_axis_0 = const()[name = tensor<string, []>("concat_8_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_8_interleave_0 = const()[name = tensor<string, []>("concat_8_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_8 = concat(axis = concat_8_axis_0, interleave = concat_8_interleave_0, values = (var_295, gather_4, gather_4))[name = tensor<string, []>("concat_8")]; |
| tensor<fp32, []> fill_0_value_0 = const()[name = tensor<string, []>("fill_0_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_0 = fill(shape = concat_8, value = fill_0_value_0)[name = tensor<string, []>("fill_0")]; |
| tensor<string, []> attn_bias_1_dtype_0 = const()[name = tensor<string, []>("attn_bias_1_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_403 = greater_equal(x = delta_1, y = var_277)[name = tensor<string, []>("op_403")]; |
| tensor<bool, [1, ?, ?]> attn_bias_1 = cast(dtype = attn_bias_1_dtype_0, x = fill_0)[name = tensor<string, []>("cast_474")]; |
| tensor<bool, [1, ?, ?]> attn_bias_3 = logical_and(x = attn_bias_1, y = var_403)[name = tensor<string, []>("attn_bias_3")]; |
| tensor<bool, [1, ?, ?]> var_405 = less(x = delta_1, y = var_284)[name = tensor<string, []>("op_405")]; |
| tensor<bool, [1, ?, ?]> attn_bias_5 = logical_and(x = attn_bias_3, y = var_405)[name = tensor<string, []>("attn_bias_5")]; |
| tensor<bool, [1, ?, ?]> var_407 = logical_and(x = attn_bias_5, y = valid_k_1)[name = tensor<string, []>("op_407")]; |
| tensor<int32, [1]> var_409_axes_0 = const()[name = tensor<string, []>("op_409_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_409 = expand_dims(axes = var_409_axes_0, x = var_407)[name = tensor<string, []>("op_409")]; |
| tensor<string, []> cast_20_dtype_0 = const()[name = tensor<string, []>("cast_20_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_0_x_0 = const()[name = tensor<string, []>("sub_0_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_20 = cast(dtype = cast_20_dtype_0, x = var_409)[name = tensor<string, []>("cast_473")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_0 = sub(x = sub_0_x_0, y = cast_20)[name = tensor<string, []>("sub_0")]; |
| tensor<fp32, []> mul_0_x_0 = const()[name = tensor<string, []>("mul_0_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_0 = mul(x = mul_0_x_0, y = sub_0)[name = tensor<string, []>("mul_0")]; |
| tensor<fp32, []> mul_1_y_0 = const()[name = tensor<string, []>("mul_1_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_1 = mul(x = q_3, y = mul_1_y_0)[name = tensor<string, []>("mul_1")]; |
| tensor<bool, []> matmul_0_transpose_y_0 = const()[name = tensor<string, []>("matmul_0_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_0_transpose_x_0 = const()[name = tensor<string, []>("matmul_0_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_3)[name = tensor<string, []>("matmul_0")]; |
| tensor<fp32, [1, 4, ?, ?]> add_0 = add(x = matmul_0, y = mul_0)[name = tensor<string, []>("add_0")]; |
| tensor<int32, []> softmax_0_axis_0 = const()[name = tensor<string, []>("softmax_0_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_0 = softmax(axis = softmax_0_axis_0, x = add_0)[name = tensor<string, []>("softmax_0")]; |
| 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<fp32, [1, 4, ?, 64]> out_1 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = softmax_0, y = v_1)[name = tensor<string, []>("out_1")]; |
| tensor<int32, [2]> concat_9x = const()[name = tensor<string, []>("concat_9x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_415 = reshape(shape = concat_9x, x = positions_1)[name = tensor<string, []>("op_415")]; |
| tensor<int32, [2]> var_416 = const()[name = tensor<string, []>("op_416"), val = tensor<int32, [2]>([-1, 1])]; |
| tensor<fp32, [1, 1]> var_417 = reshape(shape = var_416, x = input_lengths_1)[name = tensor<string, []>("op_417")]; |
| tensor<string, []> var_415_promoted_dtype_0 = const()[name = tensor<string, []>("op_415_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_415_promoted = cast(dtype = var_415_promoted_dtype_0, x = var_415)[name = tensor<string, []>("cast_472")]; |
| tensor<bool, [1, ?]> var_418 = less(x = var_415_promoted, y = var_417)[name = tensor<string, []>("op_418")]; |
| tensor<int32, [4]> concat_10x = const()[name = tensor<string, []>("concat_10x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_1 = reshape(shape = concat_10x, x = var_418)[name = tensor<string, []>("valid_q_1")]; |
| tensor<fp32, []> var_422 = const()[name = tensor<string, []>("op_422"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_3 = select(a = out_1, b = var_422, cond = valid_q_1)[name = tensor<string, []>("out_3")]; |
| tensor<int32, [4]> var_424_perm_0 = const()[name = tensor<string, []>("op_424_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_11x = const()[name = tensor<string, []>("concat_11x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_424 = transpose(perm = var_424_perm_0, x = out_3)[name = tensor<string, []>("transpose_80")]; |
| tensor<fp32, [1, ?, 256]> input_3 = reshape(shape = concat_11x, x = var_424)[name = tensor<string, []>("input_3")]; |
| tensor<fp32, [1, ?, 256]> x_9 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_0_self_attn_out_proj_weight, x = input_3)[name = tensor<string, []>("linear_2")]; |
| tensor<fp32, [1, ?, 256]> var_430 = mul(x = encoder_1_transformer_layers_0_layer_scale_1_scale, y = x_9)[name = tensor<string, []>("op_430")]; |
| tensor<fp32, [1, ?, 256]> input_5 = add(x = x_5, y = var_430)[name = tensor<string, []>("input_5")]; |
| tensor<int32, [1]> input_7_axes_0 = const()[name = tensor<string, []>("input_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_7 = layer_norm(axes = input_7_axes_0, beta = encoder_1_transformer_layers_0_norm2_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_0_norm2_weight, x = input_5)[name = tensor<string, []>("input_7")]; |
| tensor<fp32, [1024]> linear_3_bias_0 = const()[name = tensor<string, []>("linear_3_bias_0"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44425536)))]; |
| tensor<fp32, [1, ?, 1024]> input_9 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_0_ffn_0_weight, x = input_7)[name = tensor<string, []>("linear_3")]; |
| tensor<string, []> input_11_mode_0 = const()[name = tensor<string, []>("input_11_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_11 = gelu(mode = input_11_mode_0, x = input_9)[name = tensor<string, []>("input_11")]; |
| tensor<fp32, [1, ?, 256]> x_11 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_0_ffn_2_weight, x = input_11)[name = tensor<string, []>("linear_4")]; |
| tensor<fp32, [1, ?, 256]> var_445 = mul(x = encoder_1_transformer_layers_0_layer_scale_2_scale, y = x_11)[name = tensor<string, []>("op_445")]; |
| tensor<fp32, [1, ?, 256]> input_13 = add(x = input_5, y = var_445)[name = tensor<string, []>("input_13")]; |
| tensor<int32, [1]> x_13_axes_0 = const()[name = tensor<string, []>("x_13_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_1_transformer_layers_1_norm1_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_1_norm1_weight, x = input_13)[name = tensor<string, []>("x_13")]; |
| tensor<int32, [3]> var_460_shape = shape(x = x_13)[name = tensor<string, []>("op_460_shape")]; |
| tensor<int32, []> gather_14_batch_dims_0 = const()[name = tensor<string, []>("gather_14_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_14_validate_indices_0 = const()[name = tensor<string, []>("gather_14_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_18 = const()[name = tensor<string, []>("select_18"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_14_axis_1 = const()[name = tensor<string, []>("gather_14_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_14 = gather(axis = gather_14_axis_1, batch_dims = gather_14_batch_dims_0, indices = select_18, validate_indices = gather_14_validate_indices_0, x = var_460_shape)[name = tensor<string, []>("gather_14")]; |
| tensor<fp32, [1, ?, 768]> projected_5 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_1_self_attn_in_proj_weight, x = x_13)[name = tensor<string, []>("linear_5")]; |
| tensor<int32, [5]> concat_12x = const()[name = tensor<string, []>("concat_12x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_467 = reshape(shape = concat_12x, x = projected_5)[name = tensor<string, []>("op_467")]; |
| tensor<int32, [5]> var_468 = const()[name = tensor<string, []>("op_468"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_5_begin_0 = const()[name = tensor<string, []>("q_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_5_end_0 = const()[name = tensor<string, []>("q_5_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_5_end_mask_0 = const()[name = tensor<string, []>("q_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_5_squeeze_mask_0 = const()[name = tensor<string, []>("q_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_7 = transpose(perm = var_468, x = var_467)[name = tensor<string, []>("transpose_79")]; |
| tensor<fp32, [1, 4, ?, 64]> q_5 = slice_by_index(begin = q_5_begin_0, end = q_5_end_0, end_mask = q_5_end_mask_0, squeeze_mask = q_5_squeeze_mask_0, x = projected_7)[name = tensor<string, []>("q_5")]; |
| tensor<int32, [5]> k_5_begin_0 = const()[name = tensor<string, []>("k_5_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_5_end_0 = const()[name = tensor<string, []>("k_5_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_5_end_mask_0 = const()[name = tensor<string, []>("k_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_5_squeeze_mask_0 = const()[name = tensor<string, []>("k_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_5 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = projected_7)[name = tensor<string, []>("k_5")]; |
| tensor<int32, [5]> v_3_begin_0 = const()[name = tensor<string, []>("v_3_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_3_end_0 = const()[name = tensor<string, []>("v_3_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_3_end_mask_0 = const()[name = tensor<string, []>("v_3_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_3_squeeze_mask_0 = const()[name = tensor<string, []>("v_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_3 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = projected_7)[name = tensor<string, []>("v_3")]; |
| tensor<fp32, [32]> freqs_3 = const()[name = tensor<string, []>("freqs_3"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44429696)))]; |
| tensor<int32, [4]> var_480_shape = shape(x = q_5)[name = tensor<string, []>("op_480_shape")]; |
| tensor<int32, []> gather_18_batch_dims_0 = const()[name = tensor<string, []>("gather_18_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_18_validate_indices_0 = const()[name = tensor<string, []>("gather_18_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_19 = const()[name = tensor<string, []>("select_19"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_18_axis_1 = const()[name = tensor<string, []>("gather_18_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_18 = gather(axis = gather_18_axis_1, batch_dims = gather_18_batch_dims_0, indices = select_19, validate_indices = gather_18_validate_indices_0, x = var_480_shape)[name = tensor<string, []>("gather_18")]; |
| tensor<fp32, [1, 1]> var_483 = const()[name = tensor<string, []>("op_483"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_12 = const()[name = tensor<string, []>("const_12"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_13 = const()[name = tensor<string, []>("const_13"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_484 = range_1d(end = gather_18, start = const_12, step = const_13)[name = tensor<string, []>("op_484")]; |
| tensor<string, []> var_484_promoted_dtype_0 = const()[name = tensor<string, []>("op_484_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_484_promoted = cast(dtype = var_484_promoted_dtype_0, x = var_484)[name = tensor<string, []>("cast_471")]; |
| tensor<fp32, [1, ?]> ts_5 = add(x = var_483, y = var_484_promoted)[name = tensor<string, []>("ts_5")]; |
| tensor<int32, [4]> var_487 = const()[name = tensor<string, []>("op_487"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_7 = reshape(shape = var_487, x = ts_5)[name = tensor<string, []>("ts_7")]; |
| tensor<int32, [5]> concat_13x = const()[name = tensor<string, []>("concat_13x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_3 = reshape(shape = concat_13x, x = q_5)[name = tensor<string, []>("q2_3")]; |
| tensor<int32, [5]> concat_14x = const()[name = tensor<string, []>("concat_14x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_3 = reshape(shape = concat_14x, x = k_5)[name = tensor<string, []>("k2_3")]; |
| tensor<int32, [5]> qr_3_begin_0 = const()[name = tensor<string, []>("qr_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_3_end_0 = const()[name = tensor<string, []>("qr_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_3_end_mask_0 = const()[name = tensor<string, []>("qr_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_3_squeeze_mask_0 = const()[name = tensor<string, []>("qr_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_3 = slice_by_index(begin = qr_3_begin_0, end = qr_3_end_0, end_mask = qr_3_end_mask_0, squeeze_mask = qr_3_squeeze_mask_0, x = q2_3)[name = tensor<string, []>("qr_3")]; |
| tensor<int32, [5]> qi_3_begin_0 = const()[name = tensor<string, []>("qi_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_3_end_0 = const()[name = tensor<string, []>("qi_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_3_end_mask_0 = const()[name = tensor<string, []>("qi_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_3_squeeze_mask_0 = const()[name = tensor<string, []>("qi_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_3 = slice_by_index(begin = qi_3_begin_0, end = qi_3_end_0, end_mask = qi_3_end_mask_0, squeeze_mask = qi_3_squeeze_mask_0, x = q2_3)[name = tensor<string, []>("qi_3")]; |
| tensor<int32, [5]> kr_3_begin_0 = const()[name = tensor<string, []>("kr_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_3_end_0 = const()[name = tensor<string, []>("kr_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_3_end_mask_0 = const()[name = tensor<string, []>("kr_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_3_squeeze_mask_0 = const()[name = tensor<string, []>("kr_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_3 = slice_by_index(begin = kr_3_begin_0, end = kr_3_end_0, end_mask = kr_3_end_mask_0, squeeze_mask = kr_3_squeeze_mask_0, x = k2_3)[name = tensor<string, []>("kr_3")]; |
| tensor<int32, [5]> ki_3_begin_0 = const()[name = tensor<string, []>("ki_3_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_3_end_0 = const()[name = tensor<string, []>("ki_3_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_3_end_mask_0 = const()[name = tensor<string, []>("ki_3_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_3_squeeze_mask_0 = const()[name = tensor<string, []>("ki_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_3 = slice_by_index(begin = ki_3_begin_0, end = ki_3_end_0, end_mask = ki_3_end_mask_0, squeeze_mask = ki_3_squeeze_mask_0, x = k2_3)[name = tensor<string, []>("ki_3")]; |
| tensor<fp32, [1, 1, ?, 32]> var_500 = mul(x = freqs_3, y = ts_7)[name = tensor<string, []>("op_500")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_3 = cos(x = var_500)[name = tensor<string, []>("rotr_3")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_3 = sin(x = var_500)[name = tensor<string, []>("roti_3")]; |
| tensor<fp32, [1, 4, ?, 32]> var_504 = mul(x = qr_3, y = rotr_3)[name = tensor<string, []>("op_504")]; |
| tensor<fp32, [1, 4, ?, 32]> var_505 = mul(x = qi_3, y = roti_3)[name = tensor<string, []>("op_505")]; |
| tensor<fp32, [1, 4, ?, 32]> var_506 = sub(x = var_504, y = var_505)[name = tensor<string, []>("op_506")]; |
| tensor<fp32, [1, 4, ?, 32]> var_507 = mul(x = qr_3, y = roti_3)[name = tensor<string, []>("op_507")]; |
| tensor<fp32, [1, 4, ?, 32]> var_508 = mul(x = qi_3, y = rotr_3)[name = tensor<string, []>("op_508")]; |
| tensor<fp32, [1, 4, ?, 32]> var_509 = add(x = var_507, y = var_508)[name = tensor<string, []>("op_509")]; |
| tensor<int32, []> qo_3_axis_0 = const()[name = tensor<string, []>("qo_3_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_3 = stack(axis = qo_3_axis_0, values = (var_506, var_509))[name = tensor<string, []>("qo_3")]; |
| tensor<fp32, [1, 4, ?, 32]> var_512 = mul(x = kr_3, y = rotr_3)[name = tensor<string, []>("op_512")]; |
| tensor<fp32, [1, 4, ?, 32]> var_513 = mul(x = ki_3, y = roti_3)[name = tensor<string, []>("op_513")]; |
| tensor<fp32, [1, 4, ?, 32]> var_514 = sub(x = var_512, y = var_513)[name = tensor<string, []>("op_514")]; |
| tensor<fp32, [1, 4, ?, 32]> var_515 = mul(x = kr_3, y = roti_3)[name = tensor<string, []>("op_515")]; |
| tensor<fp32, [1, 4, ?, 32]> var_516 = mul(x = ki_3, y = rotr_3)[name = tensor<string, []>("op_516")]; |
| tensor<fp32, [1, 4, ?, 32]> var_517 = add(x = var_515, y = var_516)[name = tensor<string, []>("op_517")]; |
| tensor<int32, []> ko_3_axis_0 = const()[name = tensor<string, []>("ko_3_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_3 = stack(axis = ko_3_axis_0, values = (var_514, var_517))[name = tensor<string, []>("ko_3")]; |
| tensor<int32, [4]> concat_15x = const()[name = tensor<string, []>("concat_15x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_7 = reshape(shape = concat_15x, x = qo_3)[name = tensor<string, []>("q_7")]; |
| tensor<int32, [4]> concat_16x = const()[name = tensor<string, []>("concat_16x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_7 = reshape(shape = concat_16x, x = ko_3)[name = tensor<string, []>("k_7")]; |
| tensor<int32, []> const_14 = const()[name = tensor<string, []>("const_14"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_15 = const()[name = tensor<string, []>("const_15"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_3 = range_1d(end = gather_14, start = const_14, step = const_15)[name = tensor<string, []>("positions_3")]; |
| tensor<int32, [3]> concat_17x = const()[name = tensor<string, []>("concat_17x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_529 = reshape(shape = concat_17x, x = positions_3)[name = tensor<string, []>("op_529")]; |
| tensor<string, []> var_529_promoted_dtype_0 = const()[name = tensor<string, []>("op_529_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_529_promoted = cast(dtype = var_529_promoted_dtype_0, x = var_529)[name = tensor<string, []>("cast_470")]; |
| tensor<bool, [1, 1, ?]> valid_k_3 = less(x = var_529_promoted, y = var_394)[name = tensor<string, []>("valid_k_3")]; |
| tensor<int32, [3]> concat_18x = const()[name = tensor<string, []>("concat_18x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_534 = reshape(shape = concat_18x, x = positions_3)[name = tensor<string, []>("op_534")]; |
| tensor<int32, [1, ?, ?]> delta_3 = sub(x = var_534, y = var_529)[name = tensor<string, []>("delta_3")]; |
| tensor<int32, []> concat_20_axis_0 = const()[name = tensor<string, []>("concat_20_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_20_interleave_0 = const()[name = tensor<string, []>("concat_20_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_20 = concat(axis = concat_20_axis_0, interleave = concat_20_interleave_0, values = (var_295, gather_14, gather_14))[name = tensor<string, []>("concat_20")]; |
| tensor<fp32, []> fill_1_value_0 = const()[name = tensor<string, []>("fill_1_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_1 = fill(shape = concat_20, value = fill_1_value_0)[name = tensor<string, []>("fill_1")]; |
| tensor<string, []> attn_bias_9_dtype_0 = const()[name = tensor<string, []>("attn_bias_9_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_540 = greater_equal(x = delta_3, y = var_277)[name = tensor<string, []>("op_540")]; |
| tensor<bool, [1, ?, ?]> attn_bias_9 = cast(dtype = attn_bias_9_dtype_0, x = fill_1)[name = tensor<string, []>("cast_469")]; |
| tensor<bool, [1, ?, ?]> attn_bias_11 = logical_and(x = attn_bias_9, y = var_540)[name = tensor<string, []>("attn_bias_11")]; |
| tensor<bool, [1, ?, ?]> var_542 = less(x = delta_3, y = var_284)[name = tensor<string, []>("op_542")]; |
| tensor<bool, [1, ?, ?]> attn_bias_13 = logical_and(x = attn_bias_11, y = var_542)[name = tensor<string, []>("attn_bias_13")]; |
| tensor<bool, [1, ?, ?]> var_544 = logical_and(x = attn_bias_13, y = valid_k_3)[name = tensor<string, []>("op_544")]; |
| tensor<int32, [1]> var_546_axes_0 = const()[name = tensor<string, []>("op_546_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_546 = expand_dims(axes = var_546_axes_0, x = var_544)[name = tensor<string, []>("op_546")]; |
| tensor<string, []> cast_42_dtype_0 = const()[name = tensor<string, []>("cast_42_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_1_x_0 = const()[name = tensor<string, []>("sub_1_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_42 = cast(dtype = cast_42_dtype_0, x = var_546)[name = tensor<string, []>("cast_468")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_1 = sub(x = sub_1_x_0, y = cast_42)[name = tensor<string, []>("sub_1")]; |
| tensor<fp32, []> mul_2_x_0 = const()[name = tensor<string, []>("mul_2_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_2 = mul(x = mul_2_x_0, y = sub_1)[name = tensor<string, []>("mul_2")]; |
| tensor<fp32, []> mul_3_y_0 = const()[name = tensor<string, []>("mul_3_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_3 = mul(x = q_7, y = mul_3_y_0)[name = tensor<string, []>("mul_3")]; |
| tensor<bool, []> matmul_1_transpose_y_0 = const()[name = tensor<string, []>("matmul_1_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_1_transpose_x_0 = const()[name = tensor<string, []>("matmul_1_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k_7)[name = tensor<string, []>("matmul_1")]; |
| tensor<fp32, [1, 4, ?, ?]> add_1 = add(x = matmul_1, y = mul_2)[name = tensor<string, []>("add_1")]; |
| tensor<int32, []> softmax_1_axis_0 = const()[name = tensor<string, []>("softmax_1_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_1 = softmax(axis = softmax_1_axis_0, x = add_1)[name = tensor<string, []>("softmax_1")]; |
| 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<fp32, [1, 4, ?, 64]> out_5 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = softmax_1, y = v_3)[name = tensor<string, []>("out_5")]; |
| tensor<int32, [2]> concat_21x = const()[name = tensor<string, []>("concat_21x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_552 = reshape(shape = concat_21x, x = positions_3)[name = tensor<string, []>("op_552")]; |
| tensor<string, []> var_552_promoted_dtype_0 = const()[name = tensor<string, []>("op_552_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_552_promoted = cast(dtype = var_552_promoted_dtype_0, x = var_552)[name = tensor<string, []>("cast_467")]; |
| tensor<bool, [1, ?]> var_555 = less(x = var_552_promoted, y = var_417)[name = tensor<string, []>("op_555")]; |
| tensor<int32, [4]> concat_22x = const()[name = tensor<string, []>("concat_22x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_3 = reshape(shape = concat_22x, x = var_555)[name = tensor<string, []>("valid_q_3")]; |
| tensor<fp32, []> var_559 = const()[name = tensor<string, []>("op_559"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_7 = select(a = out_5, b = var_559, cond = valid_q_3)[name = tensor<string, []>("out_7")]; |
| tensor<int32, [4]> var_561_perm_0 = const()[name = tensor<string, []>("op_561_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_23x = const()[name = tensor<string, []>("concat_23x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_561 = transpose(perm = var_561_perm_0, x = out_7)[name = tensor<string, []>("transpose_78")]; |
| tensor<fp32, [1, ?, 256]> input_15 = reshape(shape = concat_23x, x = var_561)[name = tensor<string, []>("input_15")]; |
| tensor<fp32, [1, ?, 256]> x_15 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor<string, []>("linear_6")]; |
| tensor<fp32, [1, ?, 256]> var_567 = mul(x = encoder_1_transformer_layers_1_layer_scale_1_scale, y = x_15)[name = tensor<string, []>("op_567")]; |
| tensor<fp32, [1, ?, 256]> input_17 = add(x = input_13, y = var_567)[name = tensor<string, []>("input_17")]; |
| tensor<int32, [1]> input_19_axes_0 = const()[name = tensor<string, []>("input_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_19 = layer_norm(axes = input_19_axes_0, beta = encoder_1_transformer_layers_1_norm2_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_1_norm2_weight, x = input_17)[name = tensor<string, []>("input_19")]; |
| tensor<fp32, [1, ?, 1024]> input_21 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_1_ffn_0_weight, x = input_19)[name = tensor<string, []>("linear_7")]; |
| tensor<string, []> input_23_mode_0 = const()[name = tensor<string, []>("input_23_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_23 = gelu(mode = input_23_mode_0, x = input_21)[name = tensor<string, []>("input_23")]; |
| tensor<fp32, [1, ?, 256]> x_17 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_1_ffn_2_weight, x = input_23)[name = tensor<string, []>("linear_8")]; |
| tensor<fp32, [1, ?, 256]> var_582 = mul(x = encoder_1_transformer_layers_1_layer_scale_2_scale, y = x_17)[name = tensor<string, []>("op_582")]; |
| tensor<fp32, [1, ?, 256]> input_25 = add(x = input_17, y = var_582)[name = tensor<string, []>("input_25")]; |
| tensor<int32, [1]> x_19_axes_0 = const()[name = tensor<string, []>("x_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_1_transformer_layers_2_norm1_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_2_norm1_weight, x = input_25)[name = tensor<string, []>("x_19")]; |
| tensor<int32, [3]> var_597_shape = shape(x = x_19)[name = tensor<string, []>("op_597_shape")]; |
| tensor<int32, []> gather_24_batch_dims_0 = const()[name = tensor<string, []>("gather_24_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_24_validate_indices_0 = const()[name = tensor<string, []>("gather_24_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_20 = const()[name = tensor<string, []>("select_20"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_24_axis_1 = const()[name = tensor<string, []>("gather_24_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_24 = gather(axis = gather_24_axis_1, batch_dims = gather_24_batch_dims_0, indices = select_20, validate_indices = gather_24_validate_indices_0, x = var_597_shape)[name = tensor<string, []>("gather_24")]; |
| tensor<fp32, [1, ?, 768]> projected_9 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_2_self_attn_in_proj_weight, x = x_19)[name = tensor<string, []>("linear_9")]; |
| tensor<int32, [5]> concat_24x = const()[name = tensor<string, []>("concat_24x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_604 = reshape(shape = concat_24x, x = projected_9)[name = tensor<string, []>("op_604")]; |
| tensor<int32, [5]> var_605 = const()[name = tensor<string, []>("op_605"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_9_begin_0 = const()[name = tensor<string, []>("q_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_9_end_0 = const()[name = tensor<string, []>("q_9_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_9_end_mask_0 = const()[name = tensor<string, []>("q_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_9_squeeze_mask_0 = const()[name = tensor<string, []>("q_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_11 = transpose(perm = var_605, x = var_604)[name = tensor<string, []>("transpose_77")]; |
| tensor<fp32, [1, 4, ?, 64]> q_9 = slice_by_index(begin = q_9_begin_0, end = q_9_end_0, end_mask = q_9_end_mask_0, squeeze_mask = q_9_squeeze_mask_0, x = projected_11)[name = tensor<string, []>("q_9")]; |
| tensor<int32, [5]> k_9_begin_0 = const()[name = tensor<string, []>("k_9_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_9_end_0 = const()[name = tensor<string, []>("k_9_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_9_end_mask_0 = const()[name = tensor<string, []>("k_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_9_squeeze_mask_0 = const()[name = tensor<string, []>("k_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_9 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = projected_11)[name = tensor<string, []>("k_9")]; |
| tensor<int32, [5]> v_5_begin_0 = const()[name = tensor<string, []>("v_5_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_5_end_0 = const()[name = tensor<string, []>("v_5_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_5_end_mask_0 = const()[name = tensor<string, []>("v_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_5_squeeze_mask_0 = const()[name = tensor<string, []>("v_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_5 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = projected_11)[name = tensor<string, []>("v_5")]; |
| tensor<fp32, [32]> freqs_5 = const()[name = tensor<string, []>("freqs_5"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44429888)))]; |
| tensor<int32, [4]> var_617_shape = shape(x = q_9)[name = tensor<string, []>("op_617_shape")]; |
| tensor<int32, []> gather_28_batch_dims_0 = const()[name = tensor<string, []>("gather_28_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_28_validate_indices_0 = const()[name = tensor<string, []>("gather_28_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_21 = const()[name = tensor<string, []>("select_21"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_28_axis_1 = const()[name = tensor<string, []>("gather_28_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_28 = gather(axis = gather_28_axis_1, batch_dims = gather_28_batch_dims_0, indices = select_21, validate_indices = gather_28_validate_indices_0, x = var_617_shape)[name = tensor<string, []>("gather_28")]; |
| tensor<fp32, [1, 1]> var_620 = const()[name = tensor<string, []>("op_620"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_21 = const()[name = tensor<string, []>("const_21"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_22 = const()[name = tensor<string, []>("const_22"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_621 = range_1d(end = gather_28, start = const_21, step = const_22)[name = tensor<string, []>("op_621")]; |
| tensor<string, []> var_621_promoted_dtype_0 = const()[name = tensor<string, []>("op_621_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_621_promoted = cast(dtype = var_621_promoted_dtype_0, x = var_621)[name = tensor<string, []>("cast_466")]; |
| tensor<fp32, [1, ?]> ts_9 = add(x = var_620, y = var_621_promoted)[name = tensor<string, []>("ts_9")]; |
| tensor<int32, [4]> var_624 = const()[name = tensor<string, []>("op_624"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_11 = reshape(shape = var_624, x = ts_9)[name = tensor<string, []>("ts_11")]; |
| tensor<int32, [5]> concat_25x = const()[name = tensor<string, []>("concat_25x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_5 = reshape(shape = concat_25x, x = q_9)[name = tensor<string, []>("q2_5")]; |
| tensor<int32, [5]> concat_26x = const()[name = tensor<string, []>("concat_26x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_5 = reshape(shape = concat_26x, x = k_9)[name = tensor<string, []>("k2_5")]; |
| tensor<int32, [5]> qr_5_begin_0 = const()[name = tensor<string, []>("qr_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_5_end_0 = const()[name = tensor<string, []>("qr_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_5_end_mask_0 = const()[name = tensor<string, []>("qr_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_5_squeeze_mask_0 = const()[name = tensor<string, []>("qr_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_5 = slice_by_index(begin = qr_5_begin_0, end = qr_5_end_0, end_mask = qr_5_end_mask_0, squeeze_mask = qr_5_squeeze_mask_0, x = q2_5)[name = tensor<string, []>("qr_5")]; |
| tensor<int32, [5]> qi_5_begin_0 = const()[name = tensor<string, []>("qi_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_5_end_0 = const()[name = tensor<string, []>("qi_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_5_end_mask_0 = const()[name = tensor<string, []>("qi_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_5_squeeze_mask_0 = const()[name = tensor<string, []>("qi_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_5 = slice_by_index(begin = qi_5_begin_0, end = qi_5_end_0, end_mask = qi_5_end_mask_0, squeeze_mask = qi_5_squeeze_mask_0, x = q2_5)[name = tensor<string, []>("qi_5")]; |
| tensor<int32, [5]> kr_5_begin_0 = const()[name = tensor<string, []>("kr_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_5_end_0 = const()[name = tensor<string, []>("kr_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_5_end_mask_0 = const()[name = tensor<string, []>("kr_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_5_squeeze_mask_0 = const()[name = tensor<string, []>("kr_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_5 = slice_by_index(begin = kr_5_begin_0, end = kr_5_end_0, end_mask = kr_5_end_mask_0, squeeze_mask = kr_5_squeeze_mask_0, x = k2_5)[name = tensor<string, []>("kr_5")]; |
| tensor<int32, [5]> ki_5_begin_0 = const()[name = tensor<string, []>("ki_5_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_5_end_0 = const()[name = tensor<string, []>("ki_5_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_5_end_mask_0 = const()[name = tensor<string, []>("ki_5_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_5_squeeze_mask_0 = const()[name = tensor<string, []>("ki_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_5 = slice_by_index(begin = ki_5_begin_0, end = ki_5_end_0, end_mask = ki_5_end_mask_0, squeeze_mask = ki_5_squeeze_mask_0, x = k2_5)[name = tensor<string, []>("ki_5")]; |
| tensor<fp32, [1, 1, ?, 32]> var_637 = mul(x = freqs_5, y = ts_11)[name = tensor<string, []>("op_637")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_5 = cos(x = var_637)[name = tensor<string, []>("rotr_5")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_5 = sin(x = var_637)[name = tensor<string, []>("roti_5")]; |
| tensor<fp32, [1, 4, ?, 32]> var_641 = mul(x = qr_5, y = rotr_5)[name = tensor<string, []>("op_641")]; |
| tensor<fp32, [1, 4, ?, 32]> var_642 = mul(x = qi_5, y = roti_5)[name = tensor<string, []>("op_642")]; |
| tensor<fp32, [1, 4, ?, 32]> var_643 = sub(x = var_641, y = var_642)[name = tensor<string, []>("op_643")]; |
| tensor<fp32, [1, 4, ?, 32]> var_644 = mul(x = qr_5, y = roti_5)[name = tensor<string, []>("op_644")]; |
| tensor<fp32, [1, 4, ?, 32]> var_645 = mul(x = qi_5, y = rotr_5)[name = tensor<string, []>("op_645")]; |
| tensor<fp32, [1, 4, ?, 32]> var_646 = add(x = var_644, y = var_645)[name = tensor<string, []>("op_646")]; |
| tensor<int32, []> qo_5_axis_0 = const()[name = tensor<string, []>("qo_5_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_5 = stack(axis = qo_5_axis_0, values = (var_643, var_646))[name = tensor<string, []>("qo_5")]; |
| tensor<fp32, [1, 4, ?, 32]> var_649 = mul(x = kr_5, y = rotr_5)[name = tensor<string, []>("op_649")]; |
| tensor<fp32, [1, 4, ?, 32]> var_650 = mul(x = ki_5, y = roti_5)[name = tensor<string, []>("op_650")]; |
| tensor<fp32, [1, 4, ?, 32]> var_651 = sub(x = var_649, y = var_650)[name = tensor<string, []>("op_651")]; |
| tensor<fp32, [1, 4, ?, 32]> var_652 = mul(x = kr_5, y = roti_5)[name = tensor<string, []>("op_652")]; |
| tensor<fp32, [1, 4, ?, 32]> var_653 = mul(x = ki_5, y = rotr_5)[name = tensor<string, []>("op_653")]; |
| tensor<fp32, [1, 4, ?, 32]> var_654 = add(x = var_652, y = var_653)[name = tensor<string, []>("op_654")]; |
| tensor<int32, []> ko_5_axis_0 = const()[name = tensor<string, []>("ko_5_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_5 = stack(axis = ko_5_axis_0, values = (var_651, var_654))[name = tensor<string, []>("ko_5")]; |
| tensor<int32, [4]> concat_27x = const()[name = tensor<string, []>("concat_27x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_11 = reshape(shape = concat_27x, x = qo_5)[name = tensor<string, []>("q_11")]; |
| tensor<int32, [4]> concat_28x = const()[name = tensor<string, []>("concat_28x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_11 = reshape(shape = concat_28x, x = ko_5)[name = tensor<string, []>("k_11")]; |
| tensor<int32, []> const_23 = const()[name = tensor<string, []>("const_23"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_24 = const()[name = tensor<string, []>("const_24"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_5 = range_1d(end = gather_24, start = const_23, step = const_24)[name = tensor<string, []>("positions_5")]; |
| tensor<int32, [3]> concat_29x = const()[name = tensor<string, []>("concat_29x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_666 = reshape(shape = concat_29x, x = positions_5)[name = tensor<string, []>("op_666")]; |
| tensor<string, []> var_666_promoted_dtype_0 = const()[name = tensor<string, []>("op_666_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_666_promoted = cast(dtype = var_666_promoted_dtype_0, x = var_666)[name = tensor<string, []>("cast_465")]; |
| tensor<bool, [1, 1, ?]> valid_k_5 = less(x = var_666_promoted, y = var_394)[name = tensor<string, []>("valid_k_5")]; |
| tensor<int32, [3]> concat_30x = const()[name = tensor<string, []>("concat_30x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_671 = reshape(shape = concat_30x, x = positions_5)[name = tensor<string, []>("op_671")]; |
| tensor<int32, [1, ?, ?]> delta_5 = sub(x = var_671, y = var_666)[name = tensor<string, []>("delta_5")]; |
| tensor<int32, []> concat_32_axis_0 = const()[name = tensor<string, []>("concat_32_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_32_interleave_0 = const()[name = tensor<string, []>("concat_32_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_32 = concat(axis = concat_32_axis_0, interleave = concat_32_interleave_0, values = (var_295, gather_24, gather_24))[name = tensor<string, []>("concat_32")]; |
| tensor<fp32, []> fill_2_value_0 = const()[name = tensor<string, []>("fill_2_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_2 = fill(shape = concat_32, value = fill_2_value_0)[name = tensor<string, []>("fill_2")]; |
| tensor<string, []> attn_bias_17_dtype_0 = const()[name = tensor<string, []>("attn_bias_17_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_677 = greater_equal(x = delta_5, y = var_277)[name = tensor<string, []>("op_677")]; |
| tensor<bool, [1, ?, ?]> attn_bias_17 = cast(dtype = attn_bias_17_dtype_0, x = fill_2)[name = tensor<string, []>("cast_464")]; |
| tensor<bool, [1, ?, ?]> attn_bias_19 = logical_and(x = attn_bias_17, y = var_677)[name = tensor<string, []>("attn_bias_19")]; |
| tensor<bool, [1, ?, ?]> var_679 = less(x = delta_5, y = var_284)[name = tensor<string, []>("op_679")]; |
| tensor<bool, [1, ?, ?]> attn_bias_21 = logical_and(x = attn_bias_19, y = var_679)[name = tensor<string, []>("attn_bias_21")]; |
| tensor<bool, [1, ?, ?]> var_681 = logical_and(x = attn_bias_21, y = valid_k_5)[name = tensor<string, []>("op_681")]; |
| tensor<int32, [1]> var_683_axes_0 = const()[name = tensor<string, []>("op_683_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_683 = expand_dims(axes = var_683_axes_0, x = var_681)[name = tensor<string, []>("op_683")]; |
| tensor<string, []> cast_64_dtype_0 = const()[name = tensor<string, []>("cast_64_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_2_x_0 = const()[name = tensor<string, []>("sub_2_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_64 = cast(dtype = cast_64_dtype_0, x = var_683)[name = tensor<string, []>("cast_463")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_2 = sub(x = sub_2_x_0, y = cast_64)[name = tensor<string, []>("sub_2")]; |
| tensor<fp32, []> mul_4_x_0 = const()[name = tensor<string, []>("mul_4_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_4 = mul(x = mul_4_x_0, y = sub_2)[name = tensor<string, []>("mul_4")]; |
| tensor<fp32, []> mul_5_y_0 = const()[name = tensor<string, []>("mul_5_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_5 = mul(x = q_11, y = mul_5_y_0)[name = tensor<string, []>("mul_5")]; |
| tensor<bool, []> matmul_2_transpose_y_0 = const()[name = tensor<string, []>("matmul_2_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_2_transpose_x_0 = const()[name = tensor<string, []>("matmul_2_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = mul_5, y = k_11)[name = tensor<string, []>("matmul_2")]; |
| tensor<fp32, [1, 4, ?, ?]> add_2 = add(x = matmul_2, y = mul_4)[name = tensor<string, []>("add_2")]; |
| tensor<int32, []> softmax_2_axis_0 = const()[name = tensor<string, []>("softmax_2_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_2 = softmax(axis = softmax_2_axis_0, x = add_2)[name = tensor<string, []>("softmax_2")]; |
| 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<fp32, [1, 4, ?, 64]> out_9 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = softmax_2, y = v_5)[name = tensor<string, []>("out_9")]; |
| tensor<int32, [2]> concat_33x = const()[name = tensor<string, []>("concat_33x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_689 = reshape(shape = concat_33x, x = positions_5)[name = tensor<string, []>("op_689")]; |
| tensor<string, []> var_689_promoted_dtype_0 = const()[name = tensor<string, []>("op_689_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_689_promoted = cast(dtype = var_689_promoted_dtype_0, x = var_689)[name = tensor<string, []>("cast_462")]; |
| tensor<bool, [1, ?]> var_692 = less(x = var_689_promoted, y = var_417)[name = tensor<string, []>("op_692")]; |
| tensor<int32, [4]> concat_34x = const()[name = tensor<string, []>("concat_34x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_5 = reshape(shape = concat_34x, x = var_692)[name = tensor<string, []>("valid_q_5")]; |
| tensor<fp32, []> var_696 = const()[name = tensor<string, []>("op_696"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_11 = select(a = out_9, b = var_696, cond = valid_q_5)[name = tensor<string, []>("out_11")]; |
| tensor<int32, [4]> var_698_perm_0 = const()[name = tensor<string, []>("op_698_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_35x = const()[name = tensor<string, []>("concat_35x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_698 = transpose(perm = var_698_perm_0, x = out_11)[name = tensor<string, []>("transpose_76")]; |
| tensor<fp32, [1, ?, 256]> input_27 = reshape(shape = concat_35x, x = var_698)[name = tensor<string, []>("input_27")]; |
| tensor<fp32, [1, ?, 256]> x_21 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_2_self_attn_out_proj_weight, x = input_27)[name = tensor<string, []>("linear_10")]; |
| tensor<fp32, [1, ?, 256]> var_704 = mul(x = encoder_1_transformer_layers_2_layer_scale_1_scale, y = x_21)[name = tensor<string, []>("op_704")]; |
| tensor<fp32, [1, ?, 256]> input_29 = add(x = input_25, y = var_704)[name = tensor<string, []>("input_29")]; |
| tensor<int32, [1]> input_31_axes_0 = const()[name = tensor<string, []>("input_31_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_31 = layer_norm(axes = input_31_axes_0, beta = encoder_1_transformer_layers_2_norm2_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_2_norm2_weight, x = input_29)[name = tensor<string, []>("input_31")]; |
| tensor<fp32, [1, ?, 1024]> input_33 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_2_ffn_0_weight, x = input_31)[name = tensor<string, []>("linear_11")]; |
| tensor<string, []> input_35_mode_0 = const()[name = tensor<string, []>("input_35_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_35 = gelu(mode = input_35_mode_0, x = input_33)[name = tensor<string, []>("input_35")]; |
| tensor<fp32, [1, ?, 256]> x_23 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_2_ffn_2_weight, x = input_35)[name = tensor<string, []>("linear_12")]; |
| tensor<fp32, [1, ?, 256]> var_719 = mul(x = encoder_1_transformer_layers_2_layer_scale_2_scale, y = x_23)[name = tensor<string, []>("op_719")]; |
| tensor<fp32, [1, ?, 256]> input_37 = add(x = input_29, y = var_719)[name = tensor<string, []>("input_37")]; |
| tensor<int32, [1]> x_25_axes_0 = const()[name = tensor<string, []>("x_25_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_1_transformer_layers_3_norm1_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_3_norm1_weight, x = input_37)[name = tensor<string, []>("x_25")]; |
| tensor<int32, [3]> var_734_shape = shape(x = x_25)[name = tensor<string, []>("op_734_shape")]; |
| tensor<int32, []> gather_34_batch_dims_0 = const()[name = tensor<string, []>("gather_34_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_34_validate_indices_0 = const()[name = tensor<string, []>("gather_34_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_22 = const()[name = tensor<string, []>("select_22"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_34_axis_1 = const()[name = tensor<string, []>("gather_34_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_34 = gather(axis = gather_34_axis_1, batch_dims = gather_34_batch_dims_0, indices = select_22, validate_indices = gather_34_validate_indices_0, x = var_734_shape)[name = tensor<string, []>("gather_34")]; |
| tensor<fp32, [1, ?, 768]> projected_13 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_3_self_attn_in_proj_weight, x = x_25)[name = tensor<string, []>("linear_13")]; |
| tensor<int32, [5]> concat_36x = const()[name = tensor<string, []>("concat_36x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_741 = reshape(shape = concat_36x, x = projected_13)[name = tensor<string, []>("op_741")]; |
| tensor<int32, [5]> var_742 = const()[name = tensor<string, []>("op_742"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_13_begin_0 = const()[name = tensor<string, []>("q_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_13_end_0 = const()[name = tensor<string, []>("q_13_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_13_end_mask_0 = const()[name = tensor<string, []>("q_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_13_squeeze_mask_0 = const()[name = tensor<string, []>("q_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_15 = transpose(perm = var_742, x = var_741)[name = tensor<string, []>("transpose_75")]; |
| tensor<fp32, [1, 4, ?, 64]> q_13 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = projected_15)[name = tensor<string, []>("q_13")]; |
| tensor<int32, [5]> k_13_begin_0 = const()[name = tensor<string, []>("k_13_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_13_end_0 = const()[name = tensor<string, []>("k_13_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_13_end_mask_0 = const()[name = tensor<string, []>("k_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_13_squeeze_mask_0 = const()[name = tensor<string, []>("k_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_13 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = projected_15)[name = tensor<string, []>("k_13")]; |
| tensor<int32, [5]> v_7_begin_0 = const()[name = tensor<string, []>("v_7_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_7_end_0 = const()[name = tensor<string, []>("v_7_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_7_end_mask_0 = const()[name = tensor<string, []>("v_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_7_squeeze_mask_0 = const()[name = tensor<string, []>("v_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_7 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = projected_15)[name = tensor<string, []>("v_7")]; |
| tensor<fp32, [32]> freqs_7 = const()[name = tensor<string, []>("freqs_7"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44430080)))]; |
| tensor<int32, [4]> var_754_shape = shape(x = q_13)[name = tensor<string, []>("op_754_shape")]; |
| tensor<int32, []> gather_38_batch_dims_0 = const()[name = tensor<string, []>("gather_38_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_38_validate_indices_0 = const()[name = tensor<string, []>("gather_38_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_23 = const()[name = tensor<string, []>("select_23"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_38_axis_1 = const()[name = tensor<string, []>("gather_38_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_38 = gather(axis = gather_38_axis_1, batch_dims = gather_38_batch_dims_0, indices = select_23, validate_indices = gather_38_validate_indices_0, x = var_754_shape)[name = tensor<string, []>("gather_38")]; |
| tensor<fp32, [1, 1]> var_757 = const()[name = tensor<string, []>("op_757"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_30 = const()[name = tensor<string, []>("const_30"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_31 = const()[name = tensor<string, []>("const_31"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_758 = range_1d(end = gather_38, start = const_30, step = const_31)[name = tensor<string, []>("op_758")]; |
| tensor<string, []> var_758_promoted_dtype_0 = const()[name = tensor<string, []>("op_758_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_758_promoted = cast(dtype = var_758_promoted_dtype_0, x = var_758)[name = tensor<string, []>("cast_461")]; |
| tensor<fp32, [1, ?]> ts_13 = add(x = var_757, y = var_758_promoted)[name = tensor<string, []>("ts_13")]; |
| tensor<int32, [4]> var_761 = const()[name = tensor<string, []>("op_761"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_15 = reshape(shape = var_761, x = ts_13)[name = tensor<string, []>("ts_15")]; |
| tensor<int32, [5]> concat_37x = const()[name = tensor<string, []>("concat_37x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_7 = reshape(shape = concat_37x, x = q_13)[name = tensor<string, []>("q2_7")]; |
| tensor<int32, [5]> concat_38x = const()[name = tensor<string, []>("concat_38x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_7 = reshape(shape = concat_38x, x = k_13)[name = tensor<string, []>("k2_7")]; |
| tensor<int32, [5]> qr_7_begin_0 = const()[name = tensor<string, []>("qr_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_7_end_0 = const()[name = tensor<string, []>("qr_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_7_end_mask_0 = const()[name = tensor<string, []>("qr_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_7_squeeze_mask_0 = const()[name = tensor<string, []>("qr_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_7 = slice_by_index(begin = qr_7_begin_0, end = qr_7_end_0, end_mask = qr_7_end_mask_0, squeeze_mask = qr_7_squeeze_mask_0, x = q2_7)[name = tensor<string, []>("qr_7")]; |
| tensor<int32, [5]> qi_7_begin_0 = const()[name = tensor<string, []>("qi_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_7_end_0 = const()[name = tensor<string, []>("qi_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_7_end_mask_0 = const()[name = tensor<string, []>("qi_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_7_squeeze_mask_0 = const()[name = tensor<string, []>("qi_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_7 = slice_by_index(begin = qi_7_begin_0, end = qi_7_end_0, end_mask = qi_7_end_mask_0, squeeze_mask = qi_7_squeeze_mask_0, x = q2_7)[name = tensor<string, []>("qi_7")]; |
| tensor<int32, [5]> kr_7_begin_0 = const()[name = tensor<string, []>("kr_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_7_end_0 = const()[name = tensor<string, []>("kr_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_7_end_mask_0 = const()[name = tensor<string, []>("kr_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_7_squeeze_mask_0 = const()[name = tensor<string, []>("kr_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_7 = slice_by_index(begin = kr_7_begin_0, end = kr_7_end_0, end_mask = kr_7_end_mask_0, squeeze_mask = kr_7_squeeze_mask_0, x = k2_7)[name = tensor<string, []>("kr_7")]; |
| tensor<int32, [5]> ki_7_begin_0 = const()[name = tensor<string, []>("ki_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_7_end_0 = const()[name = tensor<string, []>("ki_7_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_7_end_mask_0 = const()[name = tensor<string, []>("ki_7_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_7_squeeze_mask_0 = const()[name = tensor<string, []>("ki_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_7 = slice_by_index(begin = ki_7_begin_0, end = ki_7_end_0, end_mask = ki_7_end_mask_0, squeeze_mask = ki_7_squeeze_mask_0, x = k2_7)[name = tensor<string, []>("ki_7")]; |
| tensor<fp32, [1, 1, ?, 32]> var_774 = mul(x = freqs_7, y = ts_15)[name = tensor<string, []>("op_774")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_7 = cos(x = var_774)[name = tensor<string, []>("rotr_7")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_7 = sin(x = var_774)[name = tensor<string, []>("roti_7")]; |
| tensor<fp32, [1, 4, ?, 32]> var_778 = mul(x = qr_7, y = rotr_7)[name = tensor<string, []>("op_778")]; |
| tensor<fp32, [1, 4, ?, 32]> var_779 = mul(x = qi_7, y = roti_7)[name = tensor<string, []>("op_779")]; |
| tensor<fp32, [1, 4, ?, 32]> var_780 = sub(x = var_778, y = var_779)[name = tensor<string, []>("op_780")]; |
| tensor<fp32, [1, 4, ?, 32]> var_781 = mul(x = qr_7, y = roti_7)[name = tensor<string, []>("op_781")]; |
| tensor<fp32, [1, 4, ?, 32]> var_782 = mul(x = qi_7, y = rotr_7)[name = tensor<string, []>("op_782")]; |
| tensor<fp32, [1, 4, ?, 32]> var_783 = add(x = var_781, y = var_782)[name = tensor<string, []>("op_783")]; |
| tensor<int32, []> qo_7_axis_0 = const()[name = tensor<string, []>("qo_7_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_7 = stack(axis = qo_7_axis_0, values = (var_780, var_783))[name = tensor<string, []>("qo_7")]; |
| tensor<fp32, [1, 4, ?, 32]> var_786 = mul(x = kr_7, y = rotr_7)[name = tensor<string, []>("op_786")]; |
| tensor<fp32, [1, 4, ?, 32]> var_787 = mul(x = ki_7, y = roti_7)[name = tensor<string, []>("op_787")]; |
| tensor<fp32, [1, 4, ?, 32]> var_788 = sub(x = var_786, y = var_787)[name = tensor<string, []>("op_788")]; |
| tensor<fp32, [1, 4, ?, 32]> var_789 = mul(x = kr_7, y = roti_7)[name = tensor<string, []>("op_789")]; |
| tensor<fp32, [1, 4, ?, 32]> var_790 = mul(x = ki_7, y = rotr_7)[name = tensor<string, []>("op_790")]; |
| tensor<fp32, [1, 4, ?, 32]> var_791 = add(x = var_789, y = var_790)[name = tensor<string, []>("op_791")]; |
| tensor<int32, []> ko_7_axis_0 = const()[name = tensor<string, []>("ko_7_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_7 = stack(axis = ko_7_axis_0, values = (var_788, var_791))[name = tensor<string, []>("ko_7")]; |
| tensor<int32, [4]> concat_39x = const()[name = tensor<string, []>("concat_39x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_15 = reshape(shape = concat_39x, x = qo_7)[name = tensor<string, []>("q_15")]; |
| tensor<int32, [4]> concat_40x = const()[name = tensor<string, []>("concat_40x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_15 = reshape(shape = concat_40x, x = ko_7)[name = tensor<string, []>("k_15")]; |
| tensor<int32, []> const_32 = const()[name = tensor<string, []>("const_32"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_33 = const()[name = tensor<string, []>("const_33"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_7 = range_1d(end = gather_34, start = const_32, step = const_33)[name = tensor<string, []>("positions_7")]; |
| tensor<int32, [3]> concat_41x = const()[name = tensor<string, []>("concat_41x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_803 = reshape(shape = concat_41x, x = positions_7)[name = tensor<string, []>("op_803")]; |
| tensor<string, []> var_803_promoted_dtype_0 = const()[name = tensor<string, []>("op_803_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_803_promoted = cast(dtype = var_803_promoted_dtype_0, x = var_803)[name = tensor<string, []>("cast_460")]; |
| tensor<bool, [1, 1, ?]> valid_k_7 = less(x = var_803_promoted, y = var_394)[name = tensor<string, []>("valid_k_7")]; |
| tensor<int32, [3]> concat_42x = const()[name = tensor<string, []>("concat_42x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_808 = reshape(shape = concat_42x, x = positions_7)[name = tensor<string, []>("op_808")]; |
| tensor<int32, [1, ?, ?]> delta_7 = sub(x = var_808, y = var_803)[name = tensor<string, []>("delta_7")]; |
| tensor<int32, []> concat_44_axis_0 = const()[name = tensor<string, []>("concat_44_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_44_interleave_0 = const()[name = tensor<string, []>("concat_44_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_44 = concat(axis = concat_44_axis_0, interleave = concat_44_interleave_0, values = (var_295, gather_34, gather_34))[name = tensor<string, []>("concat_44")]; |
| tensor<fp32, []> fill_3_value_0 = const()[name = tensor<string, []>("fill_3_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_3 = fill(shape = concat_44, value = fill_3_value_0)[name = tensor<string, []>("fill_3")]; |
| tensor<string, []> attn_bias_25_dtype_0 = const()[name = tensor<string, []>("attn_bias_25_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_814 = greater_equal(x = delta_7, y = var_277)[name = tensor<string, []>("op_814")]; |
| tensor<bool, [1, ?, ?]> attn_bias_25 = cast(dtype = attn_bias_25_dtype_0, x = fill_3)[name = tensor<string, []>("cast_459")]; |
| tensor<bool, [1, ?, ?]> attn_bias_27 = logical_and(x = attn_bias_25, y = var_814)[name = tensor<string, []>("attn_bias_27")]; |
| tensor<bool, [1, ?, ?]> var_816 = less(x = delta_7, y = var_284)[name = tensor<string, []>("op_816")]; |
| tensor<bool, [1, ?, ?]> attn_bias_29 = logical_and(x = attn_bias_27, y = var_816)[name = tensor<string, []>("attn_bias_29")]; |
| tensor<bool, [1, ?, ?]> var_818 = logical_and(x = attn_bias_29, y = valid_k_7)[name = tensor<string, []>("op_818")]; |
| tensor<int32, [1]> var_820_axes_0 = const()[name = tensor<string, []>("op_820_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_820 = expand_dims(axes = var_820_axes_0, x = var_818)[name = tensor<string, []>("op_820")]; |
| tensor<string, []> cast_86_dtype_0 = const()[name = tensor<string, []>("cast_86_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_3_x_0 = const()[name = tensor<string, []>("sub_3_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_86 = cast(dtype = cast_86_dtype_0, x = var_820)[name = tensor<string, []>("cast_458")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_3 = sub(x = sub_3_x_0, y = cast_86)[name = tensor<string, []>("sub_3")]; |
| tensor<fp32, []> mul_6_x_0 = const()[name = tensor<string, []>("mul_6_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_6 = mul(x = mul_6_x_0, y = sub_3)[name = tensor<string, []>("mul_6")]; |
| tensor<fp32, []> mul_7_y_0 = const()[name = tensor<string, []>("mul_7_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_7 = mul(x = q_15, y = mul_7_y_0)[name = tensor<string, []>("mul_7")]; |
| tensor<bool, []> matmul_3_transpose_y_0 = const()[name = tensor<string, []>("matmul_3_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_3_transpose_x_0 = const()[name = tensor<string, []>("matmul_3_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_3 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = mul_7, y = k_15)[name = tensor<string, []>("matmul_3")]; |
| tensor<fp32, [1, 4, ?, ?]> add_3 = add(x = matmul_3, y = mul_6)[name = tensor<string, []>("add_3")]; |
| tensor<int32, []> softmax_3_axis_0 = const()[name = tensor<string, []>("softmax_3_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_3 = softmax(axis = softmax_3_axis_0, x = add_3)[name = tensor<string, []>("softmax_3")]; |
| 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<fp32, [1, 4, ?, 64]> out_13 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = softmax_3, y = v_7)[name = tensor<string, []>("out_13")]; |
| tensor<int32, [2]> concat_45x = const()[name = tensor<string, []>("concat_45x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_826 = reshape(shape = concat_45x, x = positions_7)[name = tensor<string, []>("op_826")]; |
| tensor<string, []> var_826_promoted_dtype_0 = const()[name = tensor<string, []>("op_826_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_826_promoted = cast(dtype = var_826_promoted_dtype_0, x = var_826)[name = tensor<string, []>("cast_457")]; |
| tensor<bool, [1, ?]> var_829 = less(x = var_826_promoted, y = var_417)[name = tensor<string, []>("op_829")]; |
| tensor<int32, [4]> concat_46x = const()[name = tensor<string, []>("concat_46x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_7 = reshape(shape = concat_46x, x = var_829)[name = tensor<string, []>("valid_q_7")]; |
| tensor<fp32, []> var_833 = const()[name = tensor<string, []>("op_833"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_15 = select(a = out_13, b = var_833, cond = valid_q_7)[name = tensor<string, []>("out_15")]; |
| tensor<int32, [4]> var_835_perm_0 = const()[name = tensor<string, []>("op_835_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_47x = const()[name = tensor<string, []>("concat_47x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_835 = transpose(perm = var_835_perm_0, x = out_15)[name = tensor<string, []>("transpose_74")]; |
| tensor<fp32, [1, ?, 256]> input_39 = reshape(shape = concat_47x, x = var_835)[name = tensor<string, []>("input_39")]; |
| tensor<fp32, [1, ?, 256]> x_27 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_3_self_attn_out_proj_weight, x = input_39)[name = tensor<string, []>("linear_14")]; |
| tensor<fp32, [1, ?, 256]> var_841 = mul(x = encoder_1_transformer_layers_3_layer_scale_1_scale, y = x_27)[name = tensor<string, []>("op_841")]; |
| tensor<fp32, [1, ?, 256]> input_41 = add(x = input_37, y = var_841)[name = tensor<string, []>("input_41")]; |
| tensor<int32, [1]> input_43_axes_0 = const()[name = tensor<string, []>("input_43_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_1_transformer_layers_3_norm2_bias, epsilon = var_291, gamma = encoder_1_transformer_layers_3_norm2_weight, x = input_41)[name = tensor<string, []>("input_43")]; |
| tensor<fp32, [1, ?, 1024]> input_45 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_3_ffn_0_weight, x = input_43)[name = tensor<string, []>("linear_15")]; |
| tensor<string, []> input_47_mode_0 = const()[name = tensor<string, []>("input_47_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_47 = gelu(mode = input_47_mode_0, x = input_45)[name = tensor<string, []>("input_47")]; |
| tensor<fp32, [1, ?, 256]> x_29 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_3_ffn_2_weight, x = input_47)[name = tensor<string, []>("linear_16")]; |
| tensor<fp32, [1, ?, 256]> var_856 = mul(x = encoder_1_transformer_layers_3_layer_scale_2_scale, y = x_29)[name = tensor<string, []>("op_856")]; |
| tensor<fp32, [1, ?, 256]> input_49 = add(x = input_41, y = var_856)[name = tensor<string, []>("input_49")]; |
| tensor<fp32, [384]> linear_17_bias_0 = const()[name = tensor<string, []>("linear_17_bias_0"), val = tensor<fp32, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44430272)))]; |
| tensor<fp32, [1, ?, 384]> var_859 = linear(bias = linear_17_bias_0, weight = encoder_1_output_proj_weight, x = input_49)[name = tensor<string, []>("linear_17")]; |
| tensor<int32, [3]> x_31_perm_0 = const()[name = tensor<string, []>("x_31_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [4]> var_870 = const()[name = tensor<string, []>("op_870"), val = tensor<int32, [4]>([1, 384, -1, 2])]; |
| tensor<fp32, [1, 384, ?]> x_31 = transpose(perm = x_31_perm_0, x = var_859)[name = tensor<string, []>("transpose_73")]; |
| tensor<fp32, [1, 384, ?, 2]> var_871 = reshape(shape = var_870, x = x_31)[name = tensor<string, []>("op_871")]; |
| tensor<int32, [4]> var_872 = const()[name = tensor<string, []>("op_872"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| tensor<int32, [3]> var_877 = const()[name = tensor<string, []>("op_877"), val = tensor<int32, [3]>([1, 768, -1])]; |
| tensor<fp32, [1, 384, 2, ?]> var_873 = transpose(perm = var_872, x = var_871)[name = tensor<string, []>("transpose_72")]; |
| tensor<fp32, [1, 768, ?]> x_33 = reshape(shape = var_877, x = var_873)[name = tensor<string, []>("x_33")]; |
| tensor<fp32, []> var_861_promoted = const()[name = tensor<string, []>("op_861_promoted"), val = tensor<fp32, []>(0x1p+1)]; |
| tensor<fp32, [1]> floor_div_1 = floor_div(x = input_lengths_1, y = var_861_promoted)[name = tensor<string, []>("floor_div_1")]; |
| tensor<int32, []> var_884 = const()[name = tensor<string, []>("op_884"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> var_891 = const()[name = tensor<string, []>("op_891"), val = tensor<int32, []>(1200)]; |
| tensor<fp32, []> var_898 = const()[name = tensor<string, []>("op_898"), val = tensor<fp32, []>(0x1.4f8b58p-17)]; |
| tensor<int32, []> var_902 = const()[name = tensor<string, []>("op_902"), val = tensor<int32, []>(1)]; |
| tensor<int32, [3]> input_51_perm_0 = const()[name = tensor<string, []>("input_51_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, ?, 768]> input_51 = transpose(perm = input_51_perm_0, x = x_33)[name = tensor<string, []>("transpose_71")]; |
| tensor<fp32, [1, ?, 256]> x_35 = linear(bias = linear_0_bias_0, weight = encoder_3_input_proj_weight, x = input_51)[name = tensor<string, []>("linear_18")]; |
| tensor<int32, [1]> x_37_axes_0 = const()[name = tensor<string, []>("x_37_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_37 = layer_norm(axes = x_37_axes_0, beta = encoder_3_transformer_layers_0_norm1_bias, epsilon = var_898, gamma = encoder_3_transformer_layers_0_norm1_weight, x = x_35)[name = tensor<string, []>("x_37")]; |
| tensor<int32, [3]> var_926_shape = shape(x = x_37)[name = tensor<string, []>("op_926_shape")]; |
| tensor<int32, []> gather_46_batch_dims_0 = const()[name = tensor<string, []>("gather_46_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_46_validate_indices_0 = const()[name = tensor<string, []>("gather_46_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_24 = const()[name = tensor<string, []>("select_24"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_46_axis_1 = const()[name = tensor<string, []>("gather_46_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_46 = gather(axis = gather_46_axis_1, batch_dims = gather_46_batch_dims_0, indices = select_24, validate_indices = gather_46_validate_indices_0, x = var_926_shape)[name = tensor<string, []>("gather_46")]; |
| tensor<fp32, [1, ?, 768]> projected_17 = linear(bias = linear_1_bias_0, weight = encoder_3_transformer_layers_0_self_attn_in_proj_weight, x = x_37)[name = tensor<string, []>("linear_19")]; |
| tensor<int32, [5]> concat_48x = const()[name = tensor<string, []>("concat_48x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_933 = reshape(shape = concat_48x, x = projected_17)[name = tensor<string, []>("op_933")]; |
| tensor<int32, [5]> var_934 = const()[name = tensor<string, []>("op_934"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_17_begin_0 = const()[name = tensor<string, []>("q_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_17_end_0 = const()[name = tensor<string, []>("q_17_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_17_end_mask_0 = const()[name = tensor<string, []>("q_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_17_squeeze_mask_0 = const()[name = tensor<string, []>("q_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_19 = transpose(perm = var_934, x = var_933)[name = tensor<string, []>("transpose_70")]; |
| tensor<fp32, [1, 4, ?, 64]> q_17 = slice_by_index(begin = q_17_begin_0, end = q_17_end_0, end_mask = q_17_end_mask_0, squeeze_mask = q_17_squeeze_mask_0, x = projected_19)[name = tensor<string, []>("q_17")]; |
| tensor<int32, [5]> k_17_begin_0 = const()[name = tensor<string, []>("k_17_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_17_end_0 = const()[name = tensor<string, []>("k_17_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_17_end_mask_0 = const()[name = tensor<string, []>("k_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_17_squeeze_mask_0 = const()[name = tensor<string, []>("k_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_17 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = projected_19)[name = tensor<string, []>("k_17")]; |
| tensor<int32, [5]> v_9_begin_0 = const()[name = tensor<string, []>("v_9_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_9_end_0 = const()[name = tensor<string, []>("v_9_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_9_end_mask_0 = const()[name = tensor<string, []>("v_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_9_squeeze_mask_0 = const()[name = tensor<string, []>("v_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_9 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = projected_19)[name = tensor<string, []>("v_9")]; |
| tensor<fp32, [32]> freqs_9 = const()[name = tensor<string, []>("freqs_9"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44431872)))]; |
| tensor<int32, [4]> var_946_shape = shape(x = q_17)[name = tensor<string, []>("op_946_shape")]; |
| tensor<int32, []> gather_50_batch_dims_0 = const()[name = tensor<string, []>("gather_50_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_50_validate_indices_0 = const()[name = tensor<string, []>("gather_50_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_25 = const()[name = tensor<string, []>("select_25"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_50_axis_1 = const()[name = tensor<string, []>("gather_50_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_50 = gather(axis = gather_50_axis_1, batch_dims = gather_50_batch_dims_0, indices = select_25, validate_indices = gather_50_validate_indices_0, x = var_946_shape)[name = tensor<string, []>("gather_50")]; |
| tensor<fp32, [1, 1]> var_949 = const()[name = tensor<string, []>("op_949"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_39 = const()[name = tensor<string, []>("const_39"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_40 = const()[name = tensor<string, []>("const_40"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_950 = range_1d(end = gather_50, start = const_39, step = const_40)[name = tensor<string, []>("op_950")]; |
| tensor<string, []> var_950_promoted_dtype_0 = const()[name = tensor<string, []>("op_950_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_950_promoted = cast(dtype = var_950_promoted_dtype_0, x = var_950)[name = tensor<string, []>("cast_456")]; |
| tensor<fp32, [1, ?]> ts_17 = add(x = var_949, y = var_950_promoted)[name = tensor<string, []>("ts_17")]; |
| tensor<int32, [4]> var_953 = const()[name = tensor<string, []>("op_953"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_19 = reshape(shape = var_953, x = ts_17)[name = tensor<string, []>("ts_19")]; |
| tensor<int32, [5]> concat_49x = const()[name = tensor<string, []>("concat_49x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_9 = reshape(shape = concat_49x, x = q_17)[name = tensor<string, []>("q2_9")]; |
| tensor<int32, [5]> concat_50x = const()[name = tensor<string, []>("concat_50x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_9 = reshape(shape = concat_50x, x = k_17)[name = tensor<string, []>("k2_9")]; |
| tensor<int32, [5]> qr_9_begin_0 = const()[name = tensor<string, []>("qr_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_9_end_0 = const()[name = tensor<string, []>("qr_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_9_end_mask_0 = const()[name = tensor<string, []>("qr_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_9_squeeze_mask_0 = const()[name = tensor<string, []>("qr_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_9 = slice_by_index(begin = qr_9_begin_0, end = qr_9_end_0, end_mask = qr_9_end_mask_0, squeeze_mask = qr_9_squeeze_mask_0, x = q2_9)[name = tensor<string, []>("qr_9")]; |
| tensor<int32, [5]> qi_9_begin_0 = const()[name = tensor<string, []>("qi_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_9_end_0 = const()[name = tensor<string, []>("qi_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_9_end_mask_0 = const()[name = tensor<string, []>("qi_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_9_squeeze_mask_0 = const()[name = tensor<string, []>("qi_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_9 = slice_by_index(begin = qi_9_begin_0, end = qi_9_end_0, end_mask = qi_9_end_mask_0, squeeze_mask = qi_9_squeeze_mask_0, x = q2_9)[name = tensor<string, []>("qi_9")]; |
| tensor<int32, [5]> kr_9_begin_0 = const()[name = tensor<string, []>("kr_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_9_end_0 = const()[name = tensor<string, []>("kr_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_9_end_mask_0 = const()[name = tensor<string, []>("kr_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_9_squeeze_mask_0 = const()[name = tensor<string, []>("kr_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_9 = slice_by_index(begin = kr_9_begin_0, end = kr_9_end_0, end_mask = kr_9_end_mask_0, squeeze_mask = kr_9_squeeze_mask_0, x = k2_9)[name = tensor<string, []>("kr_9")]; |
| tensor<int32, [5]> ki_9_begin_0 = const()[name = tensor<string, []>("ki_9_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_9_end_0 = const()[name = tensor<string, []>("ki_9_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_9_end_mask_0 = const()[name = tensor<string, []>("ki_9_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_9_squeeze_mask_0 = const()[name = tensor<string, []>("ki_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_9 = slice_by_index(begin = ki_9_begin_0, end = ki_9_end_0, end_mask = ki_9_end_mask_0, squeeze_mask = ki_9_squeeze_mask_0, x = k2_9)[name = tensor<string, []>("ki_9")]; |
| tensor<fp32, [1, 1, ?, 32]> var_966 = mul(x = freqs_9, y = ts_19)[name = tensor<string, []>("op_966")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_9 = cos(x = var_966)[name = tensor<string, []>("rotr_9")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_9 = sin(x = var_966)[name = tensor<string, []>("roti_9")]; |
| tensor<fp32, [1, 4, ?, 32]> var_970 = mul(x = qr_9, y = rotr_9)[name = tensor<string, []>("op_970")]; |
| tensor<fp32, [1, 4, ?, 32]> var_971 = mul(x = qi_9, y = roti_9)[name = tensor<string, []>("op_971")]; |
| tensor<fp32, [1, 4, ?, 32]> var_972 = sub(x = var_970, y = var_971)[name = tensor<string, []>("op_972")]; |
| tensor<fp32, [1, 4, ?, 32]> var_973 = mul(x = qr_9, y = roti_9)[name = tensor<string, []>("op_973")]; |
| tensor<fp32, [1, 4, ?, 32]> var_974 = mul(x = qi_9, y = rotr_9)[name = tensor<string, []>("op_974")]; |
| tensor<fp32, [1, 4, ?, 32]> var_975 = add(x = var_973, y = var_974)[name = tensor<string, []>("op_975")]; |
| tensor<int32, []> qo_9_axis_0 = const()[name = tensor<string, []>("qo_9_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_9 = stack(axis = qo_9_axis_0, values = (var_972, var_975))[name = tensor<string, []>("qo_9")]; |
| tensor<fp32, [1, 4, ?, 32]> var_978 = mul(x = kr_9, y = rotr_9)[name = tensor<string, []>("op_978")]; |
| tensor<fp32, [1, 4, ?, 32]> var_979 = mul(x = ki_9, y = roti_9)[name = tensor<string, []>("op_979")]; |
| tensor<fp32, [1, 4, ?, 32]> var_980 = sub(x = var_978, y = var_979)[name = tensor<string, []>("op_980")]; |
| tensor<fp32, [1, 4, ?, 32]> var_981 = mul(x = kr_9, y = roti_9)[name = tensor<string, []>("op_981")]; |
| tensor<fp32, [1, 4, ?, 32]> var_982 = mul(x = ki_9, y = rotr_9)[name = tensor<string, []>("op_982")]; |
| tensor<fp32, [1, 4, ?, 32]> var_983 = add(x = var_981, y = var_982)[name = tensor<string, []>("op_983")]; |
| tensor<int32, []> ko_9_axis_0 = const()[name = tensor<string, []>("ko_9_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_9 = stack(axis = ko_9_axis_0, values = (var_980, var_983))[name = tensor<string, []>("ko_9")]; |
| tensor<int32, [4]> concat_51x = const()[name = tensor<string, []>("concat_51x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_19 = reshape(shape = concat_51x, x = qo_9)[name = tensor<string, []>("q_19")]; |
| tensor<int32, [4]> concat_52x = const()[name = tensor<string, []>("concat_52x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_19 = reshape(shape = concat_52x, x = ko_9)[name = tensor<string, []>("k_19")]; |
| tensor<int32, []> const_41 = const()[name = tensor<string, []>("const_41"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_42 = const()[name = tensor<string, []>("const_42"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_9 = range_1d(end = gather_46, start = const_41, step = const_42)[name = tensor<string, []>("positions_9")]; |
| tensor<int32, [3]> concat_53x = const()[name = tensor<string, []>("concat_53x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_995 = reshape(shape = concat_53x, x = positions_9)[name = tensor<string, []>("op_995")]; |
| tensor<int32, [3]> var_996 = const()[name = tensor<string, []>("op_996"), val = tensor<int32, [3]>([-1, 1, 1])]; |
| tensor<fp32, [1, 1, 1]> var_997 = reshape(shape = var_996, x = floor_div_1)[name = tensor<string, []>("op_997")]; |
| tensor<string, []> var_995_promoted_dtype_0 = const()[name = tensor<string, []>("op_995_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_995_promoted = cast(dtype = var_995_promoted_dtype_0, x = var_995)[name = tensor<string, []>("cast_455")]; |
| tensor<bool, [1, 1, ?]> valid_k_9 = less(x = var_995_promoted, y = var_997)[name = tensor<string, []>("valid_k_9")]; |
| tensor<int32, [3]> concat_54x = const()[name = tensor<string, []>("concat_54x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_1000 = reshape(shape = concat_54x, x = positions_9)[name = tensor<string, []>("op_1000")]; |
| tensor<int32, [1, ?, ?]> delta_9 = sub(x = var_1000, y = var_995)[name = tensor<string, []>("delta_9")]; |
| tensor<int32, []> concat_56_axis_0 = const()[name = tensor<string, []>("concat_56_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_56_interleave_0 = const()[name = tensor<string, []>("concat_56_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_56 = concat(axis = concat_56_axis_0, interleave = concat_56_interleave_0, values = (var_902, gather_46, gather_46))[name = tensor<string, []>("concat_56")]; |
| tensor<fp32, []> fill_4_value_0 = const()[name = tensor<string, []>("fill_4_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_4 = fill(shape = concat_56, value = fill_4_value_0)[name = tensor<string, []>("fill_4")]; |
| tensor<string, []> attn_bias_33_dtype_0 = const()[name = tensor<string, []>("attn_bias_33_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_1006 = greater_equal(x = delta_9, y = var_884)[name = tensor<string, []>("op_1006")]; |
| tensor<bool, [1, ?, ?]> attn_bias_33 = cast(dtype = attn_bias_33_dtype_0, x = fill_4)[name = tensor<string, []>("cast_454")]; |
| tensor<bool, [1, ?, ?]> attn_bias_35 = logical_and(x = attn_bias_33, y = var_1006)[name = tensor<string, []>("attn_bias_35")]; |
| tensor<bool, [1, ?, ?]> var_1008 = less(x = delta_9, y = var_891)[name = tensor<string, []>("op_1008")]; |
| tensor<bool, [1, ?, ?]> attn_bias_37 = logical_and(x = attn_bias_35, y = var_1008)[name = tensor<string, []>("attn_bias_37")]; |
| tensor<bool, [1, ?, ?]> var_1010 = logical_and(x = attn_bias_37, y = valid_k_9)[name = tensor<string, []>("op_1010")]; |
| tensor<int32, [1]> var_1012_axes_0 = const()[name = tensor<string, []>("op_1012_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_1012 = expand_dims(axes = var_1012_axes_0, x = var_1010)[name = tensor<string, []>("op_1012")]; |
| tensor<string, []> cast_110_dtype_0 = const()[name = tensor<string, []>("cast_110_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_4_x_0 = const()[name = tensor<string, []>("sub_4_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_110 = cast(dtype = cast_110_dtype_0, x = var_1012)[name = tensor<string, []>("cast_453")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_4 = sub(x = sub_4_x_0, y = cast_110)[name = tensor<string, []>("sub_4")]; |
| tensor<fp32, []> mul_8_x_0 = const()[name = tensor<string, []>("mul_8_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_8 = mul(x = mul_8_x_0, y = sub_4)[name = tensor<string, []>("mul_8")]; |
| tensor<fp32, []> mul_9_y_0 = const()[name = tensor<string, []>("mul_9_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_9 = mul(x = q_19, y = mul_9_y_0)[name = tensor<string, []>("mul_9")]; |
| tensor<bool, []> matmul_4_transpose_y_0 = const()[name = tensor<string, []>("matmul_4_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_4_transpose_x_0 = const()[name = tensor<string, []>("matmul_4_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_4 = matmul(transpose_x = matmul_4_transpose_x_0, transpose_y = matmul_4_transpose_y_0, x = mul_9, y = k_19)[name = tensor<string, []>("matmul_4")]; |
| tensor<fp32, [1, 4, ?, ?]> add_4 = add(x = matmul_4, y = mul_8)[name = tensor<string, []>("add_4")]; |
| tensor<int32, []> softmax_4_axis_0 = const()[name = tensor<string, []>("softmax_4_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_4 = softmax(axis = softmax_4_axis_0, x = add_4)[name = tensor<string, []>("softmax_4")]; |
| 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<fp32, [1, 4, ?, 64]> out_17 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = softmax_4, y = v_9)[name = tensor<string, []>("out_17")]; |
| tensor<int32, [2]> concat_57x = const()[name = tensor<string, []>("concat_57x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_1018 = reshape(shape = concat_57x, x = positions_9)[name = tensor<string, []>("op_1018")]; |
| tensor<int32, [2]> var_1019 = const()[name = tensor<string, []>("op_1019"), val = tensor<int32, [2]>([-1, 1])]; |
| tensor<fp32, [1, 1]> var_1020 = reshape(shape = var_1019, x = floor_div_1)[name = tensor<string, []>("op_1020")]; |
| tensor<string, []> var_1018_promoted_dtype_0 = const()[name = tensor<string, []>("op_1018_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_1018_promoted = cast(dtype = var_1018_promoted_dtype_0, x = var_1018)[name = tensor<string, []>("cast_452")]; |
| tensor<bool, [1, ?]> var_1021 = less(x = var_1018_promoted, y = var_1020)[name = tensor<string, []>("op_1021")]; |
| tensor<int32, [4]> concat_58x = const()[name = tensor<string, []>("concat_58x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_9 = reshape(shape = concat_58x, x = var_1021)[name = tensor<string, []>("valid_q_9")]; |
| tensor<fp32, []> var_1025 = const()[name = tensor<string, []>("op_1025"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_19 = select(a = out_17, b = var_1025, cond = valid_q_9)[name = tensor<string, []>("out_19")]; |
| tensor<int32, [4]> var_1027_perm_0 = const()[name = tensor<string, []>("op_1027_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_59x = const()[name = tensor<string, []>("concat_59x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_1027 = transpose(perm = var_1027_perm_0, x = out_19)[name = tensor<string, []>("transpose_69")]; |
| tensor<fp32, [1, ?, 256]> input_53 = reshape(shape = concat_59x, x = var_1027)[name = tensor<string, []>("input_53")]; |
| tensor<fp32, [1, ?, 256]> x_39 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_0_self_attn_out_proj_weight, x = input_53)[name = tensor<string, []>("linear_20")]; |
| tensor<fp32, [1, ?, 256]> var_1033 = mul(x = encoder_3_transformer_layers_0_layer_scale_1_scale, y = x_39)[name = tensor<string, []>("op_1033")]; |
| tensor<fp32, [1, ?, 256]> input_55 = add(x = x_35, y = var_1033)[name = tensor<string, []>("input_55")]; |
| tensor<int32, [1]> input_57_axes_0 = const()[name = tensor<string, []>("input_57_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_57 = layer_norm(axes = input_57_axes_0, beta = encoder_3_transformer_layers_0_norm2_bias, epsilon = var_898, gamma = encoder_3_transformer_layers_0_norm2_weight, x = input_55)[name = tensor<string, []>("input_57")]; |
| tensor<fp32, [1, ?, 1024]> input_59 = linear(bias = linear_3_bias_0, weight = encoder_3_transformer_layers_0_ffn_0_weight, x = input_57)[name = tensor<string, []>("linear_21")]; |
| tensor<string, []> input_61_mode_0 = const()[name = tensor<string, []>("input_61_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_61 = gelu(mode = input_61_mode_0, x = input_59)[name = tensor<string, []>("input_61")]; |
| tensor<fp32, [1, ?, 256]> x_41 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_0_ffn_2_weight, x = input_61)[name = tensor<string, []>("linear_22")]; |
| tensor<fp32, [1, ?, 256]> var_1048 = mul(x = encoder_3_transformer_layers_0_layer_scale_2_scale, y = x_41)[name = tensor<string, []>("op_1048")]; |
| tensor<fp32, [1, ?, 256]> input_63 = add(x = input_55, y = var_1048)[name = tensor<string, []>("input_63")]; |
| tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_43 = layer_norm(axes = x_43_axes_0, beta = encoder_3_transformer_layers_1_norm1_bias, epsilon = var_898, gamma = encoder_3_transformer_layers_1_norm1_weight, x = input_63)[name = tensor<string, []>("x_43")]; |
| tensor<int32, [3]> var_1063_shape = shape(x = x_43)[name = tensor<string, []>("op_1063_shape")]; |
| tensor<int32, []> gather_56_batch_dims_0 = const()[name = tensor<string, []>("gather_56_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_56_validate_indices_0 = const()[name = tensor<string, []>("gather_56_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_26 = const()[name = tensor<string, []>("select_26"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_56_axis_1 = const()[name = tensor<string, []>("gather_56_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_56 = gather(axis = gather_56_axis_1, batch_dims = gather_56_batch_dims_0, indices = select_26, validate_indices = gather_56_validate_indices_0, x = var_1063_shape)[name = tensor<string, []>("gather_56")]; |
| tensor<fp32, [1, ?, 768]> projected_21 = linear(bias = linear_1_bias_0, weight = encoder_3_transformer_layers_1_self_attn_in_proj_weight, x = x_43)[name = tensor<string, []>("linear_23")]; |
| tensor<int32, [5]> concat_60x = const()[name = tensor<string, []>("concat_60x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_1070 = reshape(shape = concat_60x, x = projected_21)[name = tensor<string, []>("op_1070")]; |
| tensor<int32, [5]> var_1071 = const()[name = tensor<string, []>("op_1071"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_21_begin_0 = const()[name = tensor<string, []>("q_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_21_end_0 = const()[name = tensor<string, []>("q_21_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_21_end_mask_0 = const()[name = tensor<string, []>("q_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_21_squeeze_mask_0 = const()[name = tensor<string, []>("q_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_23 = transpose(perm = var_1071, x = var_1070)[name = tensor<string, []>("transpose_68")]; |
| tensor<fp32, [1, 4, ?, 64]> q_21 = slice_by_index(begin = q_21_begin_0, end = q_21_end_0, end_mask = q_21_end_mask_0, squeeze_mask = q_21_squeeze_mask_0, x = projected_23)[name = tensor<string, []>("q_21")]; |
| tensor<int32, [5]> k_21_begin_0 = const()[name = tensor<string, []>("k_21_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_21_end_0 = const()[name = tensor<string, []>("k_21_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_21_end_mask_0 = const()[name = tensor<string, []>("k_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_21_squeeze_mask_0 = const()[name = tensor<string, []>("k_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_21 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = projected_23)[name = tensor<string, []>("k_21")]; |
| tensor<int32, [5]> v_11_begin_0 = const()[name = tensor<string, []>("v_11_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_11_end_0 = const()[name = tensor<string, []>("v_11_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_11_end_mask_0 = const()[name = tensor<string, []>("v_11_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_11_squeeze_mask_0 = const()[name = tensor<string, []>("v_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = projected_23)[name = tensor<string, []>("v_11")]; |
| tensor<fp32, [32]> freqs_11 = const()[name = tensor<string, []>("freqs_11"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44432064)))]; |
| tensor<int32, [4]> var_1083_shape = shape(x = q_21)[name = tensor<string, []>("op_1083_shape")]; |
| tensor<int32, []> gather_60_batch_dims_0 = const()[name = tensor<string, []>("gather_60_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_60_validate_indices_0 = const()[name = tensor<string, []>("gather_60_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_27 = const()[name = tensor<string, []>("select_27"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_60_axis_1 = const()[name = tensor<string, []>("gather_60_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_60 = gather(axis = gather_60_axis_1, batch_dims = gather_60_batch_dims_0, indices = select_27, validate_indices = gather_60_validate_indices_0, x = var_1083_shape)[name = tensor<string, []>("gather_60")]; |
| tensor<fp32, [1, 1]> var_1086 = const()[name = tensor<string, []>("op_1086"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_48 = const()[name = tensor<string, []>("const_48"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_49 = const()[name = tensor<string, []>("const_49"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_1087 = range_1d(end = gather_60, start = const_48, step = const_49)[name = tensor<string, []>("op_1087")]; |
| tensor<string, []> var_1087_promoted_dtype_0 = const()[name = tensor<string, []>("op_1087_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_1087_promoted = cast(dtype = var_1087_promoted_dtype_0, x = var_1087)[name = tensor<string, []>("cast_451")]; |
| tensor<fp32, [1, ?]> ts_21 = add(x = var_1086, y = var_1087_promoted)[name = tensor<string, []>("ts_21")]; |
| tensor<int32, [4]> var_1090 = const()[name = tensor<string, []>("op_1090"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_23 = reshape(shape = var_1090, x = ts_21)[name = tensor<string, []>("ts_23")]; |
| tensor<int32, [5]> concat_61x = const()[name = tensor<string, []>("concat_61x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_11 = reshape(shape = concat_61x, x = q_21)[name = tensor<string, []>("q2_11")]; |
| tensor<int32, [5]> concat_62x = const()[name = tensor<string, []>("concat_62x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_11 = reshape(shape = concat_62x, x = k_21)[name = tensor<string, []>("k2_11")]; |
| tensor<int32, [5]> qr_11_begin_0 = const()[name = tensor<string, []>("qr_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_11_end_0 = const()[name = tensor<string, []>("qr_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_11_end_mask_0 = const()[name = tensor<string, []>("qr_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_11_squeeze_mask_0 = const()[name = tensor<string, []>("qr_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_11 = slice_by_index(begin = qr_11_begin_0, end = qr_11_end_0, end_mask = qr_11_end_mask_0, squeeze_mask = qr_11_squeeze_mask_0, x = q2_11)[name = tensor<string, []>("qr_11")]; |
| tensor<int32, [5]> qi_11_begin_0 = const()[name = tensor<string, []>("qi_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_11_end_0 = const()[name = tensor<string, []>("qi_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_11_end_mask_0 = const()[name = tensor<string, []>("qi_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_11_squeeze_mask_0 = const()[name = tensor<string, []>("qi_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_11 = slice_by_index(begin = qi_11_begin_0, end = qi_11_end_0, end_mask = qi_11_end_mask_0, squeeze_mask = qi_11_squeeze_mask_0, x = q2_11)[name = tensor<string, []>("qi_11")]; |
| tensor<int32, [5]> kr_11_begin_0 = const()[name = tensor<string, []>("kr_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_11_end_0 = const()[name = tensor<string, []>("kr_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_11_end_mask_0 = const()[name = tensor<string, []>("kr_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_11_squeeze_mask_0 = const()[name = tensor<string, []>("kr_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_11 = slice_by_index(begin = kr_11_begin_0, end = kr_11_end_0, end_mask = kr_11_end_mask_0, squeeze_mask = kr_11_squeeze_mask_0, x = k2_11)[name = tensor<string, []>("kr_11")]; |
| tensor<int32, [5]> ki_11_begin_0 = const()[name = tensor<string, []>("ki_11_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_11_end_0 = const()[name = tensor<string, []>("ki_11_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_11_end_mask_0 = const()[name = tensor<string, []>("ki_11_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_11_squeeze_mask_0 = const()[name = tensor<string, []>("ki_11_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_11 = slice_by_index(begin = ki_11_begin_0, end = ki_11_end_0, end_mask = ki_11_end_mask_0, squeeze_mask = ki_11_squeeze_mask_0, x = k2_11)[name = tensor<string, []>("ki_11")]; |
| tensor<fp32, [1, 1, ?, 32]> var_1103 = mul(x = freqs_11, y = ts_23)[name = tensor<string, []>("op_1103")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_11 = cos(x = var_1103)[name = tensor<string, []>("rotr_11")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_11 = sin(x = var_1103)[name = tensor<string, []>("roti_11")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1107 = mul(x = qr_11, y = rotr_11)[name = tensor<string, []>("op_1107")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1108 = mul(x = qi_11, y = roti_11)[name = tensor<string, []>("op_1108")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1109 = sub(x = var_1107, y = var_1108)[name = tensor<string, []>("op_1109")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1110 = mul(x = qr_11, y = roti_11)[name = tensor<string, []>("op_1110")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1111 = mul(x = qi_11, y = rotr_11)[name = tensor<string, []>("op_1111")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1112 = add(x = var_1110, y = var_1111)[name = tensor<string, []>("op_1112")]; |
| tensor<int32, []> qo_11_axis_0 = const()[name = tensor<string, []>("qo_11_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_11 = stack(axis = qo_11_axis_0, values = (var_1109, var_1112))[name = tensor<string, []>("qo_11")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1115 = mul(x = kr_11, y = rotr_11)[name = tensor<string, []>("op_1115")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1116 = mul(x = ki_11, y = roti_11)[name = tensor<string, []>("op_1116")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1117 = sub(x = var_1115, y = var_1116)[name = tensor<string, []>("op_1117")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1118 = mul(x = kr_11, y = roti_11)[name = tensor<string, []>("op_1118")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1119 = mul(x = ki_11, y = rotr_11)[name = tensor<string, []>("op_1119")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1120 = add(x = var_1118, y = var_1119)[name = tensor<string, []>("op_1120")]; |
| tensor<int32, []> ko_11_axis_0 = const()[name = tensor<string, []>("ko_11_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_11 = stack(axis = ko_11_axis_0, values = (var_1117, var_1120))[name = tensor<string, []>("ko_11")]; |
| tensor<int32, [4]> concat_63x = const()[name = tensor<string, []>("concat_63x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_23 = reshape(shape = concat_63x, x = qo_11)[name = tensor<string, []>("q_23")]; |
| tensor<int32, [4]> concat_64x = const()[name = tensor<string, []>("concat_64x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_23 = reshape(shape = concat_64x, x = ko_11)[name = tensor<string, []>("k_23")]; |
| tensor<int32, []> const_50 = const()[name = tensor<string, []>("const_50"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_51 = const()[name = tensor<string, []>("const_51"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_11 = range_1d(end = gather_56, start = const_50, step = const_51)[name = tensor<string, []>("positions_11")]; |
| tensor<int32, [3]> concat_65x = const()[name = tensor<string, []>("concat_65x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_1132 = reshape(shape = concat_65x, x = positions_11)[name = tensor<string, []>("op_1132")]; |
| tensor<string, []> var_1132_promoted_dtype_0 = const()[name = tensor<string, []>("op_1132_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_1132_promoted = cast(dtype = var_1132_promoted_dtype_0, x = var_1132)[name = tensor<string, []>("cast_450")]; |
| tensor<bool, [1, 1, ?]> valid_k_11 = less(x = var_1132_promoted, y = var_997)[name = tensor<string, []>("valid_k_11")]; |
| tensor<int32, [3]> concat_66x = const()[name = tensor<string, []>("concat_66x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_1137 = reshape(shape = concat_66x, x = positions_11)[name = tensor<string, []>("op_1137")]; |
| tensor<int32, [1, ?, ?]> delta_11 = sub(x = var_1137, y = var_1132)[name = tensor<string, []>("delta_11")]; |
| tensor<int32, []> concat_68_axis_0 = const()[name = tensor<string, []>("concat_68_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_68_interleave_0 = const()[name = tensor<string, []>("concat_68_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_68 = concat(axis = concat_68_axis_0, interleave = concat_68_interleave_0, values = (var_902, gather_56, gather_56))[name = tensor<string, []>("concat_68")]; |
| tensor<fp32, []> fill_5_value_0 = const()[name = tensor<string, []>("fill_5_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_5 = fill(shape = concat_68, value = fill_5_value_0)[name = tensor<string, []>("fill_5")]; |
| tensor<string, []> attn_bias_41_dtype_0 = const()[name = tensor<string, []>("attn_bias_41_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_1143 = greater_equal(x = delta_11, y = var_884)[name = tensor<string, []>("op_1143")]; |
| tensor<bool, [1, ?, ?]> attn_bias_41 = cast(dtype = attn_bias_41_dtype_0, x = fill_5)[name = tensor<string, []>("cast_449")]; |
| tensor<bool, [1, ?, ?]> attn_bias_43 = logical_and(x = attn_bias_41, y = var_1143)[name = tensor<string, []>("attn_bias_43")]; |
| tensor<bool, [1, ?, ?]> var_1145 = less(x = delta_11, y = var_891)[name = tensor<string, []>("op_1145")]; |
| tensor<bool, [1, ?, ?]> attn_bias_45 = logical_and(x = attn_bias_43, y = var_1145)[name = tensor<string, []>("attn_bias_45")]; |
| tensor<bool, [1, ?, ?]> var_1147 = logical_and(x = attn_bias_45, y = valid_k_11)[name = tensor<string, []>("op_1147")]; |
| tensor<int32, [1]> var_1149_axes_0 = const()[name = tensor<string, []>("op_1149_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_1149 = expand_dims(axes = var_1149_axes_0, x = var_1147)[name = tensor<string, []>("op_1149")]; |
| tensor<string, []> cast_132_dtype_0 = const()[name = tensor<string, []>("cast_132_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_5_x_0 = const()[name = tensor<string, []>("sub_5_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_132 = cast(dtype = cast_132_dtype_0, x = var_1149)[name = tensor<string, []>("cast_448")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_5 = sub(x = sub_5_x_0, y = cast_132)[name = tensor<string, []>("sub_5")]; |
| tensor<fp32, []> mul_10_x_0 = const()[name = tensor<string, []>("mul_10_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_10 = mul(x = mul_10_x_0, y = sub_5)[name = tensor<string, []>("mul_10")]; |
| tensor<fp32, []> mul_11_y_0 = const()[name = tensor<string, []>("mul_11_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_11 = mul(x = q_23, y = mul_11_y_0)[name = tensor<string, []>("mul_11")]; |
| tensor<bool, []> matmul_5_transpose_y_0 = const()[name = tensor<string, []>("matmul_5_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_5_transpose_x_0 = const()[name = tensor<string, []>("matmul_5_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_5 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = mul_11, y = k_23)[name = tensor<string, []>("matmul_5")]; |
| tensor<fp32, [1, 4, ?, ?]> add_5 = add(x = matmul_5, y = mul_10)[name = tensor<string, []>("add_5")]; |
| tensor<int32, []> softmax_5_axis_0 = const()[name = tensor<string, []>("softmax_5_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_5 = softmax(axis = softmax_5_axis_0, x = add_5)[name = tensor<string, []>("softmax_5")]; |
| 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<fp32, [1, 4, ?, 64]> out_21 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = softmax_5, y = v_11)[name = tensor<string, []>("out_21")]; |
| tensor<int32, [2]> concat_69x = const()[name = tensor<string, []>("concat_69x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_1155 = reshape(shape = concat_69x, x = positions_11)[name = tensor<string, []>("op_1155")]; |
| tensor<string, []> var_1155_promoted_dtype_0 = const()[name = tensor<string, []>("op_1155_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_1155_promoted = cast(dtype = var_1155_promoted_dtype_0, x = var_1155)[name = tensor<string, []>("cast_447")]; |
| tensor<bool, [1, ?]> var_1158 = less(x = var_1155_promoted, y = var_1020)[name = tensor<string, []>("op_1158")]; |
| tensor<int32, [4]> concat_70x = const()[name = tensor<string, []>("concat_70x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_11 = reshape(shape = concat_70x, x = var_1158)[name = tensor<string, []>("valid_q_11")]; |
| tensor<fp32, []> var_1162 = const()[name = tensor<string, []>("op_1162"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_23 = select(a = out_21, b = var_1162, cond = valid_q_11)[name = tensor<string, []>("out_23")]; |
| tensor<int32, [4]> var_1164_perm_0 = const()[name = tensor<string, []>("op_1164_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_71x = const()[name = tensor<string, []>("concat_71x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_1164 = transpose(perm = var_1164_perm_0, x = out_23)[name = tensor<string, []>("transpose_67")]; |
| tensor<fp32, [1, ?, 256]> input_65 = reshape(shape = concat_71x, x = var_1164)[name = tensor<string, []>("input_65")]; |
| tensor<fp32, [1, ?, 256]> x_45 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_1_self_attn_out_proj_weight, x = input_65)[name = tensor<string, []>("linear_24")]; |
| tensor<fp32, [1, ?, 256]> var_1170 = mul(x = encoder_3_transformer_layers_1_layer_scale_1_scale, y = x_45)[name = tensor<string, []>("op_1170")]; |
| tensor<fp32, [1, ?, 256]> input_67 = add(x = input_63, y = var_1170)[name = tensor<string, []>("input_67")]; |
| tensor<int32, [1]> input_69_axes_0 = const()[name = tensor<string, []>("input_69_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_69 = layer_norm(axes = input_69_axes_0, beta = encoder_3_transformer_layers_1_norm2_bias, epsilon = var_898, gamma = encoder_3_transformer_layers_1_norm2_weight, x = input_67)[name = tensor<string, []>("input_69")]; |
| tensor<fp32, [1, ?, 1024]> input_71 = linear(bias = linear_3_bias_0, weight = encoder_3_transformer_layers_1_ffn_0_weight, x = input_69)[name = tensor<string, []>("linear_25")]; |
| tensor<string, []> input_73_mode_0 = const()[name = tensor<string, []>("input_73_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_73 = gelu(mode = input_73_mode_0, x = input_71)[name = tensor<string, []>("input_73")]; |
| tensor<fp32, [1, ?, 256]> x_47 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_1_ffn_2_weight, x = input_73)[name = tensor<string, []>("linear_26")]; |
| tensor<fp32, [1, ?, 256]> var_1185 = mul(x = encoder_3_transformer_layers_1_layer_scale_2_scale, y = x_47)[name = tensor<string, []>("op_1185")]; |
| tensor<fp32, [1, ?, 256]> input_75 = add(x = input_67, y = var_1185)[name = tensor<string, []>("input_75")]; |
| tensor<fp32, [1, ?, 384]> var_1188 = linear(bias = linear_17_bias_0, weight = encoder_3_output_proj_weight, x = input_75)[name = tensor<string, []>("linear_27")]; |
| tensor<int32, [3]> x_49_perm_0 = const()[name = tensor<string, []>("x_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [4]> var_1199 = const()[name = tensor<string, []>("op_1199"), val = tensor<int32, [4]>([1, 384, -1, 2])]; |
| tensor<fp32, [1, 384, ?]> x_49 = transpose(perm = x_49_perm_0, x = var_1188)[name = tensor<string, []>("transpose_66")]; |
| tensor<fp32, [1, 384, ?, 2]> var_1200 = reshape(shape = var_1199, x = x_49)[name = tensor<string, []>("op_1200")]; |
| tensor<int32, [4]> var_1201 = const()[name = tensor<string, []>("op_1201"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| tensor<int32, [3]> var_1206 = const()[name = tensor<string, []>("op_1206"), val = tensor<int32, [3]>([1, 768, -1])]; |
| tensor<fp32, [1, 384, 2, ?]> var_1202 = transpose(perm = var_1201, x = var_1200)[name = tensor<string, []>("transpose_65")]; |
| tensor<fp32, [1, 768, ?]> x_51 = reshape(shape = var_1206, x = var_1202)[name = tensor<string, []>("x_51")]; |
| tensor<fp32, []> var_1190_promoted = const()[name = tensor<string, []>("op_1190_promoted"), val = tensor<fp32, []>(0x1p+1)]; |
| tensor<fp32, [1]> floor_div_2 = floor_div(x = floor_div_1, y = var_1190_promoted)[name = tensor<string, []>("floor_div_2")]; |
| tensor<int32, []> var_1213 = const()[name = tensor<string, []>("op_1213"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> var_1220 = const()[name = tensor<string, []>("op_1220"), val = tensor<int32, []>(800)]; |
| tensor<fp32, []> var_1227 = const()[name = tensor<string, []>("op_1227"), val = tensor<fp32, []>(0x1.4f8b58p-17)]; |
| tensor<int32, []> var_1231 = const()[name = tensor<string, []>("op_1231"), val = tensor<int32, []>(1)]; |
| tensor<int32, [3]> input_77_perm_0 = const()[name = tensor<string, []>("input_77_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, ?, 768]> input_77 = transpose(perm = input_77_perm_0, x = x_51)[name = tensor<string, []>("transpose_64")]; |
| tensor<fp32, [1, ?, 256]> x_53 = linear(bias = linear_0_bias_0, weight = encoder_5_input_proj_weight, x = input_77)[name = tensor<string, []>("linear_28")]; |
| tensor<int32, [1]> x_55_axes_0 = const()[name = tensor<string, []>("x_55_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_55 = layer_norm(axes = x_55_axes_0, beta = encoder_5_transformer_layers_0_norm1_bias, epsilon = var_1227, gamma = encoder_5_transformer_layers_0_norm1_weight, x = x_53)[name = tensor<string, []>("x_55")]; |
| tensor<int32, [3]> var_1255_shape = shape(x = x_55)[name = tensor<string, []>("op_1255_shape")]; |
| tensor<int32, []> gather_68_batch_dims_0 = const()[name = tensor<string, []>("gather_68_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_68_validate_indices_0 = const()[name = tensor<string, []>("gather_68_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_28 = const()[name = tensor<string, []>("select_28"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_68_axis_1 = const()[name = tensor<string, []>("gather_68_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_68 = gather(axis = gather_68_axis_1, batch_dims = gather_68_batch_dims_0, indices = select_28, validate_indices = gather_68_validate_indices_0, x = var_1255_shape)[name = tensor<string, []>("gather_68")]; |
| tensor<fp32, [1, ?, 768]> projected_25 = linear(bias = linear_1_bias_0, weight = encoder_5_transformer_layers_0_self_attn_in_proj_weight, x = x_55)[name = tensor<string, []>("linear_29")]; |
| tensor<int32, [5]> concat_72x = const()[name = tensor<string, []>("concat_72x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_1262 = reshape(shape = concat_72x, x = projected_25)[name = tensor<string, []>("op_1262")]; |
| tensor<int32, [5]> var_1263 = const()[name = tensor<string, []>("op_1263"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_25_begin_0 = const()[name = tensor<string, []>("q_25_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_25_end_0 = const()[name = tensor<string, []>("q_25_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_25_end_mask_0 = const()[name = tensor<string, []>("q_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_25_squeeze_mask_0 = const()[name = tensor<string, []>("q_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_27 = transpose(perm = var_1263, x = var_1262)[name = tensor<string, []>("transpose_63")]; |
| tensor<fp32, [1, 4, ?, 64]> q_25 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = projected_27)[name = tensor<string, []>("q_25")]; |
| tensor<int32, [5]> k_25_begin_0 = const()[name = tensor<string, []>("k_25_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_25_end_0 = const()[name = tensor<string, []>("k_25_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_25_end_mask_0 = const()[name = tensor<string, []>("k_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_25_squeeze_mask_0 = const()[name = tensor<string, []>("k_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_25 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = projected_27)[name = tensor<string, []>("k_25")]; |
| tensor<int32, [5]> v_13_begin_0 = const()[name = tensor<string, []>("v_13_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_13_end_0 = const()[name = tensor<string, []>("v_13_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_13_end_mask_0 = const()[name = tensor<string, []>("v_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_13_squeeze_mask_0 = const()[name = tensor<string, []>("v_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_13 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = projected_27)[name = tensor<string, []>("v_13")]; |
| tensor<fp32, [32]> freqs_13 = const()[name = tensor<string, []>("freqs_13"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44432256)))]; |
| tensor<int32, [4]> var_1275_shape = shape(x = q_25)[name = tensor<string, []>("op_1275_shape")]; |
| tensor<int32, []> gather_72_batch_dims_0 = const()[name = tensor<string, []>("gather_72_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_72_validate_indices_0 = const()[name = tensor<string, []>("gather_72_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_29 = const()[name = tensor<string, []>("select_29"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_72_axis_1 = const()[name = tensor<string, []>("gather_72_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_72 = gather(axis = gather_72_axis_1, batch_dims = gather_72_batch_dims_0, indices = select_29, validate_indices = gather_72_validate_indices_0, x = var_1275_shape)[name = tensor<string, []>("gather_72")]; |
| tensor<fp32, [1, 1]> var_1278 = const()[name = tensor<string, []>("op_1278"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_57 = const()[name = tensor<string, []>("const_57"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_58 = const()[name = tensor<string, []>("const_58"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_1279 = range_1d(end = gather_72, start = const_57, step = const_58)[name = tensor<string, []>("op_1279")]; |
| tensor<string, []> var_1279_promoted_dtype_0 = const()[name = tensor<string, []>("op_1279_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_1279_promoted = cast(dtype = var_1279_promoted_dtype_0, x = var_1279)[name = tensor<string, []>("cast_446")]; |
| tensor<fp32, [1, ?]> ts_25 = add(x = var_1278, y = var_1279_promoted)[name = tensor<string, []>("ts_25")]; |
| tensor<int32, [4]> var_1282 = const()[name = tensor<string, []>("op_1282"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_27 = reshape(shape = var_1282, x = ts_25)[name = tensor<string, []>("ts_27")]; |
| tensor<int32, [5]> concat_73x = const()[name = tensor<string, []>("concat_73x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_13 = reshape(shape = concat_73x, x = q_25)[name = tensor<string, []>("q2_13")]; |
| tensor<int32, [5]> concat_74x = const()[name = tensor<string, []>("concat_74x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_13 = reshape(shape = concat_74x, x = k_25)[name = tensor<string, []>("k2_13")]; |
| tensor<int32, [5]> qr_13_begin_0 = const()[name = tensor<string, []>("qr_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_13_end_0 = const()[name = tensor<string, []>("qr_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_13_end_mask_0 = const()[name = tensor<string, []>("qr_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_13_squeeze_mask_0 = const()[name = tensor<string, []>("qr_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_13 = slice_by_index(begin = qr_13_begin_0, end = qr_13_end_0, end_mask = qr_13_end_mask_0, squeeze_mask = qr_13_squeeze_mask_0, x = q2_13)[name = tensor<string, []>("qr_13")]; |
| tensor<int32, [5]> qi_13_begin_0 = const()[name = tensor<string, []>("qi_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_13_end_0 = const()[name = tensor<string, []>("qi_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_13_end_mask_0 = const()[name = tensor<string, []>("qi_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_13_squeeze_mask_0 = const()[name = tensor<string, []>("qi_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_13 = slice_by_index(begin = qi_13_begin_0, end = qi_13_end_0, end_mask = qi_13_end_mask_0, squeeze_mask = qi_13_squeeze_mask_0, x = q2_13)[name = tensor<string, []>("qi_13")]; |
| tensor<int32, [5]> kr_13_begin_0 = const()[name = tensor<string, []>("kr_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_13_end_0 = const()[name = tensor<string, []>("kr_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_13_end_mask_0 = const()[name = tensor<string, []>("kr_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_13_squeeze_mask_0 = const()[name = tensor<string, []>("kr_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_13 = slice_by_index(begin = kr_13_begin_0, end = kr_13_end_0, end_mask = kr_13_end_mask_0, squeeze_mask = kr_13_squeeze_mask_0, x = k2_13)[name = tensor<string, []>("kr_13")]; |
| tensor<int32, [5]> ki_13_begin_0 = const()[name = tensor<string, []>("ki_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_13_end_0 = const()[name = tensor<string, []>("ki_13_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_13_end_mask_0 = const()[name = tensor<string, []>("ki_13_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_13_squeeze_mask_0 = const()[name = tensor<string, []>("ki_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_13 = slice_by_index(begin = ki_13_begin_0, end = ki_13_end_0, end_mask = ki_13_end_mask_0, squeeze_mask = ki_13_squeeze_mask_0, x = k2_13)[name = tensor<string, []>("ki_13")]; |
| tensor<fp32, [1, 1, ?, 32]> var_1295 = mul(x = freqs_13, y = ts_27)[name = tensor<string, []>("op_1295")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_13 = cos(x = var_1295)[name = tensor<string, []>("rotr_13")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_13 = sin(x = var_1295)[name = tensor<string, []>("roti_13")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1299 = mul(x = qr_13, y = rotr_13)[name = tensor<string, []>("op_1299")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1300 = mul(x = qi_13, y = roti_13)[name = tensor<string, []>("op_1300")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1301 = sub(x = var_1299, y = var_1300)[name = tensor<string, []>("op_1301")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1302 = mul(x = qr_13, y = roti_13)[name = tensor<string, []>("op_1302")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1303 = mul(x = qi_13, y = rotr_13)[name = tensor<string, []>("op_1303")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1304 = add(x = var_1302, y = var_1303)[name = tensor<string, []>("op_1304")]; |
| tensor<int32, []> qo_13_axis_0 = const()[name = tensor<string, []>("qo_13_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_13 = stack(axis = qo_13_axis_0, values = (var_1301, var_1304))[name = tensor<string, []>("qo_13")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1307 = mul(x = kr_13, y = rotr_13)[name = tensor<string, []>("op_1307")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1308 = mul(x = ki_13, y = roti_13)[name = tensor<string, []>("op_1308")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1309 = sub(x = var_1307, y = var_1308)[name = tensor<string, []>("op_1309")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1310 = mul(x = kr_13, y = roti_13)[name = tensor<string, []>("op_1310")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1311 = mul(x = ki_13, y = rotr_13)[name = tensor<string, []>("op_1311")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1312 = add(x = var_1310, y = var_1311)[name = tensor<string, []>("op_1312")]; |
| tensor<int32, []> ko_13_axis_0 = const()[name = tensor<string, []>("ko_13_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_13 = stack(axis = ko_13_axis_0, values = (var_1309, var_1312))[name = tensor<string, []>("ko_13")]; |
| tensor<int32, [4]> concat_75x = const()[name = tensor<string, []>("concat_75x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_27 = reshape(shape = concat_75x, x = qo_13)[name = tensor<string, []>("q_27")]; |
| tensor<int32, [4]> concat_76x = const()[name = tensor<string, []>("concat_76x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_27 = reshape(shape = concat_76x, x = ko_13)[name = tensor<string, []>("k_27")]; |
| tensor<int32, []> const_59 = const()[name = tensor<string, []>("const_59"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_60 = const()[name = tensor<string, []>("const_60"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_13 = range_1d(end = gather_68, start = const_59, step = const_60)[name = tensor<string, []>("positions_13")]; |
| tensor<int32, [3]> concat_77x = const()[name = tensor<string, []>("concat_77x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_1324 = reshape(shape = concat_77x, x = positions_13)[name = tensor<string, []>("op_1324")]; |
| tensor<int32, [3]> var_1325 = const()[name = tensor<string, []>("op_1325"), val = tensor<int32, [3]>([-1, 1, 1])]; |
| tensor<fp32, [1, 1, 1]> var_1326 = reshape(shape = var_1325, x = floor_div_2)[name = tensor<string, []>("op_1326")]; |
| tensor<string, []> var_1324_promoted_dtype_0 = const()[name = tensor<string, []>("op_1324_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_1324_promoted = cast(dtype = var_1324_promoted_dtype_0, x = var_1324)[name = tensor<string, []>("cast_445")]; |
| tensor<bool, [1, 1, ?]> valid_k_13 = less(x = var_1324_promoted, y = var_1326)[name = tensor<string, []>("valid_k_13")]; |
| tensor<int32, [3]> concat_78x = const()[name = tensor<string, []>("concat_78x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_1329 = reshape(shape = concat_78x, x = positions_13)[name = tensor<string, []>("op_1329")]; |
| tensor<int32, [1, ?, ?]> delta_13 = sub(x = var_1329, y = var_1324)[name = tensor<string, []>("delta_13")]; |
| tensor<int32, []> concat_80_axis_0 = const()[name = tensor<string, []>("concat_80_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_80_interleave_0 = const()[name = tensor<string, []>("concat_80_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_80 = concat(axis = concat_80_axis_0, interleave = concat_80_interleave_0, values = (var_1231, gather_68, gather_68))[name = tensor<string, []>("concat_80")]; |
| tensor<fp32, []> fill_6_value_0 = const()[name = tensor<string, []>("fill_6_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_6 = fill(shape = concat_80, value = fill_6_value_0)[name = tensor<string, []>("fill_6")]; |
| tensor<string, []> attn_bias_49_dtype_0 = const()[name = tensor<string, []>("attn_bias_49_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_1335 = greater_equal(x = delta_13, y = var_1213)[name = tensor<string, []>("op_1335")]; |
| tensor<bool, [1, ?, ?]> attn_bias_49 = cast(dtype = attn_bias_49_dtype_0, x = fill_6)[name = tensor<string, []>("cast_444")]; |
| tensor<bool, [1, ?, ?]> attn_bias_51 = logical_and(x = attn_bias_49, y = var_1335)[name = tensor<string, []>("attn_bias_51")]; |
| tensor<bool, [1, ?, ?]> var_1337 = less(x = delta_13, y = var_1220)[name = tensor<string, []>("op_1337")]; |
| tensor<bool, [1, ?, ?]> attn_bias_53 = logical_and(x = attn_bias_51, y = var_1337)[name = tensor<string, []>("attn_bias_53")]; |
| tensor<bool, [1, ?, ?]> var_1339 = logical_and(x = attn_bias_53, y = valid_k_13)[name = tensor<string, []>("op_1339")]; |
| tensor<int32, [1]> var_1341_axes_0 = const()[name = tensor<string, []>("op_1341_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_1341 = expand_dims(axes = var_1341_axes_0, x = var_1339)[name = tensor<string, []>("op_1341")]; |
| tensor<string, []> cast_156_dtype_0 = const()[name = tensor<string, []>("cast_156_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_6_x_0 = const()[name = tensor<string, []>("sub_6_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_156 = cast(dtype = cast_156_dtype_0, x = var_1341)[name = tensor<string, []>("cast_443")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_6 = sub(x = sub_6_x_0, y = cast_156)[name = tensor<string, []>("sub_6")]; |
| tensor<fp32, []> mul_12_x_0 = const()[name = tensor<string, []>("mul_12_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_12 = mul(x = mul_12_x_0, y = sub_6)[name = tensor<string, []>("mul_12")]; |
| tensor<fp32, []> mul_13_y_0 = const()[name = tensor<string, []>("mul_13_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_13 = mul(x = q_27, y = mul_13_y_0)[name = tensor<string, []>("mul_13")]; |
| tensor<bool, []> matmul_6_transpose_y_0 = const()[name = tensor<string, []>("matmul_6_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_6_transpose_x_0 = const()[name = tensor<string, []>("matmul_6_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_6 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = mul_13, y = k_27)[name = tensor<string, []>("matmul_6")]; |
| tensor<fp32, [1, 4, ?, ?]> add_6 = add(x = matmul_6, y = mul_12)[name = tensor<string, []>("add_6")]; |
| tensor<int32, []> softmax_6_axis_0 = const()[name = tensor<string, []>("softmax_6_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_6 = softmax(axis = softmax_6_axis_0, x = add_6)[name = tensor<string, []>("softmax_6")]; |
| tensor<bool, []> out_25_transpose_x_0 = const()[name = tensor<string, []>("out_25_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> out_25_transpose_y_0 = const()[name = tensor<string, []>("out_25_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, 64]> out_25 = matmul(transpose_x = out_25_transpose_x_0, transpose_y = out_25_transpose_y_0, x = softmax_6, y = v_13)[name = tensor<string, []>("out_25")]; |
| tensor<int32, [2]> concat_81x = const()[name = tensor<string, []>("concat_81x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_1347 = reshape(shape = concat_81x, x = positions_13)[name = tensor<string, []>("op_1347")]; |
| tensor<int32, [2]> var_1348 = const()[name = tensor<string, []>("op_1348"), val = tensor<int32, [2]>([-1, 1])]; |
| tensor<fp32, [1, 1]> var_1349 = reshape(shape = var_1348, x = floor_div_2)[name = tensor<string, []>("op_1349")]; |
| tensor<string, []> var_1347_promoted_dtype_0 = const()[name = tensor<string, []>("op_1347_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_1347_promoted = cast(dtype = var_1347_promoted_dtype_0, x = var_1347)[name = tensor<string, []>("cast_442")]; |
| tensor<bool, [1, ?]> var_1350 = less(x = var_1347_promoted, y = var_1349)[name = tensor<string, []>("op_1350")]; |
| tensor<int32, [4]> concat_82x = const()[name = tensor<string, []>("concat_82x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_13 = reshape(shape = concat_82x, x = var_1350)[name = tensor<string, []>("valid_q_13")]; |
| tensor<fp32, []> var_1354 = const()[name = tensor<string, []>("op_1354"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_27 = select(a = out_25, b = var_1354, cond = valid_q_13)[name = tensor<string, []>("out_27")]; |
| tensor<int32, [4]> var_1356_perm_0 = const()[name = tensor<string, []>("op_1356_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_83x = const()[name = tensor<string, []>("concat_83x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_1356 = transpose(perm = var_1356_perm_0, x = out_27)[name = tensor<string, []>("transpose_62")]; |
| tensor<fp32, [1, ?, 256]> input_79 = reshape(shape = concat_83x, x = var_1356)[name = tensor<string, []>("input_79")]; |
| tensor<fp32, [1, ?, 256]> x_57 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_0_self_attn_out_proj_weight, x = input_79)[name = tensor<string, []>("linear_30")]; |
| tensor<fp32, [1, ?, 256]> var_1362 = mul(x = encoder_5_transformer_layers_0_layer_scale_1_scale, y = x_57)[name = tensor<string, []>("op_1362")]; |
| tensor<fp32, [1, ?, 256]> input_81 = add(x = x_53, y = var_1362)[name = tensor<string, []>("input_81")]; |
| tensor<int32, [1]> input_83_axes_0 = const()[name = tensor<string, []>("input_83_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_5_transformer_layers_0_norm2_bias, epsilon = var_1227, gamma = encoder_5_transformer_layers_0_norm2_weight, x = input_81)[name = tensor<string, []>("input_83")]; |
| tensor<fp32, [1, ?, 1024]> input_85 = linear(bias = linear_3_bias_0, weight = encoder_5_transformer_layers_0_ffn_0_weight, x = input_83)[name = tensor<string, []>("linear_31")]; |
| tensor<string, []> input_87_mode_0 = const()[name = tensor<string, []>("input_87_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_87 = gelu(mode = input_87_mode_0, x = input_85)[name = tensor<string, []>("input_87")]; |
| tensor<fp32, [1, ?, 256]> x_59 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_0_ffn_2_weight, x = input_87)[name = tensor<string, []>("linear_32")]; |
| tensor<fp32, [1, ?, 256]> var_1377 = mul(x = encoder_5_transformer_layers_0_layer_scale_2_scale, y = x_59)[name = tensor<string, []>("op_1377")]; |
| tensor<fp32, [1, ?, 256]> input_89 = add(x = input_81, y = var_1377)[name = tensor<string, []>("input_89")]; |
| tensor<int32, [1]> x_61_axes_0 = const()[name = tensor<string, []>("x_61_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_61 = layer_norm(axes = x_61_axes_0, beta = encoder_5_transformer_layers_1_norm1_bias, epsilon = var_1227, gamma = encoder_5_transformer_layers_1_norm1_weight, x = input_89)[name = tensor<string, []>("x_61")]; |
| tensor<int32, [3]> var_1392_shape = shape(x = x_61)[name = tensor<string, []>("op_1392_shape")]; |
| tensor<int32, []> gather_78_batch_dims_0 = const()[name = tensor<string, []>("gather_78_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_78_validate_indices_0 = const()[name = tensor<string, []>("gather_78_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_30 = const()[name = tensor<string, []>("select_30"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_78_axis_1 = const()[name = tensor<string, []>("gather_78_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_78 = gather(axis = gather_78_axis_1, batch_dims = gather_78_batch_dims_0, indices = select_30, validate_indices = gather_78_validate_indices_0, x = var_1392_shape)[name = tensor<string, []>("gather_78")]; |
| tensor<fp32, [1, ?, 768]> projected_29 = linear(bias = linear_1_bias_0, weight = encoder_5_transformer_layers_1_self_attn_in_proj_weight, x = x_61)[name = tensor<string, []>("linear_33")]; |
| tensor<int32, [5]> concat_84x = const()[name = tensor<string, []>("concat_84x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_1399 = reshape(shape = concat_84x, x = projected_29)[name = tensor<string, []>("op_1399")]; |
| tensor<int32, [5]> var_1400 = const()[name = tensor<string, []>("op_1400"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_29_begin_0 = const()[name = tensor<string, []>("q_29_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_29_end_0 = const()[name = tensor<string, []>("q_29_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_29_end_mask_0 = const()[name = tensor<string, []>("q_29_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_29_squeeze_mask_0 = const()[name = tensor<string, []>("q_29_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_31 = transpose(perm = var_1400, x = var_1399)[name = tensor<string, []>("transpose_61")]; |
| tensor<fp32, [1, 4, ?, 64]> q_29 = slice_by_index(begin = q_29_begin_0, end = q_29_end_0, end_mask = q_29_end_mask_0, squeeze_mask = q_29_squeeze_mask_0, x = projected_31)[name = tensor<string, []>("q_29")]; |
| tensor<int32, [5]> k_29_begin_0 = const()[name = tensor<string, []>("k_29_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_29_end_0 = const()[name = tensor<string, []>("k_29_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_29_end_mask_0 = const()[name = tensor<string, []>("k_29_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_29_squeeze_mask_0 = const()[name = tensor<string, []>("k_29_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_29 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = projected_31)[name = tensor<string, []>("k_29")]; |
| tensor<int32, [5]> v_15_begin_0 = const()[name = tensor<string, []>("v_15_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_15_end_0 = const()[name = tensor<string, []>("v_15_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_15_end_mask_0 = const()[name = tensor<string, []>("v_15_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_15_squeeze_mask_0 = const()[name = tensor<string, []>("v_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_15 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = projected_31)[name = tensor<string, []>("v_15")]; |
| tensor<fp32, [32]> freqs_15 = const()[name = tensor<string, []>("freqs_15"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44432448)))]; |
| tensor<int32, [4]> var_1412_shape = shape(x = q_29)[name = tensor<string, []>("op_1412_shape")]; |
| tensor<int32, []> gather_82_batch_dims_0 = const()[name = tensor<string, []>("gather_82_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_82_validate_indices_0 = const()[name = tensor<string, []>("gather_82_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_31 = const()[name = tensor<string, []>("select_31"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_82_axis_1 = const()[name = tensor<string, []>("gather_82_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_82 = gather(axis = gather_82_axis_1, batch_dims = gather_82_batch_dims_0, indices = select_31, validate_indices = gather_82_validate_indices_0, x = var_1412_shape)[name = tensor<string, []>("gather_82")]; |
| tensor<fp32, [1, 1]> var_1415 = const()[name = tensor<string, []>("op_1415"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_66 = const()[name = tensor<string, []>("const_66"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_67 = const()[name = tensor<string, []>("const_67"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_1416 = range_1d(end = gather_82, start = const_66, step = const_67)[name = tensor<string, []>("op_1416")]; |
| tensor<string, []> var_1416_promoted_dtype_0 = const()[name = tensor<string, []>("op_1416_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_1416_promoted = cast(dtype = var_1416_promoted_dtype_0, x = var_1416)[name = tensor<string, []>("cast_441")]; |
| tensor<fp32, [1, ?]> ts_29 = add(x = var_1415, y = var_1416_promoted)[name = tensor<string, []>("ts_29")]; |
| tensor<int32, [4]> var_1419 = const()[name = tensor<string, []>("op_1419"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_31 = reshape(shape = var_1419, x = ts_29)[name = tensor<string, []>("ts_31")]; |
| tensor<int32, [5]> concat_85x = const()[name = tensor<string, []>("concat_85x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_15 = reshape(shape = concat_85x, x = q_29)[name = tensor<string, []>("q2_15")]; |
| tensor<int32, [5]> concat_86x = const()[name = tensor<string, []>("concat_86x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_15 = reshape(shape = concat_86x, x = k_29)[name = tensor<string, []>("k2_15")]; |
| tensor<int32, [5]> qr_15_begin_0 = const()[name = tensor<string, []>("qr_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_15_end_0 = const()[name = tensor<string, []>("qr_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_15_end_mask_0 = const()[name = tensor<string, []>("qr_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_15_squeeze_mask_0 = const()[name = tensor<string, []>("qr_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_15 = slice_by_index(begin = qr_15_begin_0, end = qr_15_end_0, end_mask = qr_15_end_mask_0, squeeze_mask = qr_15_squeeze_mask_0, x = q2_15)[name = tensor<string, []>("qr_15")]; |
| tensor<int32, [5]> qi_15_begin_0 = const()[name = tensor<string, []>("qi_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_15_end_0 = const()[name = tensor<string, []>("qi_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_15_end_mask_0 = const()[name = tensor<string, []>("qi_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_15_squeeze_mask_0 = const()[name = tensor<string, []>("qi_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_15 = slice_by_index(begin = qi_15_begin_0, end = qi_15_end_0, end_mask = qi_15_end_mask_0, squeeze_mask = qi_15_squeeze_mask_0, x = q2_15)[name = tensor<string, []>("qi_15")]; |
| tensor<int32, [5]> kr_15_begin_0 = const()[name = tensor<string, []>("kr_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_15_end_0 = const()[name = tensor<string, []>("kr_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_15_end_mask_0 = const()[name = tensor<string, []>("kr_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_15_squeeze_mask_0 = const()[name = tensor<string, []>("kr_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_15 = slice_by_index(begin = kr_15_begin_0, end = kr_15_end_0, end_mask = kr_15_end_mask_0, squeeze_mask = kr_15_squeeze_mask_0, x = k2_15)[name = tensor<string, []>("kr_15")]; |
| tensor<int32, [5]> ki_15_begin_0 = const()[name = tensor<string, []>("ki_15_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_15_end_0 = const()[name = tensor<string, []>("ki_15_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_15_end_mask_0 = const()[name = tensor<string, []>("ki_15_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_15_squeeze_mask_0 = const()[name = tensor<string, []>("ki_15_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_15 = slice_by_index(begin = ki_15_begin_0, end = ki_15_end_0, end_mask = ki_15_end_mask_0, squeeze_mask = ki_15_squeeze_mask_0, x = k2_15)[name = tensor<string, []>("ki_15")]; |
| tensor<fp32, [1, 1, ?, 32]> var_1432 = mul(x = freqs_15, y = ts_31)[name = tensor<string, []>("op_1432")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_15 = cos(x = var_1432)[name = tensor<string, []>("rotr_15")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_15 = sin(x = var_1432)[name = tensor<string, []>("roti_15")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1436 = mul(x = qr_15, y = rotr_15)[name = tensor<string, []>("op_1436")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1437 = mul(x = qi_15, y = roti_15)[name = tensor<string, []>("op_1437")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1438 = sub(x = var_1436, y = var_1437)[name = tensor<string, []>("op_1438")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1439 = mul(x = qr_15, y = roti_15)[name = tensor<string, []>("op_1439")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1440 = mul(x = qi_15, y = rotr_15)[name = tensor<string, []>("op_1440")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1441 = add(x = var_1439, y = var_1440)[name = tensor<string, []>("op_1441")]; |
| tensor<int32, []> qo_15_axis_0 = const()[name = tensor<string, []>("qo_15_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_15 = stack(axis = qo_15_axis_0, values = (var_1438, var_1441))[name = tensor<string, []>("qo_15")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1444 = mul(x = kr_15, y = rotr_15)[name = tensor<string, []>("op_1444")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1445 = mul(x = ki_15, y = roti_15)[name = tensor<string, []>("op_1445")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1446 = sub(x = var_1444, y = var_1445)[name = tensor<string, []>("op_1446")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1447 = mul(x = kr_15, y = roti_15)[name = tensor<string, []>("op_1447")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1448 = mul(x = ki_15, y = rotr_15)[name = tensor<string, []>("op_1448")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1449 = add(x = var_1447, y = var_1448)[name = tensor<string, []>("op_1449")]; |
| tensor<int32, []> ko_15_axis_0 = const()[name = tensor<string, []>("ko_15_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_15 = stack(axis = ko_15_axis_0, values = (var_1446, var_1449))[name = tensor<string, []>("ko_15")]; |
| tensor<int32, [4]> concat_87x = const()[name = tensor<string, []>("concat_87x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_31 = reshape(shape = concat_87x, x = qo_15)[name = tensor<string, []>("q_31")]; |
| tensor<int32, [4]> concat_88x = const()[name = tensor<string, []>("concat_88x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_31 = reshape(shape = concat_88x, x = ko_15)[name = tensor<string, []>("k_31")]; |
| tensor<int32, []> const_68 = const()[name = tensor<string, []>("const_68"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_69 = const()[name = tensor<string, []>("const_69"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_15 = range_1d(end = gather_78, start = const_68, step = const_69)[name = tensor<string, []>("positions_15")]; |
| tensor<int32, [3]> concat_89x = const()[name = tensor<string, []>("concat_89x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_1461 = reshape(shape = concat_89x, x = positions_15)[name = tensor<string, []>("op_1461")]; |
| tensor<string, []> var_1461_promoted_dtype_0 = const()[name = tensor<string, []>("op_1461_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_1461_promoted = cast(dtype = var_1461_promoted_dtype_0, x = var_1461)[name = tensor<string, []>("cast_440")]; |
| tensor<bool, [1, 1, ?]> valid_k_15 = less(x = var_1461_promoted, y = var_1326)[name = tensor<string, []>("valid_k_15")]; |
| tensor<int32, [3]> concat_90x = const()[name = tensor<string, []>("concat_90x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_1466 = reshape(shape = concat_90x, x = positions_15)[name = tensor<string, []>("op_1466")]; |
| tensor<int32, [1, ?, ?]> delta_15 = sub(x = var_1466, y = var_1461)[name = tensor<string, []>("delta_15")]; |
| tensor<int32, []> concat_92_axis_0 = const()[name = tensor<string, []>("concat_92_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_92_interleave_0 = const()[name = tensor<string, []>("concat_92_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_92 = concat(axis = concat_92_axis_0, interleave = concat_92_interleave_0, values = (var_1231, gather_78, gather_78))[name = tensor<string, []>("concat_92")]; |
| tensor<fp32, []> fill_7_value_0 = const()[name = tensor<string, []>("fill_7_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_7 = fill(shape = concat_92, value = fill_7_value_0)[name = tensor<string, []>("fill_7")]; |
| tensor<string, []> attn_bias_57_dtype_0 = const()[name = tensor<string, []>("attn_bias_57_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_1472 = greater_equal(x = delta_15, y = var_1213)[name = tensor<string, []>("op_1472")]; |
| tensor<bool, [1, ?, ?]> attn_bias_57 = cast(dtype = attn_bias_57_dtype_0, x = fill_7)[name = tensor<string, []>("cast_439")]; |
| tensor<bool, [1, ?, ?]> attn_bias_59 = logical_and(x = attn_bias_57, y = var_1472)[name = tensor<string, []>("attn_bias_59")]; |
| tensor<bool, [1, ?, ?]> var_1474 = less(x = delta_15, y = var_1220)[name = tensor<string, []>("op_1474")]; |
| tensor<bool, [1, ?, ?]> attn_bias_61 = logical_and(x = attn_bias_59, y = var_1474)[name = tensor<string, []>("attn_bias_61")]; |
| tensor<bool, [1, ?, ?]> var_1476 = logical_and(x = attn_bias_61, y = valid_k_15)[name = tensor<string, []>("op_1476")]; |
| tensor<int32, [1]> var_1478_axes_0 = const()[name = tensor<string, []>("op_1478_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_1478 = expand_dims(axes = var_1478_axes_0, x = var_1476)[name = tensor<string, []>("op_1478")]; |
| tensor<string, []> cast_178_dtype_0 = const()[name = tensor<string, []>("cast_178_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_7_x_0 = const()[name = tensor<string, []>("sub_7_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_178 = cast(dtype = cast_178_dtype_0, x = var_1478)[name = tensor<string, []>("cast_438")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_7 = sub(x = sub_7_x_0, y = cast_178)[name = tensor<string, []>("sub_7")]; |
| tensor<fp32, []> mul_14_x_0 = const()[name = tensor<string, []>("mul_14_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_14 = mul(x = mul_14_x_0, y = sub_7)[name = tensor<string, []>("mul_14")]; |
| tensor<fp32, []> mul_15_y_0 = const()[name = tensor<string, []>("mul_15_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_15 = mul(x = q_31, y = mul_15_y_0)[name = tensor<string, []>("mul_15")]; |
| tensor<bool, []> matmul_7_transpose_y_0 = const()[name = tensor<string, []>("matmul_7_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_7_transpose_x_0 = const()[name = tensor<string, []>("matmul_7_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_7 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = mul_15, y = k_31)[name = tensor<string, []>("matmul_7")]; |
| tensor<fp32, [1, 4, ?, ?]> add_7 = add(x = matmul_7, y = mul_14)[name = tensor<string, []>("add_7")]; |
| tensor<int32, []> softmax_7_axis_0 = const()[name = tensor<string, []>("softmax_7_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_7 = softmax(axis = softmax_7_axis_0, x = add_7)[name = tensor<string, []>("softmax_7")]; |
| tensor<bool, []> out_29_transpose_x_0 = const()[name = tensor<string, []>("out_29_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> out_29_transpose_y_0 = const()[name = tensor<string, []>("out_29_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, 64]> out_29 = matmul(transpose_x = out_29_transpose_x_0, transpose_y = out_29_transpose_y_0, x = softmax_7, y = v_15)[name = tensor<string, []>("out_29")]; |
| tensor<int32, [2]> concat_93x = const()[name = tensor<string, []>("concat_93x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_1484 = reshape(shape = concat_93x, x = positions_15)[name = tensor<string, []>("op_1484")]; |
| tensor<string, []> var_1484_promoted_dtype_0 = const()[name = tensor<string, []>("op_1484_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_1484_promoted = cast(dtype = var_1484_promoted_dtype_0, x = var_1484)[name = tensor<string, []>("cast_437")]; |
| tensor<bool, [1, ?]> var_1487 = less(x = var_1484_promoted, y = var_1349)[name = tensor<string, []>("op_1487")]; |
| tensor<int32, [4]> concat_94x = const()[name = tensor<string, []>("concat_94x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_15 = reshape(shape = concat_94x, x = var_1487)[name = tensor<string, []>("valid_q_15")]; |
| tensor<fp32, []> var_1491 = const()[name = tensor<string, []>("op_1491"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_31 = select(a = out_29, b = var_1491, cond = valid_q_15)[name = tensor<string, []>("out_31")]; |
| tensor<int32, [4]> var_1493_perm_0 = const()[name = tensor<string, []>("op_1493_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_95x = const()[name = tensor<string, []>("concat_95x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_1493 = transpose(perm = var_1493_perm_0, x = out_31)[name = tensor<string, []>("transpose_60")]; |
| tensor<fp32, [1, ?, 256]> input_91 = reshape(shape = concat_95x, x = var_1493)[name = tensor<string, []>("input_91")]; |
| tensor<fp32, [1, ?, 256]> x_63 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_1_self_attn_out_proj_weight, x = input_91)[name = tensor<string, []>("linear_34")]; |
| tensor<fp32, [1, ?, 256]> var_1499 = mul(x = encoder_5_transformer_layers_1_layer_scale_1_scale, y = x_63)[name = tensor<string, []>("op_1499")]; |
| tensor<fp32, [1, ?, 256]> input_93 = add(x = input_89, y = var_1499)[name = tensor<string, []>("input_93")]; |
| tensor<int32, [1]> input_95_axes_0 = const()[name = tensor<string, []>("input_95_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_95 = layer_norm(axes = input_95_axes_0, beta = encoder_5_transformer_layers_1_norm2_bias, epsilon = var_1227, gamma = encoder_5_transformer_layers_1_norm2_weight, x = input_93)[name = tensor<string, []>("input_95")]; |
| tensor<fp32, [1, ?, 1024]> input_97 = linear(bias = linear_3_bias_0, weight = encoder_5_transformer_layers_1_ffn_0_weight, x = input_95)[name = tensor<string, []>("linear_35")]; |
| tensor<string, []> input_99_mode_0 = const()[name = tensor<string, []>("input_99_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_99 = gelu(mode = input_99_mode_0, x = input_97)[name = tensor<string, []>("input_99")]; |
| tensor<fp32, [1, ?, 256]> x_65 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_1_ffn_2_weight, x = input_99)[name = tensor<string, []>("linear_36")]; |
| tensor<fp32, [1, ?, 256]> var_1514 = mul(x = encoder_5_transformer_layers_1_layer_scale_2_scale, y = x_65)[name = tensor<string, []>("op_1514")]; |
| tensor<fp32, [1, ?, 256]> input_101 = add(x = input_93, y = var_1514)[name = tensor<string, []>("input_101")]; |
| tensor<fp32, [1, ?, 384]> var_1517 = linear(bias = linear_17_bias_0, weight = encoder_5_output_proj_weight, x = input_101)[name = tensor<string, []>("linear_37")]; |
| tensor<int32, [3]> x_67_perm_0 = const()[name = tensor<string, []>("x_67_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [4]> var_1528 = const()[name = tensor<string, []>("op_1528"), val = tensor<int32, [4]>([1, 384, -1, 2])]; |
| tensor<fp32, [1, 384, ?]> x_67 = transpose(perm = x_67_perm_0, x = var_1517)[name = tensor<string, []>("transpose_59")]; |
| tensor<fp32, [1, 384, ?, 2]> var_1529 = reshape(shape = var_1528, x = x_67)[name = tensor<string, []>("op_1529")]; |
| tensor<int32, [4]> var_1530 = const()[name = tensor<string, []>("op_1530"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| tensor<int32, [3]> var_1535 = const()[name = tensor<string, []>("op_1535"), val = tensor<int32, [3]>([1, 768, -1])]; |
| tensor<fp32, [1, 384, 2, ?]> var_1531 = transpose(perm = var_1530, x = var_1529)[name = tensor<string, []>("transpose_58")]; |
| tensor<fp32, [1, 768, ?]> x_69 = reshape(shape = var_1535, x = var_1531)[name = tensor<string, []>("x_69")]; |
| tensor<fp32, []> var_1519_promoted = const()[name = tensor<string, []>("op_1519_promoted"), val = tensor<fp32, []>(0x1p+1)]; |
| tensor<fp32, [1]> floor_div_3 = floor_div(x = floor_div_2, y = var_1519_promoted)[name = tensor<string, []>("floor_div_3")]; |
| tensor<int32, []> var_1542 = const()[name = tensor<string, []>("op_1542"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> var_1549 = const()[name = tensor<string, []>("op_1549"), val = tensor<int32, []>(500)]; |
| tensor<fp32, []> var_1556 = const()[name = tensor<string, []>("op_1556"), val = tensor<fp32, []>(0x1.4f8b58p-17)]; |
| tensor<int32, []> var_1560 = const()[name = tensor<string, []>("op_1560"), val = tensor<int32, []>(1)]; |
| tensor<int32, [3]> input_103_perm_0 = const()[name = tensor<string, []>("input_103_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, ?, 768]> input_103 = transpose(perm = input_103_perm_0, x = x_69)[name = tensor<string, []>("transpose_57")]; |
| tensor<fp32, [1, ?, 256]> x_71 = linear(bias = linear_0_bias_0, weight = encoder_7_input_proj_weight, x = input_103)[name = tensor<string, []>("linear_38")]; |
| tensor<int32, [1]> x_73_axes_0 = const()[name = tensor<string, []>("x_73_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_73 = layer_norm(axes = x_73_axes_0, beta = encoder_7_transformer_layers_0_norm1_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_0_norm1_weight, x = x_71)[name = tensor<string, []>("x_73")]; |
| tensor<int32, [3]> var_1588_shape = shape(x = x_73)[name = tensor<string, []>("op_1588_shape")]; |
| tensor<int32, []> gather_90_batch_dims_0 = const()[name = tensor<string, []>("gather_90_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_90_validate_indices_0 = const()[name = tensor<string, []>("gather_90_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_32 = const()[name = tensor<string, []>("select_32"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_90_axis_1 = const()[name = tensor<string, []>("gather_90_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_90 = gather(axis = gather_90_axis_1, batch_dims = gather_90_batch_dims_0, indices = select_32, validate_indices = gather_90_validate_indices_0, x = var_1588_shape)[name = tensor<string, []>("gather_90")]; |
| tensor<fp32, [1, ?, 768]> projected_33 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_0_self_attn_in_proj_weight, x = x_73)[name = tensor<string, []>("linear_39")]; |
| tensor<int32, [5]> concat_96x = const()[name = tensor<string, []>("concat_96x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_1595 = reshape(shape = concat_96x, x = projected_33)[name = tensor<string, []>("op_1595")]; |
| tensor<int32, [5]> var_1596 = const()[name = tensor<string, []>("op_1596"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_33_begin_0 = const()[name = tensor<string, []>("q_33_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_33_end_0 = const()[name = tensor<string, []>("q_33_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_33_end_mask_0 = const()[name = tensor<string, []>("q_33_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_33_squeeze_mask_0 = const()[name = tensor<string, []>("q_33_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_35 = transpose(perm = var_1596, x = var_1595)[name = tensor<string, []>("transpose_56")]; |
| tensor<fp32, [1, 4, ?, 64]> q_33 = slice_by_index(begin = q_33_begin_0, end = q_33_end_0, end_mask = q_33_end_mask_0, squeeze_mask = q_33_squeeze_mask_0, x = projected_35)[name = tensor<string, []>("q_33")]; |
| tensor<int32, [5]> k_33_begin_0 = const()[name = tensor<string, []>("k_33_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_33_end_0 = const()[name = tensor<string, []>("k_33_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_33_end_mask_0 = const()[name = tensor<string, []>("k_33_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_33_squeeze_mask_0 = const()[name = tensor<string, []>("k_33_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_33 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = projected_35)[name = tensor<string, []>("k_33")]; |
| tensor<int32, [5]> v_17_begin_0 = const()[name = tensor<string, []>("v_17_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_17_end_0 = const()[name = tensor<string, []>("v_17_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_17_end_mask_0 = const()[name = tensor<string, []>("v_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_17_squeeze_mask_0 = const()[name = tensor<string, []>("v_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_17 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = projected_35)[name = tensor<string, []>("v_17")]; |
| tensor<fp32, [32]> freqs_17 = const()[name = tensor<string, []>("freqs_17"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44432640)))]; |
| tensor<int32, [4]> var_1608_shape = shape(x = q_33)[name = tensor<string, []>("op_1608_shape")]; |
| tensor<int32, []> gather_94_batch_dims_0 = const()[name = tensor<string, []>("gather_94_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_94_validate_indices_0 = const()[name = tensor<string, []>("gather_94_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_33 = const()[name = tensor<string, []>("select_33"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_94_axis_1 = const()[name = tensor<string, []>("gather_94_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_94 = gather(axis = gather_94_axis_1, batch_dims = gather_94_batch_dims_0, indices = select_33, validate_indices = gather_94_validate_indices_0, x = var_1608_shape)[name = tensor<string, []>("gather_94")]; |
| tensor<fp32, [1, 1]> var_1611 = const()[name = tensor<string, []>("op_1611"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_75 = const()[name = tensor<string, []>("const_75"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_76 = const()[name = tensor<string, []>("const_76"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_1612 = range_1d(end = gather_94, start = const_75, step = const_76)[name = tensor<string, []>("op_1612")]; |
| tensor<string, []> var_1612_promoted_dtype_0 = const()[name = tensor<string, []>("op_1612_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_1612_promoted = cast(dtype = var_1612_promoted_dtype_0, x = var_1612)[name = tensor<string, []>("cast_436")]; |
| tensor<fp32, [1, ?]> ts_33 = add(x = var_1611, y = var_1612_promoted)[name = tensor<string, []>("ts_33")]; |
| tensor<int32, [4]> var_1615 = const()[name = tensor<string, []>("op_1615"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_35 = reshape(shape = var_1615, x = ts_33)[name = tensor<string, []>("ts_35")]; |
| tensor<int32, [5]> concat_97x = const()[name = tensor<string, []>("concat_97x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_17 = reshape(shape = concat_97x, x = q_33)[name = tensor<string, []>("q2_17")]; |
| tensor<int32, [5]> concat_98x = const()[name = tensor<string, []>("concat_98x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_17 = reshape(shape = concat_98x, x = k_33)[name = tensor<string, []>("k2_17")]; |
| tensor<int32, [5]> qr_17_begin_0 = const()[name = tensor<string, []>("qr_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_17_end_0 = const()[name = tensor<string, []>("qr_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_17_end_mask_0 = const()[name = tensor<string, []>("qr_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_17_squeeze_mask_0 = const()[name = tensor<string, []>("qr_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_17 = slice_by_index(begin = qr_17_begin_0, end = qr_17_end_0, end_mask = qr_17_end_mask_0, squeeze_mask = qr_17_squeeze_mask_0, x = q2_17)[name = tensor<string, []>("qr_17")]; |
| tensor<int32, [5]> qi_17_begin_0 = const()[name = tensor<string, []>("qi_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_17_end_0 = const()[name = tensor<string, []>("qi_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_17_end_mask_0 = const()[name = tensor<string, []>("qi_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_17_squeeze_mask_0 = const()[name = tensor<string, []>("qi_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_17 = slice_by_index(begin = qi_17_begin_0, end = qi_17_end_0, end_mask = qi_17_end_mask_0, squeeze_mask = qi_17_squeeze_mask_0, x = q2_17)[name = tensor<string, []>("qi_17")]; |
| tensor<int32, [5]> kr_17_begin_0 = const()[name = tensor<string, []>("kr_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_17_end_0 = const()[name = tensor<string, []>("kr_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_17_end_mask_0 = const()[name = tensor<string, []>("kr_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_17_squeeze_mask_0 = const()[name = tensor<string, []>("kr_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_17 = slice_by_index(begin = kr_17_begin_0, end = kr_17_end_0, end_mask = kr_17_end_mask_0, squeeze_mask = kr_17_squeeze_mask_0, x = k2_17)[name = tensor<string, []>("kr_17")]; |
| tensor<int32, [5]> ki_17_begin_0 = const()[name = tensor<string, []>("ki_17_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_17_end_0 = const()[name = tensor<string, []>("ki_17_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_17_end_mask_0 = const()[name = tensor<string, []>("ki_17_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_17_squeeze_mask_0 = const()[name = tensor<string, []>("ki_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_17 = slice_by_index(begin = ki_17_begin_0, end = ki_17_end_0, end_mask = ki_17_end_mask_0, squeeze_mask = ki_17_squeeze_mask_0, x = k2_17)[name = tensor<string, []>("ki_17")]; |
| tensor<fp32, [1, 1, ?, 32]> var_1628 = mul(x = freqs_17, y = ts_35)[name = tensor<string, []>("op_1628")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_17 = cos(x = var_1628)[name = tensor<string, []>("rotr_17")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_17 = sin(x = var_1628)[name = tensor<string, []>("roti_17")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1632 = mul(x = qr_17, y = rotr_17)[name = tensor<string, []>("op_1632")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1633 = mul(x = qi_17, y = roti_17)[name = tensor<string, []>("op_1633")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1634 = sub(x = var_1632, y = var_1633)[name = tensor<string, []>("op_1634")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1635 = mul(x = qr_17, y = roti_17)[name = tensor<string, []>("op_1635")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1636 = mul(x = qi_17, y = rotr_17)[name = tensor<string, []>("op_1636")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1637 = add(x = var_1635, y = var_1636)[name = tensor<string, []>("op_1637")]; |
| tensor<int32, []> qo_17_axis_0 = const()[name = tensor<string, []>("qo_17_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_17 = stack(axis = qo_17_axis_0, values = (var_1634, var_1637))[name = tensor<string, []>("qo_17")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1640 = mul(x = kr_17, y = rotr_17)[name = tensor<string, []>("op_1640")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1641 = mul(x = ki_17, y = roti_17)[name = tensor<string, []>("op_1641")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1642 = sub(x = var_1640, y = var_1641)[name = tensor<string, []>("op_1642")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1643 = mul(x = kr_17, y = roti_17)[name = tensor<string, []>("op_1643")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1644 = mul(x = ki_17, y = rotr_17)[name = tensor<string, []>("op_1644")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1645 = add(x = var_1643, y = var_1644)[name = tensor<string, []>("op_1645")]; |
| tensor<int32, []> ko_17_axis_0 = const()[name = tensor<string, []>("ko_17_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_17 = stack(axis = ko_17_axis_0, values = (var_1642, var_1645))[name = tensor<string, []>("ko_17")]; |
| tensor<int32, [4]> concat_99x = const()[name = tensor<string, []>("concat_99x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_35 = reshape(shape = concat_99x, x = qo_17)[name = tensor<string, []>("q_35")]; |
| tensor<int32, [4]> concat_100x = const()[name = tensor<string, []>("concat_100x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_35 = reshape(shape = concat_100x, x = ko_17)[name = tensor<string, []>("k_35")]; |
| tensor<int32, []> const_77 = const()[name = tensor<string, []>("const_77"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_78 = const()[name = tensor<string, []>("const_78"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_17 = range_1d(end = gather_90, start = const_77, step = const_78)[name = tensor<string, []>("positions_17")]; |
| tensor<int32, [3]> concat_101x = const()[name = tensor<string, []>("concat_101x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_1657 = reshape(shape = concat_101x, x = positions_17)[name = tensor<string, []>("op_1657")]; |
| tensor<int32, [3]> var_1658 = const()[name = tensor<string, []>("op_1658"), val = tensor<int32, [3]>([-1, 1, 1])]; |
| tensor<fp32, [1, 1, 1]> var_1659 = reshape(shape = var_1658, x = floor_div_3)[name = tensor<string, []>("op_1659")]; |
| tensor<string, []> var_1657_promoted_dtype_0 = const()[name = tensor<string, []>("op_1657_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_1657_promoted = cast(dtype = var_1657_promoted_dtype_0, x = var_1657)[name = tensor<string, []>("cast_435")]; |
| tensor<bool, [1, 1, ?]> valid_k_17 = less(x = var_1657_promoted, y = var_1659)[name = tensor<string, []>("valid_k_17")]; |
| tensor<int32, [3]> concat_102x = const()[name = tensor<string, []>("concat_102x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_1662 = reshape(shape = concat_102x, x = positions_17)[name = tensor<string, []>("op_1662")]; |
| tensor<int32, [1, ?, ?]> delta_17 = sub(x = var_1662, y = var_1657)[name = tensor<string, []>("delta_17")]; |
| tensor<int32, []> concat_104_axis_0 = const()[name = tensor<string, []>("concat_104_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_104_interleave_0 = const()[name = tensor<string, []>("concat_104_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_104 = concat(axis = concat_104_axis_0, interleave = concat_104_interleave_0, values = (var_1560, gather_90, gather_90))[name = tensor<string, []>("concat_104")]; |
| tensor<fp32, []> fill_8_value_0 = const()[name = tensor<string, []>("fill_8_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_8 = fill(shape = concat_104, value = fill_8_value_0)[name = tensor<string, []>("fill_8")]; |
| tensor<string, []> attn_bias_65_dtype_0 = const()[name = tensor<string, []>("attn_bias_65_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_1668 = greater_equal(x = delta_17, y = var_1542)[name = tensor<string, []>("op_1668")]; |
| tensor<bool, [1, ?, ?]> attn_bias_65 = cast(dtype = attn_bias_65_dtype_0, x = fill_8)[name = tensor<string, []>("cast_434")]; |
| tensor<bool, [1, ?, ?]> attn_bias_67 = logical_and(x = attn_bias_65, y = var_1668)[name = tensor<string, []>("attn_bias_67")]; |
| tensor<bool, [1, ?, ?]> var_1670 = less(x = delta_17, y = var_1549)[name = tensor<string, []>("op_1670")]; |
| tensor<bool, [1, ?, ?]> attn_bias_69 = logical_and(x = attn_bias_67, y = var_1670)[name = tensor<string, []>("attn_bias_69")]; |
| tensor<bool, [1, ?, ?]> var_1672 = logical_and(x = attn_bias_69, y = valid_k_17)[name = tensor<string, []>("op_1672")]; |
| tensor<int32, [1]> var_1674_axes_0 = const()[name = tensor<string, []>("op_1674_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_1674 = expand_dims(axes = var_1674_axes_0, x = var_1672)[name = tensor<string, []>("op_1674")]; |
| tensor<string, []> cast_202_dtype_0 = const()[name = tensor<string, []>("cast_202_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_8_x_0 = const()[name = tensor<string, []>("sub_8_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_202 = cast(dtype = cast_202_dtype_0, x = var_1674)[name = tensor<string, []>("cast_433")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_8 = sub(x = sub_8_x_0, y = cast_202)[name = tensor<string, []>("sub_8")]; |
| tensor<fp32, []> mul_16_x_0 = const()[name = tensor<string, []>("mul_16_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_16 = mul(x = mul_16_x_0, y = sub_8)[name = tensor<string, []>("mul_16")]; |
| tensor<fp32, []> mul_17_y_0 = const()[name = tensor<string, []>("mul_17_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_17 = mul(x = q_35, y = mul_17_y_0)[name = tensor<string, []>("mul_17")]; |
| tensor<bool, []> matmul_8_transpose_y_0 = const()[name = tensor<string, []>("matmul_8_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_8_transpose_x_0 = const()[name = tensor<string, []>("matmul_8_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_8 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = mul_17, y = k_35)[name = tensor<string, []>("matmul_8")]; |
| tensor<fp32, [1, 4, ?, ?]> add_8 = add(x = matmul_8, y = mul_16)[name = tensor<string, []>("add_8")]; |
| tensor<int32, []> softmax_8_axis_0 = const()[name = tensor<string, []>("softmax_8_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_8 = softmax(axis = softmax_8_axis_0, x = add_8)[name = tensor<string, []>("softmax_8")]; |
| tensor<bool, []> out_33_transpose_x_0 = const()[name = tensor<string, []>("out_33_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> out_33_transpose_y_0 = const()[name = tensor<string, []>("out_33_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, 64]> out_33 = matmul(transpose_x = out_33_transpose_x_0, transpose_y = out_33_transpose_y_0, x = softmax_8, y = v_17)[name = tensor<string, []>("out_33")]; |
| tensor<int32, [2]> concat_105x = const()[name = tensor<string, []>("concat_105x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_1680 = reshape(shape = concat_105x, x = positions_17)[name = tensor<string, []>("op_1680")]; |
| tensor<int32, [2]> var_1681 = const()[name = tensor<string, []>("op_1681"), val = tensor<int32, [2]>([-1, 1])]; |
| tensor<fp32, [1, 1]> var_1682 = reshape(shape = var_1681, x = floor_div_3)[name = tensor<string, []>("op_1682")]; |
| tensor<string, []> var_1680_promoted_dtype_0 = const()[name = tensor<string, []>("op_1680_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_1680_promoted = cast(dtype = var_1680_promoted_dtype_0, x = var_1680)[name = tensor<string, []>("cast_432")]; |
| tensor<bool, [1, ?]> var_1683 = less(x = var_1680_promoted, y = var_1682)[name = tensor<string, []>("op_1683")]; |
| tensor<int32, [4]> concat_106x = const()[name = tensor<string, []>("concat_106x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_17 = reshape(shape = concat_106x, x = var_1683)[name = tensor<string, []>("valid_q_17")]; |
| tensor<fp32, []> var_1687 = const()[name = tensor<string, []>("op_1687"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_35 = select(a = out_33, b = var_1687, cond = valid_q_17)[name = tensor<string, []>("out_35")]; |
| tensor<int32, [4]> var_1689_perm_0 = const()[name = tensor<string, []>("op_1689_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_107x = const()[name = tensor<string, []>("concat_107x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_1689 = transpose(perm = var_1689_perm_0, x = out_35)[name = tensor<string, []>("transpose_55")]; |
| tensor<fp32, [1, ?, 256]> input_105 = reshape(shape = concat_107x, x = var_1689)[name = tensor<string, []>("input_105")]; |
| tensor<fp32, [1, ?, 256]> x_75 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_0_self_attn_out_proj_weight, x = input_105)[name = tensor<string, []>("linear_40")]; |
| tensor<fp32, [1, ?, 256]> var_1695 = mul(x = encoder_7_transformer_layers_0_layer_scale_1_scale, y = x_75)[name = tensor<string, []>("op_1695")]; |
| tensor<fp32, [1, ?, 256]> input_107 = add(x = x_71, y = var_1695)[name = tensor<string, []>("input_107")]; |
| tensor<int32, [1]> input_109_axes_0 = const()[name = tensor<string, []>("input_109_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_109 = layer_norm(axes = input_109_axes_0, beta = encoder_7_transformer_layers_0_norm2_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_0_norm2_weight, x = input_107)[name = tensor<string, []>("input_109")]; |
| tensor<fp32, [1, ?, 1024]> input_111 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_0_ffn_0_weight, x = input_109)[name = tensor<string, []>("linear_41")]; |
| tensor<string, []> input_113_mode_0 = const()[name = tensor<string, []>("input_113_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_113 = gelu(mode = input_113_mode_0, x = input_111)[name = tensor<string, []>("input_113")]; |
| tensor<fp32, [1, ?, 256]> x_77 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_0_ffn_2_weight, x = input_113)[name = tensor<string, []>("linear_42")]; |
| tensor<fp32, [1, ?, 256]> var_1710 = mul(x = encoder_7_transformer_layers_0_layer_scale_2_scale, y = x_77)[name = tensor<string, []>("op_1710")]; |
| tensor<fp32, [1, ?, 256]> input_115 = add(x = input_107, y = var_1710)[name = tensor<string, []>("input_115")]; |
| tensor<int32, [1]> x_79_axes_0 = const()[name = tensor<string, []>("x_79_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_79 = layer_norm(axes = x_79_axes_0, beta = encoder_7_transformer_layers_1_norm1_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_1_norm1_weight, x = input_115)[name = tensor<string, []>("x_79")]; |
| tensor<int32, [3]> var_1725_shape = shape(x = x_79)[name = tensor<string, []>("op_1725_shape")]; |
| tensor<int32, []> gather_100_batch_dims_0 = const()[name = tensor<string, []>("gather_100_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_100_validate_indices_0 = const()[name = tensor<string, []>("gather_100_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_34 = const()[name = tensor<string, []>("select_34"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_100_axis_1 = const()[name = tensor<string, []>("gather_100_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_100 = gather(axis = gather_100_axis_1, batch_dims = gather_100_batch_dims_0, indices = select_34, validate_indices = gather_100_validate_indices_0, x = var_1725_shape)[name = tensor<string, []>("gather_100")]; |
| tensor<fp32, [1, ?, 768]> projected_37 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_1_self_attn_in_proj_weight, x = x_79)[name = tensor<string, []>("linear_43")]; |
| tensor<int32, [5]> concat_108x = const()[name = tensor<string, []>("concat_108x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_1732 = reshape(shape = concat_108x, x = projected_37)[name = tensor<string, []>("op_1732")]; |
| tensor<int32, [5]> var_1733 = const()[name = tensor<string, []>("op_1733"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_37_begin_0 = const()[name = tensor<string, []>("q_37_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_37_end_0 = const()[name = tensor<string, []>("q_37_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_37_end_mask_0 = const()[name = tensor<string, []>("q_37_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_37_squeeze_mask_0 = const()[name = tensor<string, []>("q_37_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_39 = transpose(perm = var_1733, x = var_1732)[name = tensor<string, []>("transpose_54")]; |
| tensor<fp32, [1, 4, ?, 64]> q_37 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = projected_39)[name = tensor<string, []>("q_37")]; |
| tensor<int32, [5]> k_37_begin_0 = const()[name = tensor<string, []>("k_37_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_37_end_0 = const()[name = tensor<string, []>("k_37_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_37_end_mask_0 = const()[name = tensor<string, []>("k_37_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_37_squeeze_mask_0 = const()[name = tensor<string, []>("k_37_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_37 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = projected_39)[name = tensor<string, []>("k_37")]; |
| tensor<int32, [5]> v_19_begin_0 = const()[name = tensor<string, []>("v_19_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_19_end_0 = const()[name = tensor<string, []>("v_19_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_19_end_mask_0 = const()[name = tensor<string, []>("v_19_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_19_squeeze_mask_0 = const()[name = tensor<string, []>("v_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = projected_39)[name = tensor<string, []>("v_19")]; |
| tensor<fp32, [32]> freqs_19 = const()[name = tensor<string, []>("freqs_19"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44432832)))]; |
| tensor<int32, [4]> var_1745_shape = shape(x = q_37)[name = tensor<string, []>("op_1745_shape")]; |
| tensor<int32, []> gather_104_batch_dims_0 = const()[name = tensor<string, []>("gather_104_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_104_validate_indices_0 = const()[name = tensor<string, []>("gather_104_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_35 = const()[name = tensor<string, []>("select_35"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_104_axis_1 = const()[name = tensor<string, []>("gather_104_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_104 = gather(axis = gather_104_axis_1, batch_dims = gather_104_batch_dims_0, indices = select_35, validate_indices = gather_104_validate_indices_0, x = var_1745_shape)[name = tensor<string, []>("gather_104")]; |
| tensor<fp32, [1, 1]> var_1748 = const()[name = tensor<string, []>("op_1748"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_84 = const()[name = tensor<string, []>("const_84"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_85 = const()[name = tensor<string, []>("const_85"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_1749 = range_1d(end = gather_104, start = const_84, step = const_85)[name = tensor<string, []>("op_1749")]; |
| tensor<string, []> var_1749_promoted_dtype_0 = const()[name = tensor<string, []>("op_1749_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_1749_promoted = cast(dtype = var_1749_promoted_dtype_0, x = var_1749)[name = tensor<string, []>("cast_431")]; |
| tensor<fp32, [1, ?]> ts_37 = add(x = var_1748, y = var_1749_promoted)[name = tensor<string, []>("ts_37")]; |
| tensor<int32, [4]> var_1752 = const()[name = tensor<string, []>("op_1752"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_39 = reshape(shape = var_1752, x = ts_37)[name = tensor<string, []>("ts_39")]; |
| tensor<int32, [5]> concat_109x = const()[name = tensor<string, []>("concat_109x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_19 = reshape(shape = concat_109x, x = q_37)[name = tensor<string, []>("q2_19")]; |
| tensor<int32, [5]> concat_110x = const()[name = tensor<string, []>("concat_110x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_19 = reshape(shape = concat_110x, x = k_37)[name = tensor<string, []>("k2_19")]; |
| tensor<int32, [5]> qr_19_begin_0 = const()[name = tensor<string, []>("qr_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_19_end_0 = const()[name = tensor<string, []>("qr_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_19_end_mask_0 = const()[name = tensor<string, []>("qr_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_19_squeeze_mask_0 = const()[name = tensor<string, []>("qr_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_19 = slice_by_index(begin = qr_19_begin_0, end = qr_19_end_0, end_mask = qr_19_end_mask_0, squeeze_mask = qr_19_squeeze_mask_0, x = q2_19)[name = tensor<string, []>("qr_19")]; |
| tensor<int32, [5]> qi_19_begin_0 = const()[name = tensor<string, []>("qi_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_19_end_0 = const()[name = tensor<string, []>("qi_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_19_end_mask_0 = const()[name = tensor<string, []>("qi_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_19_squeeze_mask_0 = const()[name = tensor<string, []>("qi_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_19 = slice_by_index(begin = qi_19_begin_0, end = qi_19_end_0, end_mask = qi_19_end_mask_0, squeeze_mask = qi_19_squeeze_mask_0, x = q2_19)[name = tensor<string, []>("qi_19")]; |
| tensor<int32, [5]> kr_19_begin_0 = const()[name = tensor<string, []>("kr_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_19_end_0 = const()[name = tensor<string, []>("kr_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_19_end_mask_0 = const()[name = tensor<string, []>("kr_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_19_squeeze_mask_0 = const()[name = tensor<string, []>("kr_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_19 = slice_by_index(begin = kr_19_begin_0, end = kr_19_end_0, end_mask = kr_19_end_mask_0, squeeze_mask = kr_19_squeeze_mask_0, x = k2_19)[name = tensor<string, []>("kr_19")]; |
| tensor<int32, [5]> ki_19_begin_0 = const()[name = tensor<string, []>("ki_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_19_end_0 = const()[name = tensor<string, []>("ki_19_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_19_end_mask_0 = const()[name = tensor<string, []>("ki_19_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_19_squeeze_mask_0 = const()[name = tensor<string, []>("ki_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_19 = slice_by_index(begin = ki_19_begin_0, end = ki_19_end_0, end_mask = ki_19_end_mask_0, squeeze_mask = ki_19_squeeze_mask_0, x = k2_19)[name = tensor<string, []>("ki_19")]; |
| tensor<fp32, [1, 1, ?, 32]> var_1765 = mul(x = freqs_19, y = ts_39)[name = tensor<string, []>("op_1765")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_19 = cos(x = var_1765)[name = tensor<string, []>("rotr_19")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_19 = sin(x = var_1765)[name = tensor<string, []>("roti_19")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1769 = mul(x = qr_19, y = rotr_19)[name = tensor<string, []>("op_1769")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1770 = mul(x = qi_19, y = roti_19)[name = tensor<string, []>("op_1770")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1771 = sub(x = var_1769, y = var_1770)[name = tensor<string, []>("op_1771")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1772 = mul(x = qr_19, y = roti_19)[name = tensor<string, []>("op_1772")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1773 = mul(x = qi_19, y = rotr_19)[name = tensor<string, []>("op_1773")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1774 = add(x = var_1772, y = var_1773)[name = tensor<string, []>("op_1774")]; |
| tensor<int32, []> qo_19_axis_0 = const()[name = tensor<string, []>("qo_19_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_19 = stack(axis = qo_19_axis_0, values = (var_1771, var_1774))[name = tensor<string, []>("qo_19")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1777 = mul(x = kr_19, y = rotr_19)[name = tensor<string, []>("op_1777")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1778 = mul(x = ki_19, y = roti_19)[name = tensor<string, []>("op_1778")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1779 = sub(x = var_1777, y = var_1778)[name = tensor<string, []>("op_1779")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1780 = mul(x = kr_19, y = roti_19)[name = tensor<string, []>("op_1780")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1781 = mul(x = ki_19, y = rotr_19)[name = tensor<string, []>("op_1781")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1782 = add(x = var_1780, y = var_1781)[name = tensor<string, []>("op_1782")]; |
| tensor<int32, []> ko_19_axis_0 = const()[name = tensor<string, []>("ko_19_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_19 = stack(axis = ko_19_axis_0, values = (var_1779, var_1782))[name = tensor<string, []>("ko_19")]; |
| tensor<int32, [4]> concat_111x = const()[name = tensor<string, []>("concat_111x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_39 = reshape(shape = concat_111x, x = qo_19)[name = tensor<string, []>("q_39")]; |
| tensor<int32, [4]> concat_112x = const()[name = tensor<string, []>("concat_112x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_39 = reshape(shape = concat_112x, x = ko_19)[name = tensor<string, []>("k_39")]; |
| tensor<int32, []> const_86 = const()[name = tensor<string, []>("const_86"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_87 = const()[name = tensor<string, []>("const_87"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_19 = range_1d(end = gather_100, start = const_86, step = const_87)[name = tensor<string, []>("positions_19")]; |
| tensor<int32, [3]> concat_113x = const()[name = tensor<string, []>("concat_113x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_1794 = reshape(shape = concat_113x, x = positions_19)[name = tensor<string, []>("op_1794")]; |
| tensor<string, []> var_1794_promoted_dtype_0 = const()[name = tensor<string, []>("op_1794_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_1794_promoted = cast(dtype = var_1794_promoted_dtype_0, x = var_1794)[name = tensor<string, []>("cast_430")]; |
| tensor<bool, [1, 1, ?]> valid_k_19 = less(x = var_1794_promoted, y = var_1659)[name = tensor<string, []>("valid_k_19")]; |
| tensor<int32, [3]> concat_114x = const()[name = tensor<string, []>("concat_114x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_1799 = reshape(shape = concat_114x, x = positions_19)[name = tensor<string, []>("op_1799")]; |
| tensor<int32, [1, ?, ?]> delta_19 = sub(x = var_1799, y = var_1794)[name = tensor<string, []>("delta_19")]; |
| tensor<int32, []> concat_116_axis_0 = const()[name = tensor<string, []>("concat_116_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_116_interleave_0 = const()[name = tensor<string, []>("concat_116_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_116 = concat(axis = concat_116_axis_0, interleave = concat_116_interleave_0, values = (var_1560, gather_100, gather_100))[name = tensor<string, []>("concat_116")]; |
| tensor<fp32, []> fill_9_value_0 = const()[name = tensor<string, []>("fill_9_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_9 = fill(shape = concat_116, value = fill_9_value_0)[name = tensor<string, []>("fill_9")]; |
| tensor<string, []> attn_bias_73_dtype_0 = const()[name = tensor<string, []>("attn_bias_73_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_1805 = greater_equal(x = delta_19, y = var_1542)[name = tensor<string, []>("op_1805")]; |
| tensor<bool, [1, ?, ?]> attn_bias_73 = cast(dtype = attn_bias_73_dtype_0, x = fill_9)[name = tensor<string, []>("cast_429")]; |
| tensor<bool, [1, ?, ?]> attn_bias_75 = logical_and(x = attn_bias_73, y = var_1805)[name = tensor<string, []>("attn_bias_75")]; |
| tensor<bool, [1, ?, ?]> var_1807 = less(x = delta_19, y = var_1549)[name = tensor<string, []>("op_1807")]; |
| tensor<bool, [1, ?, ?]> attn_bias_77 = logical_and(x = attn_bias_75, y = var_1807)[name = tensor<string, []>("attn_bias_77")]; |
| tensor<bool, [1, ?, ?]> var_1809 = logical_and(x = attn_bias_77, y = valid_k_19)[name = tensor<string, []>("op_1809")]; |
| tensor<int32, [1]> var_1811_axes_0 = const()[name = tensor<string, []>("op_1811_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_1811 = expand_dims(axes = var_1811_axes_0, x = var_1809)[name = tensor<string, []>("op_1811")]; |
| tensor<string, []> cast_224_dtype_0 = const()[name = tensor<string, []>("cast_224_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_9_x_0 = const()[name = tensor<string, []>("sub_9_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_224 = cast(dtype = cast_224_dtype_0, x = var_1811)[name = tensor<string, []>("cast_428")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_9 = sub(x = sub_9_x_0, y = cast_224)[name = tensor<string, []>("sub_9")]; |
| tensor<fp32, []> mul_18_x_0 = const()[name = tensor<string, []>("mul_18_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_18 = mul(x = mul_18_x_0, y = sub_9)[name = tensor<string, []>("mul_18")]; |
| tensor<fp32, []> mul_19_y_0 = const()[name = tensor<string, []>("mul_19_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_19 = mul(x = q_39, y = mul_19_y_0)[name = tensor<string, []>("mul_19")]; |
| tensor<bool, []> matmul_9_transpose_y_0 = const()[name = tensor<string, []>("matmul_9_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_9_transpose_x_0 = const()[name = tensor<string, []>("matmul_9_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_9 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = mul_19, y = k_39)[name = tensor<string, []>("matmul_9")]; |
| tensor<fp32, [1, 4, ?, ?]> add_9 = add(x = matmul_9, y = mul_18)[name = tensor<string, []>("add_9")]; |
| tensor<int32, []> softmax_9_axis_0 = const()[name = tensor<string, []>("softmax_9_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_9 = softmax(axis = softmax_9_axis_0, x = add_9)[name = tensor<string, []>("softmax_9")]; |
| tensor<bool, []> out_37_transpose_x_0 = const()[name = tensor<string, []>("out_37_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> out_37_transpose_y_0 = const()[name = tensor<string, []>("out_37_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, 64]> out_37 = matmul(transpose_x = out_37_transpose_x_0, transpose_y = out_37_transpose_y_0, x = softmax_9, y = v_19)[name = tensor<string, []>("out_37")]; |
| tensor<int32, [2]> concat_117x = const()[name = tensor<string, []>("concat_117x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_1817 = reshape(shape = concat_117x, x = positions_19)[name = tensor<string, []>("op_1817")]; |
| tensor<string, []> var_1817_promoted_dtype_0 = const()[name = tensor<string, []>("op_1817_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_1817_promoted = cast(dtype = var_1817_promoted_dtype_0, x = var_1817)[name = tensor<string, []>("cast_427")]; |
| tensor<bool, [1, ?]> var_1820 = less(x = var_1817_promoted, y = var_1682)[name = tensor<string, []>("op_1820")]; |
| tensor<int32, [4]> concat_118x = const()[name = tensor<string, []>("concat_118x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_19 = reshape(shape = concat_118x, x = var_1820)[name = tensor<string, []>("valid_q_19")]; |
| tensor<fp32, []> var_1824 = const()[name = tensor<string, []>("op_1824"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_39 = select(a = out_37, b = var_1824, cond = valid_q_19)[name = tensor<string, []>("out_39")]; |
| tensor<int32, [4]> var_1826_perm_0 = const()[name = tensor<string, []>("op_1826_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_119x = const()[name = tensor<string, []>("concat_119x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_1826 = transpose(perm = var_1826_perm_0, x = out_39)[name = tensor<string, []>("transpose_53")]; |
| tensor<fp32, [1, ?, 256]> input_117 = reshape(shape = concat_119x, x = var_1826)[name = tensor<string, []>("input_117")]; |
| tensor<fp32, [1, ?, 256]> x_81 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_1_self_attn_out_proj_weight, x = input_117)[name = tensor<string, []>("linear_44")]; |
| tensor<fp32, [1, ?, 256]> var_1832 = mul(x = encoder_7_transformer_layers_1_layer_scale_1_scale, y = x_81)[name = tensor<string, []>("op_1832")]; |
| tensor<fp32, [1, ?, 256]> input_119 = add(x = input_115, y = var_1832)[name = tensor<string, []>("input_119")]; |
| tensor<int32, [1]> input_121_axes_0 = const()[name = tensor<string, []>("input_121_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_121 = layer_norm(axes = input_121_axes_0, beta = encoder_7_transformer_layers_1_norm2_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_1_norm2_weight, x = input_119)[name = tensor<string, []>("input_121")]; |
| tensor<fp32, [1, ?, 1024]> input_123 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_1_ffn_0_weight, x = input_121)[name = tensor<string, []>("linear_45")]; |
| tensor<string, []> input_125_mode_0 = const()[name = tensor<string, []>("input_125_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_125 = gelu(mode = input_125_mode_0, x = input_123)[name = tensor<string, []>("input_125")]; |
| tensor<fp32, [1, ?, 256]> x_83 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_1_ffn_2_weight, x = input_125)[name = tensor<string, []>("linear_46")]; |
| tensor<fp32, [1, ?, 256]> var_1847 = mul(x = encoder_7_transformer_layers_1_layer_scale_2_scale, y = x_83)[name = tensor<string, []>("op_1847")]; |
| tensor<fp32, [1, ?, 256]> input_127 = add(x = input_119, y = var_1847)[name = tensor<string, []>("input_127")]; |
| tensor<int32, [1]> x_85_axes_0 = const()[name = tensor<string, []>("x_85_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_85 = layer_norm(axes = x_85_axes_0, beta = encoder_7_transformer_layers_2_norm1_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_2_norm1_weight, x = input_127)[name = tensor<string, []>("x_85")]; |
| tensor<int32, [3]> var_1862_shape = shape(x = x_85)[name = tensor<string, []>("op_1862_shape")]; |
| tensor<int32, []> gather_110_batch_dims_0 = const()[name = tensor<string, []>("gather_110_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_110_validate_indices_0 = const()[name = tensor<string, []>("gather_110_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_36 = const()[name = tensor<string, []>("select_36"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_110_axis_1 = const()[name = tensor<string, []>("gather_110_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_110 = gather(axis = gather_110_axis_1, batch_dims = gather_110_batch_dims_0, indices = select_36, validate_indices = gather_110_validate_indices_0, x = var_1862_shape)[name = tensor<string, []>("gather_110")]; |
| tensor<fp32, [1, ?, 768]> projected_41 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_2_self_attn_in_proj_weight, x = x_85)[name = tensor<string, []>("linear_47")]; |
| tensor<int32, [5]> concat_120x = const()[name = tensor<string, []>("concat_120x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_1869 = reshape(shape = concat_120x, x = projected_41)[name = tensor<string, []>("op_1869")]; |
| tensor<int32, [5]> var_1870 = const()[name = tensor<string, []>("op_1870"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_41_begin_0 = const()[name = tensor<string, []>("q_41_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_41_end_0 = const()[name = tensor<string, []>("q_41_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_41_end_mask_0 = const()[name = tensor<string, []>("q_41_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_41_squeeze_mask_0 = const()[name = tensor<string, []>("q_41_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected_43 = transpose(perm = var_1870, x = var_1869)[name = tensor<string, []>("transpose_52")]; |
| tensor<fp32, [1, 4, ?, 64]> q_41 = slice_by_index(begin = q_41_begin_0, end = q_41_end_0, end_mask = q_41_end_mask_0, squeeze_mask = q_41_squeeze_mask_0, x = projected_43)[name = tensor<string, []>("q_41")]; |
| tensor<int32, [5]> k_41_begin_0 = const()[name = tensor<string, []>("k_41_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_41_end_0 = const()[name = tensor<string, []>("k_41_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_41_end_mask_0 = const()[name = tensor<string, []>("k_41_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_41_squeeze_mask_0 = const()[name = tensor<string, []>("k_41_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_41 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = projected_43)[name = tensor<string, []>("k_41")]; |
| tensor<int32, [5]> v_21_begin_0 = const()[name = tensor<string, []>("v_21_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_21_end_0 = const()[name = tensor<string, []>("v_21_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_21_end_mask_0 = const()[name = tensor<string, []>("v_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_21_squeeze_mask_0 = const()[name = tensor<string, []>("v_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v_21 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = projected_43)[name = tensor<string, []>("v_21")]; |
| tensor<fp32, [32]> freqs_21 = const()[name = tensor<string, []>("freqs_21"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44433024)))]; |
| tensor<int32, [4]> var_1882_shape = shape(x = q_41)[name = tensor<string, []>("op_1882_shape")]; |
| tensor<int32, []> gather_114_batch_dims_0 = const()[name = tensor<string, []>("gather_114_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_114_validate_indices_0 = const()[name = tensor<string, []>("gather_114_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_37 = const()[name = tensor<string, []>("select_37"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_114_axis_1 = const()[name = tensor<string, []>("gather_114_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_114 = gather(axis = gather_114_axis_1, batch_dims = gather_114_batch_dims_0, indices = select_37, validate_indices = gather_114_validate_indices_0, x = var_1882_shape)[name = tensor<string, []>("gather_114")]; |
| tensor<fp32, [1, 1]> var_1885 = const()[name = tensor<string, []>("op_1885"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_93 = const()[name = tensor<string, []>("const_93"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_94 = const()[name = tensor<string, []>("const_94"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_1886 = range_1d(end = gather_114, start = const_93, step = const_94)[name = tensor<string, []>("op_1886")]; |
| tensor<string, []> var_1886_promoted_dtype_0 = const()[name = tensor<string, []>("op_1886_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_1886_promoted = cast(dtype = var_1886_promoted_dtype_0, x = var_1886)[name = tensor<string, []>("cast_426")]; |
| tensor<fp32, [1, ?]> ts_41 = add(x = var_1885, y = var_1886_promoted)[name = tensor<string, []>("ts_41")]; |
| tensor<int32, [4]> var_1889 = const()[name = tensor<string, []>("op_1889"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts_43 = reshape(shape = var_1889, x = ts_41)[name = tensor<string, []>("ts_43")]; |
| tensor<int32, [5]> concat_121x = const()[name = tensor<string, []>("concat_121x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2_21 = reshape(shape = concat_121x, x = q_41)[name = tensor<string, []>("q2_21")]; |
| tensor<int32, [5]> concat_122x = const()[name = tensor<string, []>("concat_122x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2_21 = reshape(shape = concat_122x, x = k_41)[name = tensor<string, []>("k2_21")]; |
| tensor<int32, [5]> qr_21_begin_0 = const()[name = tensor<string, []>("qr_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_21_end_0 = const()[name = tensor<string, []>("qr_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_21_end_mask_0 = const()[name = tensor<string, []>("qr_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_21_squeeze_mask_0 = const()[name = tensor<string, []>("qr_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr_21 = slice_by_index(begin = qr_21_begin_0, end = qr_21_end_0, end_mask = qr_21_end_mask_0, squeeze_mask = qr_21_squeeze_mask_0, x = q2_21)[name = tensor<string, []>("qr_21")]; |
| tensor<int32, [5]> qi_21_begin_0 = const()[name = tensor<string, []>("qi_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_21_end_0 = const()[name = tensor<string, []>("qi_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_21_end_mask_0 = const()[name = tensor<string, []>("qi_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_21_squeeze_mask_0 = const()[name = tensor<string, []>("qi_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi_21 = slice_by_index(begin = qi_21_begin_0, end = qi_21_end_0, end_mask = qi_21_end_mask_0, squeeze_mask = qi_21_squeeze_mask_0, x = q2_21)[name = tensor<string, []>("qi_21")]; |
| tensor<int32, [5]> kr_21_begin_0 = const()[name = tensor<string, []>("kr_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_21_end_0 = const()[name = tensor<string, []>("kr_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_21_end_mask_0 = const()[name = tensor<string, []>("kr_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_21_squeeze_mask_0 = const()[name = tensor<string, []>("kr_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr_21 = slice_by_index(begin = kr_21_begin_0, end = kr_21_end_0, end_mask = kr_21_end_mask_0, squeeze_mask = kr_21_squeeze_mask_0, x = k2_21)[name = tensor<string, []>("kr_21")]; |
| tensor<int32, [5]> ki_21_begin_0 = const()[name = tensor<string, []>("ki_21_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_21_end_0 = const()[name = tensor<string, []>("ki_21_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_21_end_mask_0 = const()[name = tensor<string, []>("ki_21_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_21_squeeze_mask_0 = const()[name = tensor<string, []>("ki_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki_21 = slice_by_index(begin = ki_21_begin_0, end = ki_21_end_0, end_mask = ki_21_end_mask_0, squeeze_mask = ki_21_squeeze_mask_0, x = k2_21)[name = tensor<string, []>("ki_21")]; |
| tensor<fp32, [1, 1, ?, 32]> var_1902 = mul(x = freqs_21, y = ts_43)[name = tensor<string, []>("op_1902")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr_21 = cos(x = var_1902)[name = tensor<string, []>("rotr_21")]; |
| tensor<fp32, [1, 1, ?, 32]> roti_21 = sin(x = var_1902)[name = tensor<string, []>("roti_21")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1906 = mul(x = qr_21, y = rotr_21)[name = tensor<string, []>("op_1906")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1907 = mul(x = qi_21, y = roti_21)[name = tensor<string, []>("op_1907")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1908 = sub(x = var_1906, y = var_1907)[name = tensor<string, []>("op_1908")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1909 = mul(x = qr_21, y = roti_21)[name = tensor<string, []>("op_1909")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1910 = mul(x = qi_21, y = rotr_21)[name = tensor<string, []>("op_1910")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1911 = add(x = var_1909, y = var_1910)[name = tensor<string, []>("op_1911")]; |
| tensor<int32, []> qo_21_axis_0 = const()[name = tensor<string, []>("qo_21_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo_21 = stack(axis = qo_21_axis_0, values = (var_1908, var_1911))[name = tensor<string, []>("qo_21")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1914 = mul(x = kr_21, y = rotr_21)[name = tensor<string, []>("op_1914")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1915 = mul(x = ki_21, y = roti_21)[name = tensor<string, []>("op_1915")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1916 = sub(x = var_1914, y = var_1915)[name = tensor<string, []>("op_1916")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1917 = mul(x = kr_21, y = roti_21)[name = tensor<string, []>("op_1917")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1918 = mul(x = ki_21, y = rotr_21)[name = tensor<string, []>("op_1918")]; |
| tensor<fp32, [1, 4, ?, 32]> var_1919 = add(x = var_1917, y = var_1918)[name = tensor<string, []>("op_1919")]; |
| tensor<int32, []> ko_21_axis_0 = const()[name = tensor<string, []>("ko_21_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko_21 = stack(axis = ko_21_axis_0, values = (var_1916, var_1919))[name = tensor<string, []>("ko_21")]; |
| tensor<int32, [4]> concat_123x = const()[name = tensor<string, []>("concat_123x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q_43 = reshape(shape = concat_123x, x = qo_21)[name = tensor<string, []>("q_43")]; |
| tensor<int32, [4]> concat_124x = const()[name = tensor<string, []>("concat_124x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k_43 = reshape(shape = concat_124x, x = ko_21)[name = tensor<string, []>("k_43")]; |
| tensor<int32, []> const_95 = const()[name = tensor<string, []>("const_95"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_96 = const()[name = tensor<string, []>("const_96"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions_21 = range_1d(end = gather_110, start = const_95, step = const_96)[name = tensor<string, []>("positions_21")]; |
| tensor<int32, [3]> concat_125x = const()[name = tensor<string, []>("concat_125x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_1931 = reshape(shape = concat_125x, x = positions_21)[name = tensor<string, []>("op_1931")]; |
| tensor<string, []> var_1931_promoted_dtype_0 = const()[name = tensor<string, []>("op_1931_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_1931_promoted = cast(dtype = var_1931_promoted_dtype_0, x = var_1931)[name = tensor<string, []>("cast_425")]; |
| tensor<bool, [1, 1, ?]> valid_k_21 = less(x = var_1931_promoted, y = var_1659)[name = tensor<string, []>("valid_k_21")]; |
| tensor<int32, [3]> concat_126x = const()[name = tensor<string, []>("concat_126x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_1936 = reshape(shape = concat_126x, x = positions_21)[name = tensor<string, []>("op_1936")]; |
| tensor<int32, [1, ?, ?]> delta_21 = sub(x = var_1936, y = var_1931)[name = tensor<string, []>("delta_21")]; |
| tensor<int32, []> concat_128_axis_0 = const()[name = tensor<string, []>("concat_128_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_128_interleave_0 = const()[name = tensor<string, []>("concat_128_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_128 = concat(axis = concat_128_axis_0, interleave = concat_128_interleave_0, values = (var_1560, gather_110, gather_110))[name = tensor<string, []>("concat_128")]; |
| tensor<fp32, []> fill_10_value_0 = const()[name = tensor<string, []>("fill_10_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_10 = fill(shape = concat_128, value = fill_10_value_0)[name = tensor<string, []>("fill_10")]; |
| tensor<string, []> attn_bias_81_dtype_0 = const()[name = tensor<string, []>("attn_bias_81_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_1942 = greater_equal(x = delta_21, y = var_1542)[name = tensor<string, []>("op_1942")]; |
| tensor<bool, [1, ?, ?]> attn_bias_81 = cast(dtype = attn_bias_81_dtype_0, x = fill_10)[name = tensor<string, []>("cast_424")]; |
| tensor<bool, [1, ?, ?]> attn_bias_83 = logical_and(x = attn_bias_81, y = var_1942)[name = tensor<string, []>("attn_bias_83")]; |
| tensor<bool, [1, ?, ?]> var_1944 = less(x = delta_21, y = var_1549)[name = tensor<string, []>("op_1944")]; |
| tensor<bool, [1, ?, ?]> attn_bias_85 = logical_and(x = attn_bias_83, y = var_1944)[name = tensor<string, []>("attn_bias_85")]; |
| tensor<bool, [1, ?, ?]> var_1946 = logical_and(x = attn_bias_85, y = valid_k_21)[name = tensor<string, []>("op_1946")]; |
| tensor<int32, [1]> var_1948_axes_0 = const()[name = tensor<string, []>("op_1948_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_1948 = expand_dims(axes = var_1948_axes_0, x = var_1946)[name = tensor<string, []>("op_1948")]; |
| tensor<string, []> cast_246_dtype_0 = const()[name = tensor<string, []>("cast_246_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_10_x_0 = const()[name = tensor<string, []>("sub_10_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_246 = cast(dtype = cast_246_dtype_0, x = var_1948)[name = tensor<string, []>("cast_423")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_10 = sub(x = sub_10_x_0, y = cast_246)[name = tensor<string, []>("sub_10")]; |
| tensor<fp32, []> mul_20_x_0 = const()[name = tensor<string, []>("mul_20_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_20 = mul(x = mul_20_x_0, y = sub_10)[name = tensor<string, []>("mul_20")]; |
| tensor<fp32, []> mul_21_y_0 = const()[name = tensor<string, []>("mul_21_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_21 = mul(x = q_43, y = mul_21_y_0)[name = tensor<string, []>("mul_21")]; |
| tensor<bool, []> matmul_10_transpose_y_0 = const()[name = tensor<string, []>("matmul_10_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_10_transpose_x_0 = const()[name = tensor<string, []>("matmul_10_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_10 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = mul_21, y = k_43)[name = tensor<string, []>("matmul_10")]; |
| tensor<fp32, [1, 4, ?, ?]> add_10 = add(x = matmul_10, y = mul_20)[name = tensor<string, []>("add_10")]; |
| tensor<int32, []> softmax_10_axis_0 = const()[name = tensor<string, []>("softmax_10_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_10 = softmax(axis = softmax_10_axis_0, x = add_10)[name = tensor<string, []>("softmax_10")]; |
| tensor<bool, []> out_41_transpose_x_0 = const()[name = tensor<string, []>("out_41_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> out_41_transpose_y_0 = const()[name = tensor<string, []>("out_41_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, 64]> out_41 = matmul(transpose_x = out_41_transpose_x_0, transpose_y = out_41_transpose_y_0, x = softmax_10, y = v_21)[name = tensor<string, []>("out_41")]; |
| tensor<int32, [2]> concat_129x = const()[name = tensor<string, []>("concat_129x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_1954 = reshape(shape = concat_129x, x = positions_21)[name = tensor<string, []>("op_1954")]; |
| tensor<string, []> var_1954_promoted_dtype_0 = const()[name = tensor<string, []>("op_1954_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_1954_promoted = cast(dtype = var_1954_promoted_dtype_0, x = var_1954)[name = tensor<string, []>("cast_422")]; |
| tensor<bool, [1, ?]> var_1957 = less(x = var_1954_promoted, y = var_1682)[name = tensor<string, []>("op_1957")]; |
| tensor<int32, [4]> concat_130x = const()[name = tensor<string, []>("concat_130x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q_21 = reshape(shape = concat_130x, x = var_1957)[name = tensor<string, []>("valid_q_21")]; |
| tensor<fp32, []> var_1961 = const()[name = tensor<string, []>("op_1961"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out_43 = select(a = out_41, b = var_1961, cond = valid_q_21)[name = tensor<string, []>("out_43")]; |
| tensor<int32, [4]> var_1963_perm_0 = const()[name = tensor<string, []>("op_1963_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_131x = const()[name = tensor<string, []>("concat_131x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_1963 = transpose(perm = var_1963_perm_0, x = out_43)[name = tensor<string, []>("transpose_51")]; |
| tensor<fp32, [1, ?, 256]> input_129 = reshape(shape = concat_131x, x = var_1963)[name = tensor<string, []>("input_129")]; |
| tensor<fp32, [1, ?, 256]> x_87 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_2_self_attn_out_proj_weight, x = input_129)[name = tensor<string, []>("linear_48")]; |
| tensor<fp32, [1, ?, 256]> var_1969 = mul(x = encoder_7_transformer_layers_2_layer_scale_1_scale, y = x_87)[name = tensor<string, []>("op_1969")]; |
| tensor<fp32, [1, ?, 256]> input_131 = add(x = input_127, y = var_1969)[name = tensor<string, []>("input_131")]; |
| tensor<int32, [1]> input_133_axes_0 = const()[name = tensor<string, []>("input_133_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_133 = layer_norm(axes = input_133_axes_0, beta = encoder_7_transformer_layers_2_norm2_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_2_norm2_weight, x = input_131)[name = tensor<string, []>("input_133")]; |
| tensor<fp32, [1, ?, 1024]> input_135 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_2_ffn_0_weight, x = input_133)[name = tensor<string, []>("linear_49")]; |
| tensor<string, []> input_137_mode_0 = const()[name = tensor<string, []>("input_137_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_137 = gelu(mode = input_137_mode_0, x = input_135)[name = tensor<string, []>("input_137")]; |
| tensor<fp32, [1, ?, 256]> x_89 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_2_ffn_2_weight, x = input_137)[name = tensor<string, []>("linear_50")]; |
| tensor<fp32, [1, ?, 256]> var_1984 = mul(x = encoder_7_transformer_layers_2_layer_scale_2_scale, y = x_89)[name = tensor<string, []>("op_1984")]; |
| tensor<fp32, [1, ?, 256]> input_139 = add(x = input_131, y = var_1984)[name = tensor<string, []>("input_139")]; |
| tensor<int32, [1]> x_91_axes_0 = const()[name = tensor<string, []>("x_91_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> x_91 = layer_norm(axes = x_91_axes_0, beta = encoder_7_transformer_layers_3_norm1_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_3_norm1_weight, x = input_139)[name = tensor<string, []>("x_91")]; |
| tensor<int32, [3]> var_1999_shape = shape(x = x_91)[name = tensor<string, []>("op_1999_shape")]; |
| tensor<int32, []> gather_120_batch_dims_0 = const()[name = tensor<string, []>("gather_120_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_120_validate_indices_0 = const()[name = tensor<string, []>("gather_120_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_38 = const()[name = tensor<string, []>("select_38"), val = tensor<int32, []>(1)]; |
| tensor<int32, []> gather_120_axis_1 = const()[name = tensor<string, []>("gather_120_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_120 = gather(axis = gather_120_axis_1, batch_dims = gather_120_batch_dims_0, indices = select_38, validate_indices = gather_120_validate_indices_0, x = var_1999_shape)[name = tensor<string, []>("gather_120")]; |
| tensor<fp32, [1, ?, 768]> projected_45 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_3_self_attn_in_proj_weight, x = x_91)[name = tensor<string, []>("linear_51")]; |
| tensor<int32, [5]> concat_132x = const()[name = tensor<string, []>("concat_132x"), val = tensor<int32, [5]>([1, -1, 3, 4, 64])]; |
| tensor<fp32, [1, ?, 3, 4, 64]> var_2006 = reshape(shape = concat_132x, x = projected_45)[name = tensor<string, []>("op_2006")]; |
| tensor<int32, [5]> var_2007 = const()[name = tensor<string, []>("op_2007"), val = tensor<int32, [5]>([2, 0, 3, 1, 4])]; |
| tensor<int32, [5]> q_45_begin_0 = const()[name = tensor<string, []>("q_45_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> q_45_end_0 = const()[name = tensor<string, []>("q_45_end_0"), val = tensor<int32, [5]>([1, 1, 4, 0, 64])]; |
| tensor<bool, [5]> q_45_end_mask_0 = const()[name = tensor<string, []>("q_45_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> q_45_squeeze_mask_0 = const()[name = tensor<string, []>("q_45_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [3, 1, 4, ?, 64]> projected = transpose(perm = var_2007, x = var_2006)[name = tensor<string, []>("transpose_50")]; |
| tensor<fp32, [1, 4, ?, 64]> q_45 = slice_by_index(begin = q_45_begin_0, end = q_45_end_0, end_mask = q_45_end_mask_0, squeeze_mask = q_45_squeeze_mask_0, x = projected)[name = tensor<string, []>("q_45")]; |
| tensor<int32, [5]> k_45_begin_0 = const()[name = tensor<string, []>("k_45_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])]; |
| tensor<int32, [5]> k_45_end_0 = const()[name = tensor<string, []>("k_45_end_0"), val = tensor<int32, [5]>([2, 1, 4, 0, 64])]; |
| tensor<bool, [5]> k_45_end_mask_0 = const()[name = tensor<string, []>("k_45_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> k_45_squeeze_mask_0 = const()[name = tensor<string, []>("k_45_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> k_45 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = projected)[name = tensor<string, []>("k_45")]; |
| tensor<int32, [5]> v_begin_0 = const()[name = tensor<string, []>("v_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])]; |
| tensor<int32, [5]> v_end_0 = const()[name = tensor<string, []>("v_end_0"), val = tensor<int32, [5]>([3, 1, 4, 0, 64])]; |
| tensor<bool, [5]> v_end_mask_0 = const()[name = tensor<string, []>("v_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])]; |
| tensor<bool, [5]> v_squeeze_mask_0 = const()[name = tensor<string, []>("v_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])]; |
| tensor<fp32, [1, 4, ?, 64]> v = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = projected)[name = tensor<string, []>("v")]; |
| tensor<fp32, [32]> freqs = const()[name = tensor<string, []>("freqs"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44433216)))]; |
| tensor<int32, [4]> var_2019_shape = shape(x = q_45)[name = tensor<string, []>("op_2019_shape")]; |
| tensor<int32, []> gather_124_batch_dims_0 = const()[name = tensor<string, []>("gather_124_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> gather_124_validate_indices_0 = const()[name = tensor<string, []>("gather_124_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> select_39 = const()[name = tensor<string, []>("select_39"), val = tensor<int32, []>(2)]; |
| tensor<int32, []> gather_124_axis_1 = const()[name = tensor<string, []>("gather_124_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> gather_124 = gather(axis = gather_124_axis_1, batch_dims = gather_124_batch_dims_0, indices = select_39, validate_indices = gather_124_validate_indices_0, x = var_2019_shape)[name = tensor<string, []>("gather_124")]; |
| tensor<fp32, [1, 1]> var_2022 = const()[name = tensor<string, []>("op_2022"), val = tensor<fp32, [1, 1]>([[0x0p+0]])]; |
| tensor<int32, []> const_102 = const()[name = tensor<string, []>("const_102"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_103 = const()[name = tensor<string, []>("const_103"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> var_2023 = range_1d(end = gather_124, start = const_102, step = const_103)[name = tensor<string, []>("op_2023")]; |
| tensor<string, []> var_2023_promoted_dtype_0 = const()[name = tensor<string, []>("op_2023_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [?]> var_2023_promoted = cast(dtype = var_2023_promoted_dtype_0, x = var_2023)[name = tensor<string, []>("cast_421")]; |
| tensor<fp32, [1, ?]> ts_45 = add(x = var_2022, y = var_2023_promoted)[name = tensor<string, []>("ts_45")]; |
| tensor<int32, [4]> var_2026 = const()[name = tensor<string, []>("op_2026"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<fp32, [1, 1, ?, 1]> ts = reshape(shape = var_2026, x = ts_45)[name = tensor<string, []>("ts")]; |
| tensor<int32, [5]> concat_133x = const()[name = tensor<string, []>("concat_133x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> q2 = reshape(shape = concat_133x, x = q_45)[name = tensor<string, []>("q2")]; |
| tensor<int32, [5]> concat_134x = const()[name = tensor<string, []>("concat_134x"), val = tensor<int32, [5]>([1, 4, -1, 32, 2])]; |
| tensor<fp32, [1, 4, ?, 32, 2]> k2 = reshape(shape = concat_134x, x = k_45)[name = tensor<string, []>("k2")]; |
| tensor<int32, [5]> qr_begin_0 = const()[name = tensor<string, []>("qr_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> qr_end_0 = const()[name = tensor<string, []>("qr_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> qr_end_mask_0 = const()[name = tensor<string, []>("qr_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qr_squeeze_mask_0 = const()[name = tensor<string, []>("qr_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qr = slice_by_index(begin = qr_begin_0, end = qr_end_0, end_mask = qr_end_mask_0, squeeze_mask = qr_squeeze_mask_0, x = q2)[name = tensor<string, []>("qr")]; |
| tensor<int32, [5]> qi_begin_0 = const()[name = tensor<string, []>("qi_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> qi_end_0 = const()[name = tensor<string, []>("qi_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> qi_end_mask_0 = const()[name = tensor<string, []>("qi_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> qi_squeeze_mask_0 = const()[name = tensor<string, []>("qi_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> qi = slice_by_index(begin = qi_begin_0, end = qi_end_0, end_mask = qi_end_mask_0, squeeze_mask = qi_squeeze_mask_0, x = q2)[name = tensor<string, []>("qi")]; |
| tensor<int32, [5]> kr_begin_0 = const()[name = tensor<string, []>("kr_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])]; |
| tensor<int32, [5]> kr_end_0 = const()[name = tensor<string, []>("kr_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 1])]; |
| tensor<bool, [5]> kr_end_mask_0 = const()[name = tensor<string, []>("kr_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> kr_squeeze_mask_0 = const()[name = tensor<string, []>("kr_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> kr = slice_by_index(begin = kr_begin_0, end = kr_end_0, end_mask = kr_end_mask_0, squeeze_mask = kr_squeeze_mask_0, x = k2)[name = tensor<string, []>("kr")]; |
| tensor<int32, [5]> ki_begin_0 = const()[name = tensor<string, []>("ki_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])]; |
| tensor<int32, [5]> ki_end_0 = const()[name = tensor<string, []>("ki_end_0"), val = tensor<int32, [5]>([1, 4, 0, 32, 2])]; |
| tensor<bool, [5]> ki_end_mask_0 = const()[name = tensor<string, []>("ki_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])]; |
| tensor<bool, [5]> ki_squeeze_mask_0 = const()[name = tensor<string, []>("ki_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])]; |
| tensor<fp32, [1, 4, ?, 32]> ki = slice_by_index(begin = ki_begin_0, end = ki_end_0, end_mask = ki_end_mask_0, squeeze_mask = ki_squeeze_mask_0, x = k2)[name = tensor<string, []>("ki")]; |
| tensor<fp32, [1, 1, ?, 32]> var_2039 = mul(x = freqs, y = ts)[name = tensor<string, []>("op_2039")]; |
| tensor<fp32, [1, 1, ?, 32]> rotr = cos(x = var_2039)[name = tensor<string, []>("rotr")]; |
| tensor<fp32, [1, 1, ?, 32]> roti = sin(x = var_2039)[name = tensor<string, []>("roti")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2043 = mul(x = qr, y = rotr)[name = tensor<string, []>("op_2043")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2044 = mul(x = qi, y = roti)[name = tensor<string, []>("op_2044")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2045 = sub(x = var_2043, y = var_2044)[name = tensor<string, []>("op_2045")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2046 = mul(x = qr, y = roti)[name = tensor<string, []>("op_2046")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2047 = mul(x = qi, y = rotr)[name = tensor<string, []>("op_2047")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2048 = add(x = var_2046, y = var_2047)[name = tensor<string, []>("op_2048")]; |
| tensor<int32, []> qo_axis_0 = const()[name = tensor<string, []>("qo_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> qo = stack(axis = qo_axis_0, values = (var_2045, var_2048))[name = tensor<string, []>("qo")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2051 = mul(x = kr, y = rotr)[name = tensor<string, []>("op_2051")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2052 = mul(x = ki, y = roti)[name = tensor<string, []>("op_2052")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2053 = sub(x = var_2051, y = var_2052)[name = tensor<string, []>("op_2053")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2054 = mul(x = kr, y = roti)[name = tensor<string, []>("op_2054")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2055 = mul(x = ki, y = rotr)[name = tensor<string, []>("op_2055")]; |
| tensor<fp32, [1, 4, ?, 32]> var_2056 = add(x = var_2054, y = var_2055)[name = tensor<string, []>("op_2056")]; |
| tensor<int32, []> ko_axis_0 = const()[name = tensor<string, []>("ko_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, 32, 2]> ko = stack(axis = ko_axis_0, values = (var_2053, var_2056))[name = tensor<string, []>("ko")]; |
| tensor<int32, [4]> concat_135x = const()[name = tensor<string, []>("concat_135x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> q = reshape(shape = concat_135x, x = qo)[name = tensor<string, []>("q")]; |
| tensor<int32, [4]> concat_136x = const()[name = tensor<string, []>("concat_136x"), val = tensor<int32, [4]>([1, 4, -1, 64])]; |
| tensor<fp32, [1, 4, ?, 64]> k = reshape(shape = concat_136x, x = ko)[name = tensor<string, []>("k")]; |
| tensor<int32, []> const_104 = const()[name = tensor<string, []>("const_104"), val = tensor<int32, []>(0)]; |
| tensor<int32, []> const_105 = const()[name = tensor<string, []>("const_105"), val = tensor<int32, []>(1)]; |
| tensor<int32, [?]> positions = range_1d(end = gather_120, start = const_104, step = const_105)[name = tensor<string, []>("positions")]; |
| tensor<int32, [3]> concat_137x = const()[name = tensor<string, []>("concat_137x"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<int32, [1, 1, ?]> var_2068 = reshape(shape = concat_137x, x = positions)[name = tensor<string, []>("op_2068")]; |
| tensor<string, []> var_2068_promoted_dtype_0 = const()[name = tensor<string, []>("op_2068_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, 1, ?]> var_2068_promoted = cast(dtype = var_2068_promoted_dtype_0, x = var_2068)[name = tensor<string, []>("cast_420")]; |
| tensor<bool, [1, 1, ?]> valid_k = less(x = var_2068_promoted, y = var_1659)[name = tensor<string, []>("valid_k")]; |
| tensor<int32, [3]> concat_138x = const()[name = tensor<string, []>("concat_138x"), val = tensor<int32, [3]>([1, -1, 1])]; |
| tensor<int32, [1, ?, 1]> var_2073 = reshape(shape = concat_138x, x = positions)[name = tensor<string, []>("op_2073")]; |
| tensor<int32, [1, ?, ?]> delta = sub(x = var_2073, y = var_2068)[name = tensor<string, []>("delta")]; |
| tensor<int32, []> concat_140_axis_0 = const()[name = tensor<string, []>("concat_140_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> concat_140_interleave_0 = const()[name = tensor<string, []>("concat_140_interleave_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, [3]> concat_140 = concat(axis = concat_140_axis_0, interleave = concat_140_interleave_0, values = (var_1560, gather_120, gather_120))[name = tensor<string, []>("concat_140")]; |
| tensor<fp32, []> fill_11_value_0 = const()[name = tensor<string, []>("fill_11_value_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, ?, ?]> fill_11 = fill(shape = concat_140, value = fill_11_value_0)[name = tensor<string, []>("fill_11")]; |
| tensor<string, []> attn_bias_89_dtype_0 = const()[name = tensor<string, []>("attn_bias_89_dtype_0"), val = tensor<string, []>("bool")]; |
| tensor<bool, [1, ?, ?]> var_2079 = greater_equal(x = delta, y = var_1542)[name = tensor<string, []>("op_2079")]; |
| tensor<bool, [1, ?, ?]> attn_bias_89 = cast(dtype = attn_bias_89_dtype_0, x = fill_11)[name = tensor<string, []>("cast_419")]; |
| tensor<bool, [1, ?, ?]> attn_bias_91 = logical_and(x = attn_bias_89, y = var_2079)[name = tensor<string, []>("attn_bias_91")]; |
| tensor<bool, [1, ?, ?]> var_2081 = less(x = delta, y = var_1549)[name = tensor<string, []>("op_2081")]; |
| tensor<bool, [1, ?, ?]> attn_bias_93 = logical_and(x = attn_bias_91, y = var_2081)[name = tensor<string, []>("attn_bias_93")]; |
| tensor<bool, [1, ?, ?]> var_2083 = logical_and(x = attn_bias_93, y = valid_k)[name = tensor<string, []>("op_2083")]; |
| tensor<int32, [1]> var_2085_axes_0 = const()[name = tensor<string, []>("op_2085_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<bool, [1, 1, ?, ?]> var_2085 = expand_dims(axes = var_2085_axes_0, x = var_2083)[name = tensor<string, []>("op_2085")]; |
| tensor<string, []> cast_268_dtype_0 = const()[name = tensor<string, []>("cast_268_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, []> sub_11_x_0 = const()[name = tensor<string, []>("sub_11_x_0"), val = tensor<fp32, []>(0x1p+0)]; |
| tensor<fp32, [1, 1, ?, ?]> cast_268 = cast(dtype = cast_268_dtype_0, x = var_2085)[name = tensor<string, []>("cast_418")]; |
| tensor<fp32, [1, 1, ?, ?]> sub_11 = sub(x = sub_11_x_0, y = cast_268)[name = tensor<string, []>("sub_11")]; |
| tensor<fp32, []> mul_22_x_0 = const()[name = tensor<string, []>("mul_22_x_0"), val = tensor<fp32, []>(-0x1.d4cp+14)]; |
| tensor<fp32, [1, 1, ?, ?]> mul_22 = mul(x = mul_22_x_0, y = sub_11)[name = tensor<string, []>("mul_22")]; |
| tensor<fp32, []> mul_23_y_0 = const()[name = tensor<string, []>("mul_23_y_0"), val = tensor<fp32, []>(0x1p-3)]; |
| tensor<fp32, [1, 4, ?, 64]> mul_23 = mul(x = q, y = mul_23_y_0)[name = tensor<string, []>("mul_23")]; |
| tensor<bool, []> matmul_11_transpose_y_0 = const()[name = tensor<string, []>("matmul_11_transpose_y_0"), val = tensor<bool, []>(true)]; |
| tensor<bool, []> matmul_11_transpose_x_0 = const()[name = tensor<string, []>("matmul_11_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, ?]> matmul_11 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23, y = k)[name = tensor<string, []>("matmul_11")]; |
| tensor<fp32, [1, 4, ?, ?]> add_11 = add(x = matmul_11, y = mul_22)[name = tensor<string, []>("add_11")]; |
| tensor<int32, []> softmax_11_axis_0 = const()[name = tensor<string, []>("softmax_11_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<fp32, [1, 4, ?, ?]> softmax_11 = softmax(axis = softmax_11_axis_0, x = add_11)[name = tensor<string, []>("softmax_11")]; |
| tensor<bool, []> out_45_transpose_x_0 = const()[name = tensor<string, []>("out_45_transpose_x_0"), val = tensor<bool, []>(false)]; |
| tensor<bool, []> out_45_transpose_y_0 = const()[name = tensor<string, []>("out_45_transpose_y_0"), val = tensor<bool, []>(false)]; |
| tensor<fp32, [1, 4, ?, 64]> out_45 = matmul(transpose_x = out_45_transpose_x_0, transpose_y = out_45_transpose_y_0, x = softmax_11, y = v)[name = tensor<string, []>("out_45")]; |
| tensor<int32, [2]> concat_141x = const()[name = tensor<string, []>("concat_141x"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2091 = reshape(shape = concat_141x, x = positions)[name = tensor<string, []>("op_2091")]; |
| tensor<string, []> var_2091_promoted_dtype_0 = const()[name = tensor<string, []>("op_2091_promoted_dtype_0"), val = tensor<string, []>("fp32")]; |
| tensor<fp32, [1, ?]> var_2091_promoted = cast(dtype = var_2091_promoted_dtype_0, x = var_2091)[name = tensor<string, []>("cast_417")]; |
| tensor<bool, [1, ?]> var_2094 = less(x = var_2091_promoted, y = var_1682)[name = tensor<string, []>("op_2094")]; |
| tensor<int32, [4]> concat_142x = const()[name = tensor<string, []>("concat_142x"), val = tensor<int32, [4]>([1, 1, -1, 1])]; |
| tensor<bool, [1, 1, ?, 1]> valid_q = reshape(shape = concat_142x, x = var_2094)[name = tensor<string, []>("valid_q")]; |
| tensor<fp32, []> var_2098 = const()[name = tensor<string, []>("op_2098"), val = tensor<fp32, []>(0x0p+0)]; |
| tensor<fp32, [1, 4, ?, 64]> out = select(a = out_45, b = var_2098, cond = valid_q)[name = tensor<string, []>("out")]; |
| tensor<int32, [4]> var_2100_perm_0 = const()[name = tensor<string, []>("op_2100_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> concat_143x = const()[name = tensor<string, []>("concat_143x"), val = tensor<int32, [3]>([1, -1, 256])]; |
| tensor<fp32, [1, ?, 4, 64]> var_2100 = transpose(perm = var_2100_perm_0, x = out)[name = tensor<string, []>("transpose_49")]; |
| tensor<fp32, [1, ?, 256]> input_141 = reshape(shape = concat_143x, x = var_2100)[name = tensor<string, []>("input_141")]; |
| tensor<fp32, [1, ?, 256]> x_93 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_3_self_attn_out_proj_weight, x = input_141)[name = tensor<string, []>("linear_52")]; |
| tensor<fp32, [1, ?, 256]> var_2106 = mul(x = encoder_7_transformer_layers_3_layer_scale_1_scale, y = x_93)[name = tensor<string, []>("op_2106")]; |
| tensor<fp32, [1, ?, 256]> input_143 = add(x = input_139, y = var_2106)[name = tensor<string, []>("input_143")]; |
| tensor<int32, [1]> input_145_axes_0 = const()[name = tensor<string, []>("input_145_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 256]> input_145 = layer_norm(axes = input_145_axes_0, beta = encoder_7_transformer_layers_3_norm2_bias, epsilon = var_1556, gamma = encoder_7_transformer_layers_3_norm2_weight, x = input_143)[name = tensor<string, []>("input_145")]; |
| tensor<fp32, [1, ?, 1024]> input_147 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_3_ffn_0_weight, x = input_145)[name = tensor<string, []>("linear_53")]; |
| tensor<string, []> input_149_mode_0 = const()[name = tensor<string, []>("input_149_mode_0"), val = tensor<string, []>("EXACT")]; |
| tensor<fp32, [1, ?, 1024]> input_149 = gelu(mode = input_149_mode_0, x = input_147)[name = tensor<string, []>("input_149")]; |
| tensor<fp32, [1, ?, 256]> x_95 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_3_ffn_2_weight, x = input_149)[name = tensor<string, []>("linear_54")]; |
| tensor<fp32, [1, ?, 256]> var_2121 = mul(x = encoder_7_transformer_layers_3_layer_scale_2_scale, y = x_95)[name = tensor<string, []>("op_2121")]; |
| tensor<fp32, [1, ?, 256]> input_151 = add(x = input_143, y = var_2121)[name = tensor<string, []>("input_151")]; |
| tensor<fp32, [192]> linear_55_bias_0 = const()[name = tensor<string, []>("linear_55_bias_0"), val = tensor<fp32, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44433408)))]; |
| tensor<fp32, [1, ?, 192]> var_2124 = linear(bias = linear_55_bias_0, weight = encoder_7_output_proj_weight, x = input_151)[name = tensor<string, []>("linear_55")]; |
| tensor<int32, [3]> x_perm_0 = const()[name = tensor<string, []>("x_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [4]> var_2135 = const()[name = tensor<string, []>("op_2135"), val = tensor<int32, [4]>([1, 192, -1, 4])]; |
| tensor<fp32, [1, 192, ?]> x = transpose(perm = x_perm_0, x = var_2124)[name = tensor<string, []>("transpose_48")]; |
| tensor<fp32, [1, 192, ?, 4]> var_2136 = reshape(shape = var_2135, x = x)[name = tensor<string, []>("op_2136")]; |
| tensor<int32, [4]> var_2137 = const()[name = tensor<string, []>("op_2137"), val = tensor<int32, [4]>([0, 1, 3, 2])]; |
| tensor<int32, [3]> var_2142 = const()[name = tensor<string, []>("op_2142"), val = tensor<int32, [3]>([1, 768, -1])]; |
| tensor<fp32, [1, 192, 4, ?]> var_2138 = transpose(perm = var_2137, x = var_2136)[name = tensor<string, []>("transpose_47")]; |
| tensor<fp32, [1, 768, ?]> input_153 = reshape(shape = var_2142, x = var_2138)[name = tensor<string, []>("input_153")]; |
| tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 512, ?]> input_155 = conv(bias = input_proj_bias, dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = input_proj_weight, x = input_153)[name = tensor<string, []>("input_155")]; |
| tensor<string, []> z_e_1_pad_type_0 = const()[name = tensor<string, []>("z_e_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_1_strides_0 = const()[name = tensor<string, []>("z_e_1_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_1_pad_0 = const()[name = tensor<string, []>("z_e_1_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_1_dilations_0 = const()[name = tensor<string, []>("z_e_1_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_1_groups_0 = const()[name = tensor<string, []>("z_e_1_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_1 = conv(bias = quantizers_0_in_proj_bias, dilations = z_e_1_dilations_0, groups = z_e_1_groups_0, pad = z_e_1_pad_0, pad_type = z_e_1_pad_type_0, strides = z_e_1_strides_0, weight = quantizers_0_in_proj_weight, x = input_155)[name = tensor<string, []>("z_e_1")]; |
| tensor<int32, [3]> var_2168_perm_0 = const()[name = tensor<string, []>("op_2168_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2174 = const()[name = tensor<string, []>("op_2174"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2168 = transpose(perm = var_2168_perm_0, x = z_e_1)[name = tensor<string, []>("transpose_46")]; |
| tensor<fp32, [?, 8]> input_157 = reshape(shape = var_2174, x = var_2168)[name = tensor<string, []>("input_157")]; |
| tensor<int32, [1]> var_2178 = const()[name = tensor<string, []>("op_2178"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2179 = const()[name = tensor<string, []>("op_2179"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2181 = reduce_l2_norm(axes = var_2178, keep_dims = var_2179, x = input_157)[name = tensor<string, []>("op_2181")]; |
| tensor<fp32, []> var_2182 = const()[name = tensor<string, []>("op_2182"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2183 = maximum(x = var_2181, y = var_2182)[name = tensor<string, []>("op_2183")]; |
| tensor<int32, [2]> shape_0 = shape(x = input_157)[name = tensor<string, []>("shape_0")]; |
| tensor<int32, [2]> shape_1 = shape(x = var_2183)[name = tensor<string, []>("shape_1")]; |
| tensor<int32, []> equal_0_y_0 = const()[name = tensor<string, []>("equal_0_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_0 = equal(x = shape_0, y = equal_0_y_0)[name = tensor<string, []>("equal_0")]; |
| tensor<int32, [2]> select_0 = select(a = shape_1, b = shape_0, cond = equal_0)[name = tensor<string, []>("select_0")]; |
| tensor<int32, [2]> real_div_0 = real_div(x = select_0, y = shape_1)[name = tensor<string, []>("real_div_0")]; |
| tensor<fp32, [?, ?]> denom_1 = tile(reps = real_div_0, x = var_2183)[name = tensor<string, []>("denom_1")]; |
| tensor<fp32, [?, 8]> enc_1 = real_div(x = input_157, y = denom_1)[name = tensor<string, []>("enc_1")]; |
| tensor<fp32, [1024, 8]> cb_1 = const()[name = tensor<string, []>("cb_1"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44434240)))]; |
| tensor<fp32, [?, 1024]> sim_1 = linear(bias = linear_3_bias_0, weight = cb_1, x = enc_1)[name = tensor<string, []>("sim_1")]; |
| tensor<int32, []> input_161_axis_0 = const()[name = tensor<string, []>("input_161_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_161_keep_dims_0 = const()[name = tensor<string, []>("input_161_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_161_output_dtype_0 = const()[name = tensor<string, []>("input_161_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_161 = reduce_argmax(axis = input_161_axis_0, keep_dims = input_161_keep_dims_0, output_dtype = input_161_output_dtype_0, x = sim_1)[name = tensor<string, []>("input_161")]; |
| tensor<int32, []> var_2204_batch_dims_0 = const()[name = tensor<string, []>("op_2204_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2204_validate_indices_0 = const()[name = tensor<string, []>("op_2204_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_24_y_0 = const()[name = tensor<string, []>("greater_equal_24_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_24 = greater_equal(x = input_161, y = greater_equal_24_y_0)[name = tensor<string, []>("greater_equal_24")]; |
| tensor<int32, []> slice_by_index_24 = const()[name = tensor<string, []>("slice_by_index_24"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_36 = add(x = input_161, y = slice_by_index_24)[name = tensor<string, []>("add_36")]; |
| tensor<int32, [?]> select_40 = select(a = input_161, b = add_36, cond = greater_equal_24)[name = tensor<string, []>("select_40")]; |
| tensor<int32, []> var_2204_axis_1 = const()[name = tensor<string, []>("op_2204_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2204 = gather(axis = var_2204_axis_1, batch_dims = var_2204_batch_dims_0, indices = select_40, validate_indices = var_2204_validate_indices_0, x = quantizers_0_codebook_weight)[name = tensor<string, []>("op_2204")]; |
| tensor<int32, [3]> var_2214 = const()[name = tensor<string, []>("op_2214"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2215 = reshape(shape = var_2214, x = var_2204)[name = tensor<string, []>("op_2215")]; |
| tensor<int32, [3]> input_163_perm_0 = const()[name = tensor<string, []>("input_163_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_1_pad_type_0 = const()[name = tensor<string, []>("z_q_1_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_1_strides_0 = const()[name = tensor<string, []>("z_q_1_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_1_pad_0 = const()[name = tensor<string, []>("z_q_1_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_1_dilations_0 = const()[name = tensor<string, []>("z_q_1_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_1_groups_0 = const()[name = tensor<string, []>("z_q_1_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_163 = transpose(perm = input_163_perm_0, x = var_2215)[name = tensor<string, []>("transpose_45")]; |
| tensor<fp32, [1, 512, ?]> z_q_1 = conv(bias = quantizers_0_out_proj_bias, dilations = z_q_1_dilations_0, groups = z_q_1_groups_0, pad = z_q_1_pad_0, pad_type = z_q_1_pad_type_0, strides = z_q_1_strides_0, weight = quantizers_0_out_proj_weight, x = input_163)[name = tensor<string, []>("z_q_1")]; |
| tensor<fp32, [1, 512, ?]> input_165 = sub(x = input_155, y = z_q_1)[name = tensor<string, []>("input_165")]; |
| tensor<int32, [2]> var_2237 = const()[name = tensor<string, []>("op_2237"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2238 = reshape(shape = var_2237, x = input_161)[name = tensor<string, []>("op_2238")]; |
| tensor<string, []> z_e_3_pad_type_0 = const()[name = tensor<string, []>("z_e_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_3_strides_0 = const()[name = tensor<string, []>("z_e_3_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_3_pad_0 = const()[name = tensor<string, []>("z_e_3_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_3_dilations_0 = const()[name = tensor<string, []>("z_e_3_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_3_groups_0 = const()[name = tensor<string, []>("z_e_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_3 = conv(bias = quantizers_1_in_proj_bias, dilations = z_e_3_dilations_0, groups = z_e_3_groups_0, pad = z_e_3_pad_0, pad_type = z_e_3_pad_type_0, strides = z_e_3_strides_0, weight = quantizers_1_in_proj_weight, x = input_165)[name = tensor<string, []>("z_e_3")]; |
| tensor<int32, [3]> var_2252_perm_0 = const()[name = tensor<string, []>("op_2252_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2258 = const()[name = tensor<string, []>("op_2258"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2252 = transpose(perm = var_2252_perm_0, x = z_e_3)[name = tensor<string, []>("transpose_44")]; |
| tensor<fp32, [?, 8]> input_167 = reshape(shape = var_2258, x = var_2252)[name = tensor<string, []>("input_167")]; |
| tensor<int32, [1]> var_2262 = const()[name = tensor<string, []>("op_2262"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2263 = const()[name = tensor<string, []>("op_2263"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2265 = reduce_l2_norm(axes = var_2262, keep_dims = var_2263, x = input_167)[name = tensor<string, []>("op_2265")]; |
| tensor<fp32, []> var_2266 = const()[name = tensor<string, []>("op_2266"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2267 = maximum(x = var_2265, y = var_2266)[name = tensor<string, []>("op_2267")]; |
| tensor<int32, [2]> shape_3 = shape(x = input_167)[name = tensor<string, []>("shape_3")]; |
| tensor<int32, [2]> shape_4 = shape(x = var_2267)[name = tensor<string, []>("shape_4")]; |
| tensor<int32, []> equal_1_y_0 = const()[name = tensor<string, []>("equal_1_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_1 = equal(x = shape_3, y = equal_1_y_0)[name = tensor<string, []>("equal_1")]; |
| tensor<int32, [2]> select_1 = select(a = shape_4, b = shape_3, cond = equal_1)[name = tensor<string, []>("select_1")]; |
| tensor<int32, [2]> real_div_1 = real_div(x = select_1, y = shape_4)[name = tensor<string, []>("real_div_1")]; |
| tensor<fp32, [?, ?]> denom_5 = tile(reps = real_div_1, x = var_2267)[name = tensor<string, []>("denom_5")]; |
| tensor<fp32, [?, 8]> enc_3 = real_div(x = input_167, y = denom_5)[name = tensor<string, []>("enc_3")]; |
| tensor<fp32, [1024, 8]> cb_3 = const()[name = tensor<string, []>("cb_3"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44467072)))]; |
| tensor<fp32, [?, 1024]> sim_3 = linear(bias = linear_3_bias_0, weight = cb_3, x = enc_3)[name = tensor<string, []>("sim_3")]; |
| tensor<int32, []> input_171_axis_0 = const()[name = tensor<string, []>("input_171_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_171_keep_dims_0 = const()[name = tensor<string, []>("input_171_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_171_output_dtype_0 = const()[name = tensor<string, []>("input_171_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_171 = reduce_argmax(axis = input_171_axis_0, keep_dims = input_171_keep_dims_0, output_dtype = input_171_output_dtype_0, x = sim_3)[name = tensor<string, []>("input_171")]; |
| tensor<int32, []> var_2288_batch_dims_0 = const()[name = tensor<string, []>("op_2288_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2288_validate_indices_0 = const()[name = tensor<string, []>("op_2288_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_25_y_0 = const()[name = tensor<string, []>("greater_equal_25_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_25 = greater_equal(x = input_171, y = greater_equal_25_y_0)[name = tensor<string, []>("greater_equal_25")]; |
| tensor<int32, []> slice_by_index_25 = const()[name = tensor<string, []>("slice_by_index_25"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_37 = add(x = input_171, y = slice_by_index_25)[name = tensor<string, []>("add_37")]; |
| tensor<int32, [?]> select_41 = select(a = input_171, b = add_37, cond = greater_equal_25)[name = tensor<string, []>("select_41")]; |
| tensor<int32, []> var_2288_axis_1 = const()[name = tensor<string, []>("op_2288_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2288 = gather(axis = var_2288_axis_1, batch_dims = var_2288_batch_dims_0, indices = select_41, validate_indices = var_2288_validate_indices_0, x = quantizers_1_codebook_weight)[name = tensor<string, []>("op_2288")]; |
| tensor<int32, [3]> var_2298 = const()[name = tensor<string, []>("op_2298"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2299 = reshape(shape = var_2298, x = var_2288)[name = tensor<string, []>("op_2299")]; |
| tensor<int32, [3]> input_173_perm_0 = const()[name = tensor<string, []>("input_173_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_3_pad_type_0 = const()[name = tensor<string, []>("z_q_3_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_3_strides_0 = const()[name = tensor<string, []>("z_q_3_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_3_pad_0 = const()[name = tensor<string, []>("z_q_3_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_3_dilations_0 = const()[name = tensor<string, []>("z_q_3_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_3_groups_0 = const()[name = tensor<string, []>("z_q_3_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_173 = transpose(perm = input_173_perm_0, x = var_2299)[name = tensor<string, []>("transpose_43")]; |
| tensor<fp32, [1, 512, ?]> z_q_3 = conv(bias = quantizers_1_out_proj_bias, dilations = z_q_3_dilations_0, groups = z_q_3_groups_0, pad = z_q_3_pad_0, pad_type = z_q_3_pad_type_0, strides = z_q_3_strides_0, weight = quantizers_1_out_proj_weight, x = input_173)[name = tensor<string, []>("z_q_3")]; |
| tensor<fp32, [1, 512, ?]> input_175 = sub(x = input_165, y = z_q_3)[name = tensor<string, []>("input_175")]; |
| tensor<int32, [2]> var_2321 = const()[name = tensor<string, []>("op_2321"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2322 = reshape(shape = var_2321, x = input_171)[name = tensor<string, []>("op_2322")]; |
| tensor<string, []> z_e_5_pad_type_0 = const()[name = tensor<string, []>("z_e_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_5_strides_0 = const()[name = tensor<string, []>("z_e_5_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_5_pad_0 = const()[name = tensor<string, []>("z_e_5_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_5_dilations_0 = const()[name = tensor<string, []>("z_e_5_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_5_groups_0 = const()[name = tensor<string, []>("z_e_5_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_5 = conv(bias = quantizers_2_in_proj_bias, dilations = z_e_5_dilations_0, groups = z_e_5_groups_0, pad = z_e_5_pad_0, pad_type = z_e_5_pad_type_0, strides = z_e_5_strides_0, weight = quantizers_2_in_proj_weight, x = input_175)[name = tensor<string, []>("z_e_5")]; |
| tensor<int32, [3]> var_2336_perm_0 = const()[name = tensor<string, []>("op_2336_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2342 = const()[name = tensor<string, []>("op_2342"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2336 = transpose(perm = var_2336_perm_0, x = z_e_5)[name = tensor<string, []>("transpose_42")]; |
| tensor<fp32, [?, 8]> input_177 = reshape(shape = var_2342, x = var_2336)[name = tensor<string, []>("input_177")]; |
| tensor<int32, [1]> var_2346 = const()[name = tensor<string, []>("op_2346"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2347 = const()[name = tensor<string, []>("op_2347"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2349 = reduce_l2_norm(axes = var_2346, keep_dims = var_2347, x = input_177)[name = tensor<string, []>("op_2349")]; |
| tensor<fp32, []> var_2350 = const()[name = tensor<string, []>("op_2350"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2351 = maximum(x = var_2349, y = var_2350)[name = tensor<string, []>("op_2351")]; |
| tensor<int32, [2]> shape_6 = shape(x = input_177)[name = tensor<string, []>("shape_6")]; |
| tensor<int32, [2]> shape_7 = shape(x = var_2351)[name = tensor<string, []>("shape_7")]; |
| tensor<int32, []> equal_2_y_0 = const()[name = tensor<string, []>("equal_2_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_2 = equal(x = shape_6, y = equal_2_y_0)[name = tensor<string, []>("equal_2")]; |
| tensor<int32, [2]> select_2 = select(a = shape_7, b = shape_6, cond = equal_2)[name = tensor<string, []>("select_2")]; |
| tensor<int32, [2]> real_div_2 = real_div(x = select_2, y = shape_7)[name = tensor<string, []>("real_div_2")]; |
| tensor<fp32, [?, ?]> denom_9 = tile(reps = real_div_2, x = var_2351)[name = tensor<string, []>("denom_9")]; |
| tensor<fp32, [?, 8]> enc_5 = real_div(x = input_177, y = denom_9)[name = tensor<string, []>("enc_5")]; |
| tensor<fp32, [1024, 8]> cb_5 = const()[name = tensor<string, []>("cb_5"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44499904)))]; |
| tensor<fp32, [?, 1024]> sim_5 = linear(bias = linear_3_bias_0, weight = cb_5, x = enc_5)[name = tensor<string, []>("sim_5")]; |
| tensor<int32, []> input_181_axis_0 = const()[name = tensor<string, []>("input_181_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_181_keep_dims_0 = const()[name = tensor<string, []>("input_181_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_181_output_dtype_0 = const()[name = tensor<string, []>("input_181_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_181 = reduce_argmax(axis = input_181_axis_0, keep_dims = input_181_keep_dims_0, output_dtype = input_181_output_dtype_0, x = sim_5)[name = tensor<string, []>("input_181")]; |
| tensor<int32, []> var_2372_batch_dims_0 = const()[name = tensor<string, []>("op_2372_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2372_validate_indices_0 = const()[name = tensor<string, []>("op_2372_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_26_y_0 = const()[name = tensor<string, []>("greater_equal_26_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_26 = greater_equal(x = input_181, y = greater_equal_26_y_0)[name = tensor<string, []>("greater_equal_26")]; |
| tensor<int32, []> slice_by_index_26 = const()[name = tensor<string, []>("slice_by_index_26"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_38 = add(x = input_181, y = slice_by_index_26)[name = tensor<string, []>("add_38")]; |
| tensor<int32, [?]> select_42 = select(a = input_181, b = add_38, cond = greater_equal_26)[name = tensor<string, []>("select_42")]; |
| tensor<int32, []> var_2372_axis_1 = const()[name = tensor<string, []>("op_2372_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2372 = gather(axis = var_2372_axis_1, batch_dims = var_2372_batch_dims_0, indices = select_42, validate_indices = var_2372_validate_indices_0, x = quantizers_2_codebook_weight)[name = tensor<string, []>("op_2372")]; |
| tensor<int32, [3]> var_2382 = const()[name = tensor<string, []>("op_2382"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2383 = reshape(shape = var_2382, x = var_2372)[name = tensor<string, []>("op_2383")]; |
| tensor<int32, [3]> input_183_perm_0 = const()[name = tensor<string, []>("input_183_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_5_pad_type_0 = const()[name = tensor<string, []>("z_q_5_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_5_strides_0 = const()[name = tensor<string, []>("z_q_5_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_5_pad_0 = const()[name = tensor<string, []>("z_q_5_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_5_dilations_0 = const()[name = tensor<string, []>("z_q_5_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_5_groups_0 = const()[name = tensor<string, []>("z_q_5_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_183 = transpose(perm = input_183_perm_0, x = var_2383)[name = tensor<string, []>("transpose_41")]; |
| tensor<fp32, [1, 512, ?]> z_q_5 = conv(bias = quantizers_2_out_proj_bias, dilations = z_q_5_dilations_0, groups = z_q_5_groups_0, pad = z_q_5_pad_0, pad_type = z_q_5_pad_type_0, strides = z_q_5_strides_0, weight = quantizers_2_out_proj_weight, x = input_183)[name = tensor<string, []>("z_q_5")]; |
| tensor<fp32, [1, 512, ?]> input_185 = sub(x = input_175, y = z_q_5)[name = tensor<string, []>("input_185")]; |
| tensor<int32, [2]> var_2405 = const()[name = tensor<string, []>("op_2405"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2406 = reshape(shape = var_2405, x = input_181)[name = tensor<string, []>("op_2406")]; |
| tensor<string, []> z_e_7_pad_type_0 = const()[name = tensor<string, []>("z_e_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_7_strides_0 = const()[name = tensor<string, []>("z_e_7_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_7_pad_0 = const()[name = tensor<string, []>("z_e_7_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_7_dilations_0 = const()[name = tensor<string, []>("z_e_7_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_7_groups_0 = const()[name = tensor<string, []>("z_e_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_7 = conv(bias = quantizers_3_in_proj_bias, dilations = z_e_7_dilations_0, groups = z_e_7_groups_0, pad = z_e_7_pad_0, pad_type = z_e_7_pad_type_0, strides = z_e_7_strides_0, weight = quantizers_3_in_proj_weight, x = input_185)[name = tensor<string, []>("z_e_7")]; |
| tensor<int32, [3]> var_2420_perm_0 = const()[name = tensor<string, []>("op_2420_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2426 = const()[name = tensor<string, []>("op_2426"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2420 = transpose(perm = var_2420_perm_0, x = z_e_7)[name = tensor<string, []>("transpose_40")]; |
| tensor<fp32, [?, 8]> input_187 = reshape(shape = var_2426, x = var_2420)[name = tensor<string, []>("input_187")]; |
| tensor<int32, [1]> var_2430 = const()[name = tensor<string, []>("op_2430"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2431 = const()[name = tensor<string, []>("op_2431"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2433 = reduce_l2_norm(axes = var_2430, keep_dims = var_2431, x = input_187)[name = tensor<string, []>("op_2433")]; |
| tensor<fp32, []> var_2434 = const()[name = tensor<string, []>("op_2434"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2435 = maximum(x = var_2433, y = var_2434)[name = tensor<string, []>("op_2435")]; |
| tensor<int32, [2]> shape_9 = shape(x = input_187)[name = tensor<string, []>("shape_9")]; |
| tensor<int32, [2]> shape_10 = shape(x = var_2435)[name = tensor<string, []>("shape_10")]; |
| tensor<int32, []> equal_3_y_0 = const()[name = tensor<string, []>("equal_3_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_3 = equal(x = shape_9, y = equal_3_y_0)[name = tensor<string, []>("equal_3")]; |
| tensor<int32, [2]> select_3 = select(a = shape_10, b = shape_9, cond = equal_3)[name = tensor<string, []>("select_3")]; |
| tensor<int32, [2]> real_div_3 = real_div(x = select_3, y = shape_10)[name = tensor<string, []>("real_div_3")]; |
| tensor<fp32, [?, ?]> denom_13 = tile(reps = real_div_3, x = var_2435)[name = tensor<string, []>("denom_13")]; |
| tensor<fp32, [?, 8]> enc_7 = real_div(x = input_187, y = denom_13)[name = tensor<string, []>("enc_7")]; |
| tensor<fp32, [1024, 8]> cb_7 = const()[name = tensor<string, []>("cb_7"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44532736)))]; |
| tensor<fp32, [?, 1024]> sim_7 = linear(bias = linear_3_bias_0, weight = cb_7, x = enc_7)[name = tensor<string, []>("sim_7")]; |
| tensor<int32, []> input_191_axis_0 = const()[name = tensor<string, []>("input_191_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_191_keep_dims_0 = const()[name = tensor<string, []>("input_191_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_191_output_dtype_0 = const()[name = tensor<string, []>("input_191_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_191 = reduce_argmax(axis = input_191_axis_0, keep_dims = input_191_keep_dims_0, output_dtype = input_191_output_dtype_0, x = sim_7)[name = tensor<string, []>("input_191")]; |
| tensor<int32, []> var_2456_batch_dims_0 = const()[name = tensor<string, []>("op_2456_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2456_validate_indices_0 = const()[name = tensor<string, []>("op_2456_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_27_y_0 = const()[name = tensor<string, []>("greater_equal_27_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_27 = greater_equal(x = input_191, y = greater_equal_27_y_0)[name = tensor<string, []>("greater_equal_27")]; |
| tensor<int32, []> slice_by_index_27 = const()[name = tensor<string, []>("slice_by_index_27"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_39 = add(x = input_191, y = slice_by_index_27)[name = tensor<string, []>("add_39")]; |
| tensor<int32, [?]> select_43 = select(a = input_191, b = add_39, cond = greater_equal_27)[name = tensor<string, []>("select_43")]; |
| tensor<int32, []> var_2456_axis_1 = const()[name = tensor<string, []>("op_2456_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2456 = gather(axis = var_2456_axis_1, batch_dims = var_2456_batch_dims_0, indices = select_43, validate_indices = var_2456_validate_indices_0, x = quantizers_3_codebook_weight)[name = tensor<string, []>("op_2456")]; |
| tensor<int32, [3]> var_2466 = const()[name = tensor<string, []>("op_2466"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2467 = reshape(shape = var_2466, x = var_2456)[name = tensor<string, []>("op_2467")]; |
| tensor<int32, [3]> input_193_perm_0 = const()[name = tensor<string, []>("input_193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_7_pad_type_0 = const()[name = tensor<string, []>("z_q_7_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_7_strides_0 = const()[name = tensor<string, []>("z_q_7_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_7_pad_0 = const()[name = tensor<string, []>("z_q_7_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_7_dilations_0 = const()[name = tensor<string, []>("z_q_7_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_7_groups_0 = const()[name = tensor<string, []>("z_q_7_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_193 = transpose(perm = input_193_perm_0, x = var_2467)[name = tensor<string, []>("transpose_39")]; |
| tensor<fp32, [1, 512, ?]> z_q_7 = conv(bias = quantizers_3_out_proj_bias, dilations = z_q_7_dilations_0, groups = z_q_7_groups_0, pad = z_q_7_pad_0, pad_type = z_q_7_pad_type_0, strides = z_q_7_strides_0, weight = quantizers_3_out_proj_weight, x = input_193)[name = tensor<string, []>("z_q_7")]; |
| tensor<fp32, [1, 512, ?]> input_195 = sub(x = input_185, y = z_q_7)[name = tensor<string, []>("input_195")]; |
| tensor<int32, [2]> var_2489 = const()[name = tensor<string, []>("op_2489"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2490 = reshape(shape = var_2489, x = input_191)[name = tensor<string, []>("op_2490")]; |
| tensor<string, []> z_e_9_pad_type_0 = const()[name = tensor<string, []>("z_e_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_9_strides_0 = const()[name = tensor<string, []>("z_e_9_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_9_pad_0 = const()[name = tensor<string, []>("z_e_9_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_9_dilations_0 = const()[name = tensor<string, []>("z_e_9_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_9_groups_0 = const()[name = tensor<string, []>("z_e_9_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_9 = conv(bias = quantizers_4_in_proj_bias, dilations = z_e_9_dilations_0, groups = z_e_9_groups_0, pad = z_e_9_pad_0, pad_type = z_e_9_pad_type_0, strides = z_e_9_strides_0, weight = quantizers_4_in_proj_weight, x = input_195)[name = tensor<string, []>("z_e_9")]; |
| tensor<int32, [3]> var_2504_perm_0 = const()[name = tensor<string, []>("op_2504_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2510 = const()[name = tensor<string, []>("op_2510"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2504 = transpose(perm = var_2504_perm_0, x = z_e_9)[name = tensor<string, []>("transpose_38")]; |
| tensor<fp32, [?, 8]> input_197 = reshape(shape = var_2510, x = var_2504)[name = tensor<string, []>("input_197")]; |
| tensor<int32, [1]> var_2514 = const()[name = tensor<string, []>("op_2514"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2515 = const()[name = tensor<string, []>("op_2515"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2517 = reduce_l2_norm(axes = var_2514, keep_dims = var_2515, x = input_197)[name = tensor<string, []>("op_2517")]; |
| tensor<fp32, []> var_2518 = const()[name = tensor<string, []>("op_2518"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2519 = maximum(x = var_2517, y = var_2518)[name = tensor<string, []>("op_2519")]; |
| tensor<int32, [2]> shape_12 = shape(x = input_197)[name = tensor<string, []>("shape_12")]; |
| tensor<int32, [2]> shape_13 = shape(x = var_2519)[name = tensor<string, []>("shape_13")]; |
| tensor<int32, []> equal_4_y_0 = const()[name = tensor<string, []>("equal_4_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_4 = equal(x = shape_12, y = equal_4_y_0)[name = tensor<string, []>("equal_4")]; |
| tensor<int32, [2]> select_4 = select(a = shape_13, b = shape_12, cond = equal_4)[name = tensor<string, []>("select_4")]; |
| tensor<int32, [2]> real_div_4 = real_div(x = select_4, y = shape_13)[name = tensor<string, []>("real_div_4")]; |
| tensor<fp32, [?, ?]> denom_17 = tile(reps = real_div_4, x = var_2519)[name = tensor<string, []>("denom_17")]; |
| tensor<fp32, [?, 8]> enc_9 = real_div(x = input_197, y = denom_17)[name = tensor<string, []>("enc_9")]; |
| tensor<fp32, [1024, 8]> cb_9 = const()[name = tensor<string, []>("cb_9"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44565568)))]; |
| tensor<fp32, [?, 1024]> sim_9 = linear(bias = linear_3_bias_0, weight = cb_9, x = enc_9)[name = tensor<string, []>("sim_9")]; |
| tensor<int32, []> input_201_axis_0 = const()[name = tensor<string, []>("input_201_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_201_keep_dims_0 = const()[name = tensor<string, []>("input_201_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_201_output_dtype_0 = const()[name = tensor<string, []>("input_201_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_201 = reduce_argmax(axis = input_201_axis_0, keep_dims = input_201_keep_dims_0, output_dtype = input_201_output_dtype_0, x = sim_9)[name = tensor<string, []>("input_201")]; |
| tensor<int32, []> var_2540_batch_dims_0 = const()[name = tensor<string, []>("op_2540_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2540_validate_indices_0 = const()[name = tensor<string, []>("op_2540_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_28_y_0 = const()[name = tensor<string, []>("greater_equal_28_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_28 = greater_equal(x = input_201, y = greater_equal_28_y_0)[name = tensor<string, []>("greater_equal_28")]; |
| tensor<int32, []> slice_by_index_28 = const()[name = tensor<string, []>("slice_by_index_28"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_40 = add(x = input_201, y = slice_by_index_28)[name = tensor<string, []>("add_40")]; |
| tensor<int32, [?]> select_44 = select(a = input_201, b = add_40, cond = greater_equal_28)[name = tensor<string, []>("select_44")]; |
| tensor<int32, []> var_2540_axis_1 = const()[name = tensor<string, []>("op_2540_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2540 = gather(axis = var_2540_axis_1, batch_dims = var_2540_batch_dims_0, indices = select_44, validate_indices = var_2540_validate_indices_0, x = quantizers_4_codebook_weight)[name = tensor<string, []>("op_2540")]; |
| tensor<int32, [3]> var_2550 = const()[name = tensor<string, []>("op_2550"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2551 = reshape(shape = var_2550, x = var_2540)[name = tensor<string, []>("op_2551")]; |
| tensor<int32, [3]> input_203_perm_0 = const()[name = tensor<string, []>("input_203_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_9_pad_type_0 = const()[name = tensor<string, []>("z_q_9_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_9_strides_0 = const()[name = tensor<string, []>("z_q_9_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_9_pad_0 = const()[name = tensor<string, []>("z_q_9_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_9_dilations_0 = const()[name = tensor<string, []>("z_q_9_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_9_groups_0 = const()[name = tensor<string, []>("z_q_9_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_203 = transpose(perm = input_203_perm_0, x = var_2551)[name = tensor<string, []>("transpose_37")]; |
| tensor<fp32, [1, 512, ?]> z_q_9 = conv(bias = quantizers_4_out_proj_bias, dilations = z_q_9_dilations_0, groups = z_q_9_groups_0, pad = z_q_9_pad_0, pad_type = z_q_9_pad_type_0, strides = z_q_9_strides_0, weight = quantizers_4_out_proj_weight, x = input_203)[name = tensor<string, []>("z_q_9")]; |
| tensor<fp32, [1, 512, ?]> input_205 = sub(x = input_195, y = z_q_9)[name = tensor<string, []>("input_205")]; |
| tensor<int32, [2]> var_2573 = const()[name = tensor<string, []>("op_2573"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2574 = reshape(shape = var_2573, x = input_201)[name = tensor<string, []>("op_2574")]; |
| tensor<string, []> z_e_11_pad_type_0 = const()[name = tensor<string, []>("z_e_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_11_strides_0 = const()[name = tensor<string, []>("z_e_11_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_11_pad_0 = const()[name = tensor<string, []>("z_e_11_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_11_dilations_0 = const()[name = tensor<string, []>("z_e_11_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_11_groups_0 = const()[name = tensor<string, []>("z_e_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_11 = conv(bias = quantizers_5_in_proj_bias, dilations = z_e_11_dilations_0, groups = z_e_11_groups_0, pad = z_e_11_pad_0, pad_type = z_e_11_pad_type_0, strides = z_e_11_strides_0, weight = quantizers_5_in_proj_weight, x = input_205)[name = tensor<string, []>("z_e_11")]; |
| tensor<int32, [3]> var_2588_perm_0 = const()[name = tensor<string, []>("op_2588_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2594 = const()[name = tensor<string, []>("op_2594"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2588 = transpose(perm = var_2588_perm_0, x = z_e_11)[name = tensor<string, []>("transpose_36")]; |
| tensor<fp32, [?, 8]> input_207 = reshape(shape = var_2594, x = var_2588)[name = tensor<string, []>("input_207")]; |
| tensor<int32, [1]> var_2598 = const()[name = tensor<string, []>("op_2598"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2599 = const()[name = tensor<string, []>("op_2599"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2601 = reduce_l2_norm(axes = var_2598, keep_dims = var_2599, x = input_207)[name = tensor<string, []>("op_2601")]; |
| tensor<fp32, []> var_2602 = const()[name = tensor<string, []>("op_2602"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2603 = maximum(x = var_2601, y = var_2602)[name = tensor<string, []>("op_2603")]; |
| tensor<int32, [2]> shape_15 = shape(x = input_207)[name = tensor<string, []>("shape_15")]; |
| tensor<int32, [2]> shape_16 = shape(x = var_2603)[name = tensor<string, []>("shape_16")]; |
| tensor<int32, []> equal_5_y_0 = const()[name = tensor<string, []>("equal_5_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_5 = equal(x = shape_15, y = equal_5_y_0)[name = tensor<string, []>("equal_5")]; |
| tensor<int32, [2]> select_5 = select(a = shape_16, b = shape_15, cond = equal_5)[name = tensor<string, []>("select_5")]; |
| tensor<int32, [2]> real_div_5 = real_div(x = select_5, y = shape_16)[name = tensor<string, []>("real_div_5")]; |
| tensor<fp32, [?, ?]> denom_21 = tile(reps = real_div_5, x = var_2603)[name = tensor<string, []>("denom_21")]; |
| tensor<fp32, [?, 8]> enc_11 = real_div(x = input_207, y = denom_21)[name = tensor<string, []>("enc_11")]; |
| tensor<fp32, [1024, 8]> cb_11 = const()[name = tensor<string, []>("cb_11"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44598400)))]; |
| tensor<fp32, [?, 1024]> sim_11 = linear(bias = linear_3_bias_0, weight = cb_11, x = enc_11)[name = tensor<string, []>("sim_11")]; |
| tensor<int32, []> input_211_axis_0 = const()[name = tensor<string, []>("input_211_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_211_keep_dims_0 = const()[name = tensor<string, []>("input_211_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_211_output_dtype_0 = const()[name = tensor<string, []>("input_211_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_211 = reduce_argmax(axis = input_211_axis_0, keep_dims = input_211_keep_dims_0, output_dtype = input_211_output_dtype_0, x = sim_11)[name = tensor<string, []>("input_211")]; |
| tensor<int32, []> var_2624_batch_dims_0 = const()[name = tensor<string, []>("op_2624_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2624_validate_indices_0 = const()[name = tensor<string, []>("op_2624_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_29_y_0 = const()[name = tensor<string, []>("greater_equal_29_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_29 = greater_equal(x = input_211, y = greater_equal_29_y_0)[name = tensor<string, []>("greater_equal_29")]; |
| tensor<int32, []> slice_by_index_29 = const()[name = tensor<string, []>("slice_by_index_29"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_41 = add(x = input_211, y = slice_by_index_29)[name = tensor<string, []>("add_41")]; |
| tensor<int32, [?]> select_45 = select(a = input_211, b = add_41, cond = greater_equal_29)[name = tensor<string, []>("select_45")]; |
| tensor<int32, []> var_2624_axis_1 = const()[name = tensor<string, []>("op_2624_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2624 = gather(axis = var_2624_axis_1, batch_dims = var_2624_batch_dims_0, indices = select_45, validate_indices = var_2624_validate_indices_0, x = quantizers_5_codebook_weight)[name = tensor<string, []>("op_2624")]; |
| tensor<int32, [3]> var_2634 = const()[name = tensor<string, []>("op_2634"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2635 = reshape(shape = var_2634, x = var_2624)[name = tensor<string, []>("op_2635")]; |
| tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_11_pad_type_0 = const()[name = tensor<string, []>("z_q_11_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_11_strides_0 = const()[name = tensor<string, []>("z_q_11_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_11_pad_0 = const()[name = tensor<string, []>("z_q_11_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_11_dilations_0 = const()[name = tensor<string, []>("z_q_11_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_11_groups_0 = const()[name = tensor<string, []>("z_q_11_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_213 = transpose(perm = input_213_perm_0, x = var_2635)[name = tensor<string, []>("transpose_35")]; |
| tensor<fp32, [1, 512, ?]> z_q_11 = conv(bias = quantizers_5_out_proj_bias, dilations = z_q_11_dilations_0, groups = z_q_11_groups_0, pad = z_q_11_pad_0, pad_type = z_q_11_pad_type_0, strides = z_q_11_strides_0, weight = quantizers_5_out_proj_weight, x = input_213)[name = tensor<string, []>("z_q_11")]; |
| tensor<fp32, [1, 512, ?]> input_215 = sub(x = input_205, y = z_q_11)[name = tensor<string, []>("input_215")]; |
| tensor<int32, [2]> var_2657 = const()[name = tensor<string, []>("op_2657"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2658 = reshape(shape = var_2657, x = input_211)[name = tensor<string, []>("op_2658")]; |
| tensor<string, []> z_e_13_pad_type_0 = const()[name = tensor<string, []>("z_e_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_13_strides_0 = const()[name = tensor<string, []>("z_e_13_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_13_pad_0 = const()[name = tensor<string, []>("z_e_13_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_13_dilations_0 = const()[name = tensor<string, []>("z_e_13_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_13_groups_0 = const()[name = tensor<string, []>("z_e_13_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_13 = conv(bias = quantizers_6_in_proj_bias, dilations = z_e_13_dilations_0, groups = z_e_13_groups_0, pad = z_e_13_pad_0, pad_type = z_e_13_pad_type_0, strides = z_e_13_strides_0, weight = quantizers_6_in_proj_weight, x = input_215)[name = tensor<string, []>("z_e_13")]; |
| tensor<int32, [3]> var_2672_perm_0 = const()[name = tensor<string, []>("op_2672_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2678 = const()[name = tensor<string, []>("op_2678"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2672 = transpose(perm = var_2672_perm_0, x = z_e_13)[name = tensor<string, []>("transpose_34")]; |
| tensor<fp32, [?, 8]> input_217 = reshape(shape = var_2678, x = var_2672)[name = tensor<string, []>("input_217")]; |
| tensor<int32, [1]> var_2682 = const()[name = tensor<string, []>("op_2682"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2683 = const()[name = tensor<string, []>("op_2683"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2685 = reduce_l2_norm(axes = var_2682, keep_dims = var_2683, x = input_217)[name = tensor<string, []>("op_2685")]; |
| tensor<fp32, []> var_2686 = const()[name = tensor<string, []>("op_2686"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2687 = maximum(x = var_2685, y = var_2686)[name = tensor<string, []>("op_2687")]; |
| tensor<int32, [2]> shape_18 = shape(x = input_217)[name = tensor<string, []>("shape_18")]; |
| tensor<int32, [2]> shape_19 = shape(x = var_2687)[name = tensor<string, []>("shape_19")]; |
| tensor<int32, []> equal_6_y_0 = const()[name = tensor<string, []>("equal_6_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_6 = equal(x = shape_18, y = equal_6_y_0)[name = tensor<string, []>("equal_6")]; |
| tensor<int32, [2]> select_6 = select(a = shape_19, b = shape_18, cond = equal_6)[name = tensor<string, []>("select_6")]; |
| tensor<int32, [2]> real_div_6 = real_div(x = select_6, y = shape_19)[name = tensor<string, []>("real_div_6")]; |
| tensor<fp32, [?, ?]> denom_25 = tile(reps = real_div_6, x = var_2687)[name = tensor<string, []>("denom_25")]; |
| tensor<fp32, [?, 8]> enc_13 = real_div(x = input_217, y = denom_25)[name = tensor<string, []>("enc_13")]; |
| tensor<fp32, [1024, 8]> cb_13 = const()[name = tensor<string, []>("cb_13"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44631232)))]; |
| tensor<fp32, [?, 1024]> sim_13 = linear(bias = linear_3_bias_0, weight = cb_13, x = enc_13)[name = tensor<string, []>("sim_13")]; |
| tensor<int32, []> input_221_axis_0 = const()[name = tensor<string, []>("input_221_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_221_keep_dims_0 = const()[name = tensor<string, []>("input_221_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_221_output_dtype_0 = const()[name = tensor<string, []>("input_221_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_221 = reduce_argmax(axis = input_221_axis_0, keep_dims = input_221_keep_dims_0, output_dtype = input_221_output_dtype_0, x = sim_13)[name = tensor<string, []>("input_221")]; |
| tensor<int32, []> var_2708_batch_dims_0 = const()[name = tensor<string, []>("op_2708_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2708_validate_indices_0 = const()[name = tensor<string, []>("op_2708_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_30_y_0 = const()[name = tensor<string, []>("greater_equal_30_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_30 = greater_equal(x = input_221, y = greater_equal_30_y_0)[name = tensor<string, []>("greater_equal_30")]; |
| tensor<int32, []> slice_by_index_30 = const()[name = tensor<string, []>("slice_by_index_30"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_42 = add(x = input_221, y = slice_by_index_30)[name = tensor<string, []>("add_42")]; |
| tensor<int32, [?]> select_46 = select(a = input_221, b = add_42, cond = greater_equal_30)[name = tensor<string, []>("select_46")]; |
| tensor<int32, []> var_2708_axis_1 = const()[name = tensor<string, []>("op_2708_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2708 = gather(axis = var_2708_axis_1, batch_dims = var_2708_batch_dims_0, indices = select_46, validate_indices = var_2708_validate_indices_0, x = quantizers_6_codebook_weight)[name = tensor<string, []>("op_2708")]; |
| tensor<int32, [3]> var_2718 = const()[name = tensor<string, []>("op_2718"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2719 = reshape(shape = var_2718, x = var_2708)[name = tensor<string, []>("op_2719")]; |
| tensor<int32, [3]> input_223_perm_0 = const()[name = tensor<string, []>("input_223_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_13_pad_type_0 = const()[name = tensor<string, []>("z_q_13_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_13_strides_0 = const()[name = tensor<string, []>("z_q_13_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_13_pad_0 = const()[name = tensor<string, []>("z_q_13_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_13_dilations_0 = const()[name = tensor<string, []>("z_q_13_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_13_groups_0 = const()[name = tensor<string, []>("z_q_13_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_223 = transpose(perm = input_223_perm_0, x = var_2719)[name = tensor<string, []>("transpose_33")]; |
| tensor<fp32, [1, 512, ?]> z_q_13 = conv(bias = quantizers_6_out_proj_bias, dilations = z_q_13_dilations_0, groups = z_q_13_groups_0, pad = z_q_13_pad_0, pad_type = z_q_13_pad_type_0, strides = z_q_13_strides_0, weight = quantizers_6_out_proj_weight, x = input_223)[name = tensor<string, []>("z_q_13")]; |
| tensor<fp32, [1, 512, ?]> input_225 = sub(x = input_215, y = z_q_13)[name = tensor<string, []>("input_225")]; |
| tensor<int32, [2]> var_2741 = const()[name = tensor<string, []>("op_2741"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2742 = reshape(shape = var_2741, x = input_221)[name = tensor<string, []>("op_2742")]; |
| tensor<string, []> z_e_15_pad_type_0 = const()[name = tensor<string, []>("z_e_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_15_strides_0 = const()[name = tensor<string, []>("z_e_15_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_15_pad_0 = const()[name = tensor<string, []>("z_e_15_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_15_dilations_0 = const()[name = tensor<string, []>("z_e_15_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_15_groups_0 = const()[name = tensor<string, []>("z_e_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_15 = conv(bias = quantizers_7_in_proj_bias, dilations = z_e_15_dilations_0, groups = z_e_15_groups_0, pad = z_e_15_pad_0, pad_type = z_e_15_pad_type_0, strides = z_e_15_strides_0, weight = quantizers_7_in_proj_weight, x = input_225)[name = tensor<string, []>("z_e_15")]; |
| tensor<int32, [3]> var_2756_perm_0 = const()[name = tensor<string, []>("op_2756_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2762 = const()[name = tensor<string, []>("op_2762"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2756 = transpose(perm = var_2756_perm_0, x = z_e_15)[name = tensor<string, []>("transpose_32")]; |
| tensor<fp32, [?, 8]> input_227 = reshape(shape = var_2762, x = var_2756)[name = tensor<string, []>("input_227")]; |
| tensor<int32, [1]> var_2766 = const()[name = tensor<string, []>("op_2766"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2767 = const()[name = tensor<string, []>("op_2767"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2769 = reduce_l2_norm(axes = var_2766, keep_dims = var_2767, x = input_227)[name = tensor<string, []>("op_2769")]; |
| tensor<fp32, []> var_2770 = const()[name = tensor<string, []>("op_2770"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2771 = maximum(x = var_2769, y = var_2770)[name = tensor<string, []>("op_2771")]; |
| tensor<int32, [2]> shape_21 = shape(x = input_227)[name = tensor<string, []>("shape_21")]; |
| tensor<int32, [2]> shape_22 = shape(x = var_2771)[name = tensor<string, []>("shape_22")]; |
| tensor<int32, []> equal_7_y_0 = const()[name = tensor<string, []>("equal_7_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_7 = equal(x = shape_21, y = equal_7_y_0)[name = tensor<string, []>("equal_7")]; |
| tensor<int32, [2]> select_7 = select(a = shape_22, b = shape_21, cond = equal_7)[name = tensor<string, []>("select_7")]; |
| tensor<int32, [2]> real_div_7 = real_div(x = select_7, y = shape_22)[name = tensor<string, []>("real_div_7")]; |
| tensor<fp32, [?, ?]> denom_29 = tile(reps = real_div_7, x = var_2771)[name = tensor<string, []>("denom_29")]; |
| tensor<fp32, [?, 8]> enc_15 = real_div(x = input_227, y = denom_29)[name = tensor<string, []>("enc_15")]; |
| tensor<fp32, [1024, 8]> cb_15 = const()[name = tensor<string, []>("cb_15"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44664064)))]; |
| tensor<fp32, [?, 1024]> sim_15 = linear(bias = linear_3_bias_0, weight = cb_15, x = enc_15)[name = tensor<string, []>("sim_15")]; |
| tensor<int32, []> input_231_axis_0 = const()[name = tensor<string, []>("input_231_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_231_keep_dims_0 = const()[name = tensor<string, []>("input_231_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_231_output_dtype_0 = const()[name = tensor<string, []>("input_231_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_231 = reduce_argmax(axis = input_231_axis_0, keep_dims = input_231_keep_dims_0, output_dtype = input_231_output_dtype_0, x = sim_15)[name = tensor<string, []>("input_231")]; |
| tensor<int32, []> var_2792_batch_dims_0 = const()[name = tensor<string, []>("op_2792_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2792_validate_indices_0 = const()[name = tensor<string, []>("op_2792_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_31_y_0 = const()[name = tensor<string, []>("greater_equal_31_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_31 = greater_equal(x = input_231, y = greater_equal_31_y_0)[name = tensor<string, []>("greater_equal_31")]; |
| tensor<int32, []> slice_by_index_31 = const()[name = tensor<string, []>("slice_by_index_31"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_43 = add(x = input_231, y = slice_by_index_31)[name = tensor<string, []>("add_43")]; |
| tensor<int32, [?]> select_47 = select(a = input_231, b = add_43, cond = greater_equal_31)[name = tensor<string, []>("select_47")]; |
| tensor<int32, []> var_2792_axis_1 = const()[name = tensor<string, []>("op_2792_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2792 = gather(axis = var_2792_axis_1, batch_dims = var_2792_batch_dims_0, indices = select_47, validate_indices = var_2792_validate_indices_0, x = quantizers_7_codebook_weight)[name = tensor<string, []>("op_2792")]; |
| tensor<int32, [3]> var_2802 = const()[name = tensor<string, []>("op_2802"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2803 = reshape(shape = var_2802, x = var_2792)[name = tensor<string, []>("op_2803")]; |
| tensor<int32, [3]> input_233_perm_0 = const()[name = tensor<string, []>("input_233_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_15_pad_type_0 = const()[name = tensor<string, []>("z_q_15_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_15_strides_0 = const()[name = tensor<string, []>("z_q_15_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_15_pad_0 = const()[name = tensor<string, []>("z_q_15_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_15_dilations_0 = const()[name = tensor<string, []>("z_q_15_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_15_groups_0 = const()[name = tensor<string, []>("z_q_15_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_233 = transpose(perm = input_233_perm_0, x = var_2803)[name = tensor<string, []>("transpose_31")]; |
| tensor<fp32, [1, 512, ?]> z_q_15 = conv(bias = quantizers_7_out_proj_bias, dilations = z_q_15_dilations_0, groups = z_q_15_groups_0, pad = z_q_15_pad_0, pad_type = z_q_15_pad_type_0, strides = z_q_15_strides_0, weight = quantizers_7_out_proj_weight, x = input_233)[name = tensor<string, []>("z_q_15")]; |
| tensor<fp32, [1, 512, ?]> input_235 = sub(x = input_225, y = z_q_15)[name = tensor<string, []>("input_235")]; |
| tensor<int32, [2]> var_2825 = const()[name = tensor<string, []>("op_2825"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2826 = reshape(shape = var_2825, x = input_231)[name = tensor<string, []>("op_2826")]; |
| tensor<string, []> z_e_17_pad_type_0 = const()[name = tensor<string, []>("z_e_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_17_strides_0 = const()[name = tensor<string, []>("z_e_17_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_17_pad_0 = const()[name = tensor<string, []>("z_e_17_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_17_dilations_0 = const()[name = tensor<string, []>("z_e_17_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_17_groups_0 = const()[name = tensor<string, []>("z_e_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_17 = conv(bias = quantizers_8_in_proj_bias, dilations = z_e_17_dilations_0, groups = z_e_17_groups_0, pad = z_e_17_pad_0, pad_type = z_e_17_pad_type_0, strides = z_e_17_strides_0, weight = quantizers_8_in_proj_weight, x = input_235)[name = tensor<string, []>("z_e_17")]; |
| tensor<int32, [3]> var_2840_perm_0 = const()[name = tensor<string, []>("op_2840_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2846 = const()[name = tensor<string, []>("op_2846"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2840 = transpose(perm = var_2840_perm_0, x = z_e_17)[name = tensor<string, []>("transpose_30")]; |
| tensor<fp32, [?, 8]> input_237 = reshape(shape = var_2846, x = var_2840)[name = tensor<string, []>("input_237")]; |
| tensor<int32, [1]> var_2850 = const()[name = tensor<string, []>("op_2850"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2851 = const()[name = tensor<string, []>("op_2851"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2853 = reduce_l2_norm(axes = var_2850, keep_dims = var_2851, x = input_237)[name = tensor<string, []>("op_2853")]; |
| tensor<fp32, []> var_2854 = const()[name = tensor<string, []>("op_2854"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2855 = maximum(x = var_2853, y = var_2854)[name = tensor<string, []>("op_2855")]; |
| tensor<int32, [2]> shape_24 = shape(x = input_237)[name = tensor<string, []>("shape_24")]; |
| tensor<int32, [2]> shape_25 = shape(x = var_2855)[name = tensor<string, []>("shape_25")]; |
| tensor<int32, []> equal_8_y_0 = const()[name = tensor<string, []>("equal_8_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_8 = equal(x = shape_24, y = equal_8_y_0)[name = tensor<string, []>("equal_8")]; |
| tensor<int32, [2]> select_8 = select(a = shape_25, b = shape_24, cond = equal_8)[name = tensor<string, []>("select_8")]; |
| tensor<int32, [2]> real_div_8 = real_div(x = select_8, y = shape_25)[name = tensor<string, []>("real_div_8")]; |
| tensor<fp32, [?, ?]> denom_33 = tile(reps = real_div_8, x = var_2855)[name = tensor<string, []>("denom_33")]; |
| tensor<fp32, [?, 8]> enc_17 = real_div(x = input_237, y = denom_33)[name = tensor<string, []>("enc_17")]; |
| tensor<fp32, [1024, 8]> cb_17 = const()[name = tensor<string, []>("cb_17"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44696896)))]; |
| tensor<fp32, [?, 1024]> sim_17 = linear(bias = linear_3_bias_0, weight = cb_17, x = enc_17)[name = tensor<string, []>("sim_17")]; |
| tensor<int32, []> input_241_axis_0 = const()[name = tensor<string, []>("input_241_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_241_keep_dims_0 = const()[name = tensor<string, []>("input_241_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_241_output_dtype_0 = const()[name = tensor<string, []>("input_241_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_241 = reduce_argmax(axis = input_241_axis_0, keep_dims = input_241_keep_dims_0, output_dtype = input_241_output_dtype_0, x = sim_17)[name = tensor<string, []>("input_241")]; |
| tensor<int32, []> var_2876_batch_dims_0 = const()[name = tensor<string, []>("op_2876_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2876_validate_indices_0 = const()[name = tensor<string, []>("op_2876_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_32_y_0 = const()[name = tensor<string, []>("greater_equal_32_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_32 = greater_equal(x = input_241, y = greater_equal_32_y_0)[name = tensor<string, []>("greater_equal_32")]; |
| tensor<int32, []> slice_by_index_32 = const()[name = tensor<string, []>("slice_by_index_32"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_44 = add(x = input_241, y = slice_by_index_32)[name = tensor<string, []>("add_44")]; |
| tensor<int32, [?]> select_48 = select(a = input_241, b = add_44, cond = greater_equal_32)[name = tensor<string, []>("select_48")]; |
| tensor<int32, []> var_2876_axis_1 = const()[name = tensor<string, []>("op_2876_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2876 = gather(axis = var_2876_axis_1, batch_dims = var_2876_batch_dims_0, indices = select_48, validate_indices = var_2876_validate_indices_0, x = quantizers_8_codebook_weight)[name = tensor<string, []>("op_2876")]; |
| tensor<int32, [3]> var_2886 = const()[name = tensor<string, []>("op_2886"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2887 = reshape(shape = var_2886, x = var_2876)[name = tensor<string, []>("op_2887")]; |
| tensor<int32, [3]> input_243_perm_0 = const()[name = tensor<string, []>("input_243_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_17_pad_type_0 = const()[name = tensor<string, []>("z_q_17_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_17_strides_0 = const()[name = tensor<string, []>("z_q_17_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_17_pad_0 = const()[name = tensor<string, []>("z_q_17_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_17_dilations_0 = const()[name = tensor<string, []>("z_q_17_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_17_groups_0 = const()[name = tensor<string, []>("z_q_17_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_243 = transpose(perm = input_243_perm_0, x = var_2887)[name = tensor<string, []>("transpose_29")]; |
| tensor<fp32, [1, 512, ?]> z_q_17 = conv(bias = quantizers_8_out_proj_bias, dilations = z_q_17_dilations_0, groups = z_q_17_groups_0, pad = z_q_17_pad_0, pad_type = z_q_17_pad_type_0, strides = z_q_17_strides_0, weight = quantizers_8_out_proj_weight, x = input_243)[name = tensor<string, []>("z_q_17")]; |
| tensor<fp32, [1, 512, ?]> input_245 = sub(x = input_235, y = z_q_17)[name = tensor<string, []>("input_245")]; |
| tensor<int32, [2]> var_2909 = const()[name = tensor<string, []>("op_2909"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2910 = reshape(shape = var_2909, x = input_241)[name = tensor<string, []>("op_2910")]; |
| tensor<string, []> z_e_19_pad_type_0 = const()[name = tensor<string, []>("z_e_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_19_strides_0 = const()[name = tensor<string, []>("z_e_19_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_19_pad_0 = const()[name = tensor<string, []>("z_e_19_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_19_dilations_0 = const()[name = tensor<string, []>("z_e_19_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_19_groups_0 = const()[name = tensor<string, []>("z_e_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_19 = conv(bias = quantizers_9_in_proj_bias, dilations = z_e_19_dilations_0, groups = z_e_19_groups_0, pad = z_e_19_pad_0, pad_type = z_e_19_pad_type_0, strides = z_e_19_strides_0, weight = quantizers_9_in_proj_weight, x = input_245)[name = tensor<string, []>("z_e_19")]; |
| tensor<int32, [3]> var_2924_perm_0 = const()[name = tensor<string, []>("op_2924_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_2930 = const()[name = tensor<string, []>("op_2930"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2924 = transpose(perm = var_2924_perm_0, x = z_e_19)[name = tensor<string, []>("transpose_28")]; |
| tensor<fp32, [?, 8]> input_247 = reshape(shape = var_2930, x = var_2924)[name = tensor<string, []>("input_247")]; |
| tensor<int32, [1]> var_2934 = const()[name = tensor<string, []>("op_2934"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_2935 = const()[name = tensor<string, []>("op_2935"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_2937 = reduce_l2_norm(axes = var_2934, keep_dims = var_2935, x = input_247)[name = tensor<string, []>("op_2937")]; |
| tensor<fp32, []> var_2938 = const()[name = tensor<string, []>("op_2938"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_2939 = maximum(x = var_2937, y = var_2938)[name = tensor<string, []>("op_2939")]; |
| tensor<int32, [2]> shape_27 = shape(x = input_247)[name = tensor<string, []>("shape_27")]; |
| tensor<int32, [2]> shape_28 = shape(x = var_2939)[name = tensor<string, []>("shape_28")]; |
| tensor<int32, []> equal_9_y_0 = const()[name = tensor<string, []>("equal_9_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_9 = equal(x = shape_27, y = equal_9_y_0)[name = tensor<string, []>("equal_9")]; |
| tensor<int32, [2]> select_9 = select(a = shape_28, b = shape_27, cond = equal_9)[name = tensor<string, []>("select_9")]; |
| tensor<int32, [2]> real_div_9 = real_div(x = select_9, y = shape_28)[name = tensor<string, []>("real_div_9")]; |
| tensor<fp32, [?, ?]> denom_37 = tile(reps = real_div_9, x = var_2939)[name = tensor<string, []>("denom_37")]; |
| tensor<fp32, [?, 8]> enc_19 = real_div(x = input_247, y = denom_37)[name = tensor<string, []>("enc_19")]; |
| tensor<fp32, [1024, 8]> cb_19 = const()[name = tensor<string, []>("cb_19"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44729728)))]; |
| tensor<fp32, [?, 1024]> sim_19 = linear(bias = linear_3_bias_0, weight = cb_19, x = enc_19)[name = tensor<string, []>("sim_19")]; |
| tensor<int32, []> input_251_axis_0 = const()[name = tensor<string, []>("input_251_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_251_keep_dims_0 = const()[name = tensor<string, []>("input_251_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_251_output_dtype_0 = const()[name = tensor<string, []>("input_251_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_251 = reduce_argmax(axis = input_251_axis_0, keep_dims = input_251_keep_dims_0, output_dtype = input_251_output_dtype_0, x = sim_19)[name = tensor<string, []>("input_251")]; |
| tensor<int32, []> var_2960_batch_dims_0 = const()[name = tensor<string, []>("op_2960_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_2960_validate_indices_0 = const()[name = tensor<string, []>("op_2960_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_33_y_0 = const()[name = tensor<string, []>("greater_equal_33_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_33 = greater_equal(x = input_251, y = greater_equal_33_y_0)[name = tensor<string, []>("greater_equal_33")]; |
| tensor<int32, []> slice_by_index_33 = const()[name = tensor<string, []>("slice_by_index_33"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_45 = add(x = input_251, y = slice_by_index_33)[name = tensor<string, []>("add_45")]; |
| tensor<int32, [?]> select_49 = select(a = input_251, b = add_45, cond = greater_equal_33)[name = tensor<string, []>("select_49")]; |
| tensor<int32, []> var_2960_axis_1 = const()[name = tensor<string, []>("op_2960_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_2960 = gather(axis = var_2960_axis_1, batch_dims = var_2960_batch_dims_0, indices = select_49, validate_indices = var_2960_validate_indices_0, x = quantizers_9_codebook_weight)[name = tensor<string, []>("op_2960")]; |
| tensor<int32, [3]> var_2970 = const()[name = tensor<string, []>("op_2970"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_2971 = reshape(shape = var_2970, x = var_2960)[name = tensor<string, []>("op_2971")]; |
| tensor<int32, [3]> input_253_perm_0 = const()[name = tensor<string, []>("input_253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_19_pad_type_0 = const()[name = tensor<string, []>("z_q_19_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_19_strides_0 = const()[name = tensor<string, []>("z_q_19_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_19_pad_0 = const()[name = tensor<string, []>("z_q_19_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_19_dilations_0 = const()[name = tensor<string, []>("z_q_19_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_19_groups_0 = const()[name = tensor<string, []>("z_q_19_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_253 = transpose(perm = input_253_perm_0, x = var_2971)[name = tensor<string, []>("transpose_27")]; |
| tensor<fp32, [1, 512, ?]> z_q_19 = conv(bias = quantizers_9_out_proj_bias, dilations = z_q_19_dilations_0, groups = z_q_19_groups_0, pad = z_q_19_pad_0, pad_type = z_q_19_pad_type_0, strides = z_q_19_strides_0, weight = quantizers_9_out_proj_weight, x = input_253)[name = tensor<string, []>("z_q_19")]; |
| tensor<fp32, [1, 512, ?]> input_255 = sub(x = input_245, y = z_q_19)[name = tensor<string, []>("input_255")]; |
| tensor<int32, [2]> var_2993 = const()[name = tensor<string, []>("op_2993"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_2994 = reshape(shape = var_2993, x = input_251)[name = tensor<string, []>("op_2994")]; |
| tensor<string, []> z_e_21_pad_type_0 = const()[name = tensor<string, []>("z_e_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_21_strides_0 = const()[name = tensor<string, []>("z_e_21_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_21_pad_0 = const()[name = tensor<string, []>("z_e_21_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_21_dilations_0 = const()[name = tensor<string, []>("z_e_21_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_21_groups_0 = const()[name = tensor<string, []>("z_e_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_21 = conv(bias = quantizers_10_in_proj_bias, dilations = z_e_21_dilations_0, groups = z_e_21_groups_0, pad = z_e_21_pad_0, pad_type = z_e_21_pad_type_0, strides = z_e_21_strides_0, weight = quantizers_10_in_proj_weight, x = input_255)[name = tensor<string, []>("z_e_21")]; |
| tensor<int32, [3]> var_3008_perm_0 = const()[name = tensor<string, []>("op_3008_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_3014 = const()[name = tensor<string, []>("op_3014"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3008 = transpose(perm = var_3008_perm_0, x = z_e_21)[name = tensor<string, []>("transpose_26")]; |
| tensor<fp32, [?, 8]> input_257 = reshape(shape = var_3014, x = var_3008)[name = tensor<string, []>("input_257")]; |
| tensor<int32, [1]> var_3018 = const()[name = tensor<string, []>("op_3018"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_3019 = const()[name = tensor<string, []>("op_3019"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_3021 = reduce_l2_norm(axes = var_3018, keep_dims = var_3019, x = input_257)[name = tensor<string, []>("op_3021")]; |
| tensor<fp32, []> var_3022 = const()[name = tensor<string, []>("op_3022"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_3023 = maximum(x = var_3021, y = var_3022)[name = tensor<string, []>("op_3023")]; |
| tensor<int32, [2]> shape_30 = shape(x = input_257)[name = tensor<string, []>("shape_30")]; |
| tensor<int32, [2]> shape_31 = shape(x = var_3023)[name = tensor<string, []>("shape_31")]; |
| tensor<int32, []> equal_10_y_0 = const()[name = tensor<string, []>("equal_10_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_10 = equal(x = shape_30, y = equal_10_y_0)[name = tensor<string, []>("equal_10")]; |
| tensor<int32, [2]> select_10 = select(a = shape_31, b = shape_30, cond = equal_10)[name = tensor<string, []>("select_10")]; |
| tensor<int32, [2]> real_div_10 = real_div(x = select_10, y = shape_31)[name = tensor<string, []>("real_div_10")]; |
| tensor<fp32, [?, ?]> denom_41 = tile(reps = real_div_10, x = var_3023)[name = tensor<string, []>("denom_41")]; |
| tensor<fp32, [?, 8]> enc_21 = real_div(x = input_257, y = denom_41)[name = tensor<string, []>("enc_21")]; |
| tensor<fp32, [1024, 8]> cb_21 = const()[name = tensor<string, []>("cb_21"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44762560)))]; |
| tensor<fp32, [?, 1024]> sim_21 = linear(bias = linear_3_bias_0, weight = cb_21, x = enc_21)[name = tensor<string, []>("sim_21")]; |
| tensor<int32, []> input_261_axis_0 = const()[name = tensor<string, []>("input_261_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_261_keep_dims_0 = const()[name = tensor<string, []>("input_261_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_261_output_dtype_0 = const()[name = tensor<string, []>("input_261_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_261 = reduce_argmax(axis = input_261_axis_0, keep_dims = input_261_keep_dims_0, output_dtype = input_261_output_dtype_0, x = sim_21)[name = tensor<string, []>("input_261")]; |
| tensor<int32, []> var_3044_batch_dims_0 = const()[name = tensor<string, []>("op_3044_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_3044_validate_indices_0 = const()[name = tensor<string, []>("op_3044_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_34_y_0 = const()[name = tensor<string, []>("greater_equal_34_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_34 = greater_equal(x = input_261, y = greater_equal_34_y_0)[name = tensor<string, []>("greater_equal_34")]; |
| tensor<int32, []> slice_by_index_34 = const()[name = tensor<string, []>("slice_by_index_34"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_46 = add(x = input_261, y = slice_by_index_34)[name = tensor<string, []>("add_46")]; |
| tensor<int32, [?]> select_50 = select(a = input_261, b = add_46, cond = greater_equal_34)[name = tensor<string, []>("select_50")]; |
| tensor<int32, []> var_3044_axis_1 = const()[name = tensor<string, []>("op_3044_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_3044 = gather(axis = var_3044_axis_1, batch_dims = var_3044_batch_dims_0, indices = select_50, validate_indices = var_3044_validate_indices_0, x = quantizers_10_codebook_weight)[name = tensor<string, []>("op_3044")]; |
| tensor<int32, [3]> var_3054 = const()[name = tensor<string, []>("op_3054"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3055 = reshape(shape = var_3054, x = var_3044)[name = tensor<string, []>("op_3055")]; |
| tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_21_pad_type_0 = const()[name = tensor<string, []>("z_q_21_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_21_strides_0 = const()[name = tensor<string, []>("z_q_21_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_21_pad_0 = const()[name = tensor<string, []>("z_q_21_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_21_dilations_0 = const()[name = tensor<string, []>("z_q_21_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_21_groups_0 = const()[name = tensor<string, []>("z_q_21_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_263 = transpose(perm = input_263_perm_0, x = var_3055)[name = tensor<string, []>("transpose_25")]; |
| tensor<fp32, [1, 512, ?]> z_q_21 = conv(bias = quantizers_10_out_proj_bias, dilations = z_q_21_dilations_0, groups = z_q_21_groups_0, pad = z_q_21_pad_0, pad_type = z_q_21_pad_type_0, strides = z_q_21_strides_0, weight = quantizers_10_out_proj_weight, x = input_263)[name = tensor<string, []>("z_q_21")]; |
| tensor<fp32, [1, 512, ?]> input_265 = sub(x = input_255, y = z_q_21)[name = tensor<string, []>("input_265")]; |
| tensor<int32, [2]> var_3077 = const()[name = tensor<string, []>("op_3077"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_3078 = reshape(shape = var_3077, x = input_261)[name = tensor<string, []>("op_3078")]; |
| tensor<string, []> z_e_23_pad_type_0 = const()[name = tensor<string, []>("z_e_23_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_23_strides_0 = const()[name = tensor<string, []>("z_e_23_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_23_pad_0 = const()[name = tensor<string, []>("z_e_23_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_23_dilations_0 = const()[name = tensor<string, []>("z_e_23_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_23_groups_0 = const()[name = tensor<string, []>("z_e_23_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_23 = conv(bias = quantizers_11_in_proj_bias, dilations = z_e_23_dilations_0, groups = z_e_23_groups_0, pad = z_e_23_pad_0, pad_type = z_e_23_pad_type_0, strides = z_e_23_strides_0, weight = quantizers_11_in_proj_weight, x = input_265)[name = tensor<string, []>("z_e_23")]; |
| tensor<int32, [3]> var_3092_perm_0 = const()[name = tensor<string, []>("op_3092_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_3098 = const()[name = tensor<string, []>("op_3098"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3092 = transpose(perm = var_3092_perm_0, x = z_e_23)[name = tensor<string, []>("transpose_24")]; |
| tensor<fp32, [?, 8]> input_267 = reshape(shape = var_3098, x = var_3092)[name = tensor<string, []>("input_267")]; |
| tensor<int32, [1]> var_3102 = const()[name = tensor<string, []>("op_3102"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_3103 = const()[name = tensor<string, []>("op_3103"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_3105 = reduce_l2_norm(axes = var_3102, keep_dims = var_3103, x = input_267)[name = tensor<string, []>("op_3105")]; |
| tensor<fp32, []> var_3106 = const()[name = tensor<string, []>("op_3106"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_3107 = maximum(x = var_3105, y = var_3106)[name = tensor<string, []>("op_3107")]; |
| tensor<int32, [2]> shape_33 = shape(x = input_267)[name = tensor<string, []>("shape_33")]; |
| tensor<int32, [2]> shape_34 = shape(x = var_3107)[name = tensor<string, []>("shape_34")]; |
| tensor<int32, []> equal_11_y_0 = const()[name = tensor<string, []>("equal_11_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_11 = equal(x = shape_33, y = equal_11_y_0)[name = tensor<string, []>("equal_11")]; |
| tensor<int32, [2]> select_11 = select(a = shape_34, b = shape_33, cond = equal_11)[name = tensor<string, []>("select_11")]; |
| tensor<int32, [2]> real_div_11 = real_div(x = select_11, y = shape_34)[name = tensor<string, []>("real_div_11")]; |
| tensor<fp32, [?, ?]> denom_45 = tile(reps = real_div_11, x = var_3107)[name = tensor<string, []>("denom_45")]; |
| tensor<fp32, [?, 8]> enc_23 = real_div(x = input_267, y = denom_45)[name = tensor<string, []>("enc_23")]; |
| tensor<fp32, [1024, 8]> cb_23 = const()[name = tensor<string, []>("cb_23"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44795392)))]; |
| tensor<fp32, [?, 1024]> sim_23 = linear(bias = linear_3_bias_0, weight = cb_23, x = enc_23)[name = tensor<string, []>("sim_23")]; |
| tensor<int32, []> input_271_axis_0 = const()[name = tensor<string, []>("input_271_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_271_keep_dims_0 = const()[name = tensor<string, []>("input_271_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_271_output_dtype_0 = const()[name = tensor<string, []>("input_271_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_271 = reduce_argmax(axis = input_271_axis_0, keep_dims = input_271_keep_dims_0, output_dtype = input_271_output_dtype_0, x = sim_23)[name = tensor<string, []>("input_271")]; |
| tensor<int32, []> var_3128_batch_dims_0 = const()[name = tensor<string, []>("op_3128_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_3128_validate_indices_0 = const()[name = tensor<string, []>("op_3128_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_35_y_0 = const()[name = tensor<string, []>("greater_equal_35_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_35 = greater_equal(x = input_271, y = greater_equal_35_y_0)[name = tensor<string, []>("greater_equal_35")]; |
| tensor<int32, []> slice_by_index_35 = const()[name = tensor<string, []>("slice_by_index_35"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_47 = add(x = input_271, y = slice_by_index_35)[name = tensor<string, []>("add_47")]; |
| tensor<int32, [?]> select_51 = select(a = input_271, b = add_47, cond = greater_equal_35)[name = tensor<string, []>("select_51")]; |
| tensor<int32, []> var_3128_axis_1 = const()[name = tensor<string, []>("op_3128_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_3128 = gather(axis = var_3128_axis_1, batch_dims = var_3128_batch_dims_0, indices = select_51, validate_indices = var_3128_validate_indices_0, x = quantizers_11_codebook_weight)[name = tensor<string, []>("op_3128")]; |
| tensor<int32, [3]> var_3138 = const()[name = tensor<string, []>("op_3138"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3139 = reshape(shape = var_3138, x = var_3128)[name = tensor<string, []>("op_3139")]; |
| tensor<int32, [3]> input_273_perm_0 = const()[name = tensor<string, []>("input_273_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_23_pad_type_0 = const()[name = tensor<string, []>("z_q_23_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_23_strides_0 = const()[name = tensor<string, []>("z_q_23_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_23_pad_0 = const()[name = tensor<string, []>("z_q_23_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_23_dilations_0 = const()[name = tensor<string, []>("z_q_23_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_23_groups_0 = const()[name = tensor<string, []>("z_q_23_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_273 = transpose(perm = input_273_perm_0, x = var_3139)[name = tensor<string, []>("transpose_23")]; |
| tensor<fp32, [1, 512, ?]> z_q_23 = conv(bias = quantizers_11_out_proj_bias, dilations = z_q_23_dilations_0, groups = z_q_23_groups_0, pad = z_q_23_pad_0, pad_type = z_q_23_pad_type_0, strides = z_q_23_strides_0, weight = quantizers_11_out_proj_weight, x = input_273)[name = tensor<string, []>("z_q_23")]; |
| tensor<fp32, [1, 512, ?]> input_275 = sub(x = input_265, y = z_q_23)[name = tensor<string, []>("input_275")]; |
| tensor<int32, [2]> var_3161 = const()[name = tensor<string, []>("op_3161"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_3162 = reshape(shape = var_3161, x = input_271)[name = tensor<string, []>("op_3162")]; |
| tensor<string, []> z_e_25_pad_type_0 = const()[name = tensor<string, []>("z_e_25_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_25_strides_0 = const()[name = tensor<string, []>("z_e_25_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_25_pad_0 = const()[name = tensor<string, []>("z_e_25_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_25_dilations_0 = const()[name = tensor<string, []>("z_e_25_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_25_groups_0 = const()[name = tensor<string, []>("z_e_25_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_25 = conv(bias = quantizers_12_in_proj_bias, dilations = z_e_25_dilations_0, groups = z_e_25_groups_0, pad = z_e_25_pad_0, pad_type = z_e_25_pad_type_0, strides = z_e_25_strides_0, weight = quantizers_12_in_proj_weight, x = input_275)[name = tensor<string, []>("z_e_25")]; |
| tensor<int32, [3]> var_3176_perm_0 = const()[name = tensor<string, []>("op_3176_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_3182 = const()[name = tensor<string, []>("op_3182"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3176 = transpose(perm = var_3176_perm_0, x = z_e_25)[name = tensor<string, []>("transpose_22")]; |
| tensor<fp32, [?, 8]> input_277 = reshape(shape = var_3182, x = var_3176)[name = tensor<string, []>("input_277")]; |
| tensor<int32, [1]> var_3186 = const()[name = tensor<string, []>("op_3186"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_3187 = const()[name = tensor<string, []>("op_3187"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_3189 = reduce_l2_norm(axes = var_3186, keep_dims = var_3187, x = input_277)[name = tensor<string, []>("op_3189")]; |
| tensor<fp32, []> var_3190 = const()[name = tensor<string, []>("op_3190"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_3191 = maximum(x = var_3189, y = var_3190)[name = tensor<string, []>("op_3191")]; |
| tensor<int32, [2]> shape_36 = shape(x = input_277)[name = tensor<string, []>("shape_36")]; |
| tensor<int32, [2]> shape_37 = shape(x = var_3191)[name = tensor<string, []>("shape_37")]; |
| tensor<int32, []> equal_12_y_0 = const()[name = tensor<string, []>("equal_12_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_12 = equal(x = shape_36, y = equal_12_y_0)[name = tensor<string, []>("equal_12")]; |
| tensor<int32, [2]> select_12 = select(a = shape_37, b = shape_36, cond = equal_12)[name = tensor<string, []>("select_12")]; |
| tensor<int32, [2]> real_div_12 = real_div(x = select_12, y = shape_37)[name = tensor<string, []>("real_div_12")]; |
| tensor<fp32, [?, ?]> denom_49 = tile(reps = real_div_12, x = var_3191)[name = tensor<string, []>("denom_49")]; |
| tensor<fp32, [?, 8]> enc_25 = real_div(x = input_277, y = denom_49)[name = tensor<string, []>("enc_25")]; |
| tensor<fp32, [1024, 8]> cb_25 = const()[name = tensor<string, []>("cb_25"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44828224)))]; |
| tensor<fp32, [?, 1024]> sim_25 = linear(bias = linear_3_bias_0, weight = cb_25, x = enc_25)[name = tensor<string, []>("sim_25")]; |
| tensor<int32, []> input_281_axis_0 = const()[name = tensor<string, []>("input_281_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_281_keep_dims_0 = const()[name = tensor<string, []>("input_281_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_281_output_dtype_0 = const()[name = tensor<string, []>("input_281_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_281 = reduce_argmax(axis = input_281_axis_0, keep_dims = input_281_keep_dims_0, output_dtype = input_281_output_dtype_0, x = sim_25)[name = tensor<string, []>("input_281")]; |
| tensor<int32, []> var_3212_batch_dims_0 = const()[name = tensor<string, []>("op_3212_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_3212_validate_indices_0 = const()[name = tensor<string, []>("op_3212_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_36_y_0 = const()[name = tensor<string, []>("greater_equal_36_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_36 = greater_equal(x = input_281, y = greater_equal_36_y_0)[name = tensor<string, []>("greater_equal_36")]; |
| tensor<int32, []> slice_by_index_36 = const()[name = tensor<string, []>("slice_by_index_36"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_48 = add(x = input_281, y = slice_by_index_36)[name = tensor<string, []>("add_48")]; |
| tensor<int32, [?]> select_52 = select(a = input_281, b = add_48, cond = greater_equal_36)[name = tensor<string, []>("select_52")]; |
| tensor<int32, []> var_3212_axis_1 = const()[name = tensor<string, []>("op_3212_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_3212 = gather(axis = var_3212_axis_1, batch_dims = var_3212_batch_dims_0, indices = select_52, validate_indices = var_3212_validate_indices_0, x = quantizers_12_codebook_weight)[name = tensor<string, []>("op_3212")]; |
| tensor<int32, [3]> var_3222 = const()[name = tensor<string, []>("op_3222"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3223 = reshape(shape = var_3222, x = var_3212)[name = tensor<string, []>("op_3223")]; |
| tensor<int32, [3]> input_283_perm_0 = const()[name = tensor<string, []>("input_283_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_25_pad_type_0 = const()[name = tensor<string, []>("z_q_25_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_25_strides_0 = const()[name = tensor<string, []>("z_q_25_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_25_pad_0 = const()[name = tensor<string, []>("z_q_25_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_25_dilations_0 = const()[name = tensor<string, []>("z_q_25_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_25_groups_0 = const()[name = tensor<string, []>("z_q_25_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_283 = transpose(perm = input_283_perm_0, x = var_3223)[name = tensor<string, []>("transpose_21")]; |
| tensor<fp32, [1, 512, ?]> z_q_25 = conv(bias = quantizers_12_out_proj_bias, dilations = z_q_25_dilations_0, groups = z_q_25_groups_0, pad = z_q_25_pad_0, pad_type = z_q_25_pad_type_0, strides = z_q_25_strides_0, weight = quantizers_12_out_proj_weight, x = input_283)[name = tensor<string, []>("z_q_25")]; |
| tensor<fp32, [1, 512, ?]> input_285 = sub(x = input_275, y = z_q_25)[name = tensor<string, []>("input_285")]; |
| tensor<int32, [2]> var_3245 = const()[name = tensor<string, []>("op_3245"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_3246 = reshape(shape = var_3245, x = input_281)[name = tensor<string, []>("op_3246")]; |
| tensor<string, []> z_e_27_pad_type_0 = const()[name = tensor<string, []>("z_e_27_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_27_strides_0 = const()[name = tensor<string, []>("z_e_27_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_27_pad_0 = const()[name = tensor<string, []>("z_e_27_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_27_dilations_0 = const()[name = tensor<string, []>("z_e_27_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_27_groups_0 = const()[name = tensor<string, []>("z_e_27_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_27 = conv(bias = quantizers_13_in_proj_bias, dilations = z_e_27_dilations_0, groups = z_e_27_groups_0, pad = z_e_27_pad_0, pad_type = z_e_27_pad_type_0, strides = z_e_27_strides_0, weight = quantizers_13_in_proj_weight, x = input_285)[name = tensor<string, []>("z_e_27")]; |
| tensor<int32, [3]> var_3260_perm_0 = const()[name = tensor<string, []>("op_3260_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_3266 = const()[name = tensor<string, []>("op_3266"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3260 = transpose(perm = var_3260_perm_0, x = z_e_27)[name = tensor<string, []>("transpose_20")]; |
| tensor<fp32, [?, 8]> input_287 = reshape(shape = var_3266, x = var_3260)[name = tensor<string, []>("input_287")]; |
| tensor<int32, [1]> var_3270 = const()[name = tensor<string, []>("op_3270"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_3271 = const()[name = tensor<string, []>("op_3271"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_3273 = reduce_l2_norm(axes = var_3270, keep_dims = var_3271, x = input_287)[name = tensor<string, []>("op_3273")]; |
| tensor<fp32, []> var_3274 = const()[name = tensor<string, []>("op_3274"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_3275 = maximum(x = var_3273, y = var_3274)[name = tensor<string, []>("op_3275")]; |
| tensor<int32, [2]> shape_39 = shape(x = input_287)[name = tensor<string, []>("shape_39")]; |
| tensor<int32, [2]> shape_40 = shape(x = var_3275)[name = tensor<string, []>("shape_40")]; |
| tensor<int32, []> equal_13_y_0 = const()[name = tensor<string, []>("equal_13_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_13 = equal(x = shape_39, y = equal_13_y_0)[name = tensor<string, []>("equal_13")]; |
| tensor<int32, [2]> select_13 = select(a = shape_40, b = shape_39, cond = equal_13)[name = tensor<string, []>("select_13")]; |
| tensor<int32, [2]> real_div_13 = real_div(x = select_13, y = shape_40)[name = tensor<string, []>("real_div_13")]; |
| tensor<fp32, [?, ?]> denom_53 = tile(reps = real_div_13, x = var_3275)[name = tensor<string, []>("denom_53")]; |
| tensor<fp32, [?, 8]> enc_27 = real_div(x = input_287, y = denom_53)[name = tensor<string, []>("enc_27")]; |
| tensor<fp32, [1024, 8]> cb_27 = const()[name = tensor<string, []>("cb_27"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44861056)))]; |
| tensor<fp32, [?, 1024]> sim_27 = linear(bias = linear_3_bias_0, weight = cb_27, x = enc_27)[name = tensor<string, []>("sim_27")]; |
| tensor<int32, []> input_291_axis_0 = const()[name = tensor<string, []>("input_291_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_291_keep_dims_0 = const()[name = tensor<string, []>("input_291_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_291_output_dtype_0 = const()[name = tensor<string, []>("input_291_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_291 = reduce_argmax(axis = input_291_axis_0, keep_dims = input_291_keep_dims_0, output_dtype = input_291_output_dtype_0, x = sim_27)[name = tensor<string, []>("input_291")]; |
| tensor<int32, []> var_3296_batch_dims_0 = const()[name = tensor<string, []>("op_3296_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_3296_validate_indices_0 = const()[name = tensor<string, []>("op_3296_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_37_y_0 = const()[name = tensor<string, []>("greater_equal_37_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_37 = greater_equal(x = input_291, y = greater_equal_37_y_0)[name = tensor<string, []>("greater_equal_37")]; |
| tensor<int32, []> slice_by_index_37 = const()[name = tensor<string, []>("slice_by_index_37"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_49 = add(x = input_291, y = slice_by_index_37)[name = tensor<string, []>("add_49")]; |
| tensor<int32, [?]> select_53 = select(a = input_291, b = add_49, cond = greater_equal_37)[name = tensor<string, []>("select_53")]; |
| tensor<int32, []> var_3296_axis_1 = const()[name = tensor<string, []>("op_3296_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_3296 = gather(axis = var_3296_axis_1, batch_dims = var_3296_batch_dims_0, indices = select_53, validate_indices = var_3296_validate_indices_0, x = quantizers_13_codebook_weight)[name = tensor<string, []>("op_3296")]; |
| tensor<int32, [3]> var_3306 = const()[name = tensor<string, []>("op_3306"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3307 = reshape(shape = var_3306, x = var_3296)[name = tensor<string, []>("op_3307")]; |
| tensor<int32, [3]> input_293_perm_0 = const()[name = tensor<string, []>("input_293_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_27_pad_type_0 = const()[name = tensor<string, []>("z_q_27_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_27_strides_0 = const()[name = tensor<string, []>("z_q_27_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_27_pad_0 = const()[name = tensor<string, []>("z_q_27_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_27_dilations_0 = const()[name = tensor<string, []>("z_q_27_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_27_groups_0 = const()[name = tensor<string, []>("z_q_27_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_293 = transpose(perm = input_293_perm_0, x = var_3307)[name = tensor<string, []>("transpose_19")]; |
| tensor<fp32, [1, 512, ?]> z_q_27 = conv(bias = quantizers_13_out_proj_bias, dilations = z_q_27_dilations_0, groups = z_q_27_groups_0, pad = z_q_27_pad_0, pad_type = z_q_27_pad_type_0, strides = z_q_27_strides_0, weight = quantizers_13_out_proj_weight, x = input_293)[name = tensor<string, []>("z_q_27")]; |
| tensor<fp32, [1, 512, ?]> input_295 = sub(x = input_285, y = z_q_27)[name = tensor<string, []>("input_295")]; |
| tensor<int32, [2]> var_3329 = const()[name = tensor<string, []>("op_3329"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_3330 = reshape(shape = var_3329, x = input_291)[name = tensor<string, []>("op_3330")]; |
| tensor<string, []> z_e_29_pad_type_0 = const()[name = tensor<string, []>("z_e_29_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_29_strides_0 = const()[name = tensor<string, []>("z_e_29_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_29_pad_0 = const()[name = tensor<string, []>("z_e_29_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_29_dilations_0 = const()[name = tensor<string, []>("z_e_29_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_29_groups_0 = const()[name = tensor<string, []>("z_e_29_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e_29 = conv(bias = quantizers_14_in_proj_bias, dilations = z_e_29_dilations_0, groups = z_e_29_groups_0, pad = z_e_29_pad_0, pad_type = z_e_29_pad_type_0, strides = z_e_29_strides_0, weight = quantizers_14_in_proj_weight, x = input_295)[name = tensor<string, []>("z_e_29")]; |
| tensor<int32, [3]> var_3344_perm_0 = const()[name = tensor<string, []>("op_3344_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_3350 = const()[name = tensor<string, []>("op_3350"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3344 = transpose(perm = var_3344_perm_0, x = z_e_29)[name = tensor<string, []>("transpose_18")]; |
| tensor<fp32, [?, 8]> input_297 = reshape(shape = var_3350, x = var_3344)[name = tensor<string, []>("input_297")]; |
| tensor<int32, [1]> var_3354 = const()[name = tensor<string, []>("op_3354"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_3355 = const()[name = tensor<string, []>("op_3355"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_3357 = reduce_l2_norm(axes = var_3354, keep_dims = var_3355, x = input_297)[name = tensor<string, []>("op_3357")]; |
| tensor<fp32, []> var_3358 = const()[name = tensor<string, []>("op_3358"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_3359 = maximum(x = var_3357, y = var_3358)[name = tensor<string, []>("op_3359")]; |
| tensor<int32, [2]> shape_42 = shape(x = input_297)[name = tensor<string, []>("shape_42")]; |
| tensor<int32, [2]> shape_43 = shape(x = var_3359)[name = tensor<string, []>("shape_43")]; |
| tensor<int32, []> equal_14_y_0 = const()[name = tensor<string, []>("equal_14_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_14 = equal(x = shape_42, y = equal_14_y_0)[name = tensor<string, []>("equal_14")]; |
| tensor<int32, [2]> select_14 = select(a = shape_43, b = shape_42, cond = equal_14)[name = tensor<string, []>("select_14")]; |
| tensor<int32, [2]> real_div_14 = real_div(x = select_14, y = shape_43)[name = tensor<string, []>("real_div_14")]; |
| tensor<fp32, [?, ?]> denom_57 = tile(reps = real_div_14, x = var_3359)[name = tensor<string, []>("denom_57")]; |
| tensor<fp32, [?, 8]> enc_29 = real_div(x = input_297, y = denom_57)[name = tensor<string, []>("enc_29")]; |
| tensor<fp32, [1024, 8]> cb_29 = const()[name = tensor<string, []>("cb_29"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44893888)))]; |
| tensor<fp32, [?, 1024]> sim_29 = linear(bias = linear_3_bias_0, weight = cb_29, x = enc_29)[name = tensor<string, []>("sim_29")]; |
| tensor<int32, []> input_301_axis_0 = const()[name = tensor<string, []>("input_301_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_301_keep_dims_0 = const()[name = tensor<string, []>("input_301_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_301_output_dtype_0 = const()[name = tensor<string, []>("input_301_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_301 = reduce_argmax(axis = input_301_axis_0, keep_dims = input_301_keep_dims_0, output_dtype = input_301_output_dtype_0, x = sim_29)[name = tensor<string, []>("input_301")]; |
| tensor<int32, []> var_3380_batch_dims_0 = const()[name = tensor<string, []>("op_3380_batch_dims_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, []> var_3380_validate_indices_0 = const()[name = tensor<string, []>("op_3380_validate_indices_0"), val = tensor<bool, []>(false)]; |
| tensor<int32, []> greater_equal_38_y_0 = const()[name = tensor<string, []>("greater_equal_38_y_0"), val = tensor<int32, []>(0)]; |
| tensor<bool, [?]> greater_equal_38 = greater_equal(x = input_301, y = greater_equal_38_y_0)[name = tensor<string, []>("greater_equal_38")]; |
| tensor<int32, []> slice_by_index_38 = const()[name = tensor<string, []>("slice_by_index_38"), val = tensor<int32, []>(1024)]; |
| tensor<int32, [?]> add_50 = add(x = input_301, y = slice_by_index_38)[name = tensor<string, []>("add_50")]; |
| tensor<int32, [?]> select_54 = select(a = input_301, b = add_50, cond = greater_equal_38)[name = tensor<string, []>("select_54")]; |
| tensor<int32, []> var_3380_axis_1 = const()[name = tensor<string, []>("op_3380_axis_1"), val = tensor<int32, []>(0)]; |
| tensor<fp32, [?, 8]> var_3380 = gather(axis = var_3380_axis_1, batch_dims = var_3380_batch_dims_0, indices = select_54, validate_indices = var_3380_validate_indices_0, x = quantizers_14_codebook_weight)[name = tensor<string, []>("op_3380")]; |
| tensor<int32, [3]> var_3390 = const()[name = tensor<string, []>("op_3390"), val = tensor<int32, [3]>([1, -1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3391 = reshape(shape = var_3390, x = var_3380)[name = tensor<string, []>("op_3391")]; |
| tensor<int32, [3]> input_303_perm_0 = const()[name = tensor<string, []>("input_303_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<string, []> z_q_29_pad_type_0 = const()[name = tensor<string, []>("z_q_29_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_q_29_strides_0 = const()[name = tensor<string, []>("z_q_29_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_q_29_pad_0 = const()[name = tensor<string, []>("z_q_29_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_q_29_dilations_0 = const()[name = tensor<string, []>("z_q_29_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_q_29_groups_0 = const()[name = tensor<string, []>("z_q_29_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> input_303 = transpose(perm = input_303_perm_0, x = var_3391)[name = tensor<string, []>("transpose_17")]; |
| tensor<fp32, [1, 512, ?]> z_q_29 = conv(bias = quantizers_14_out_proj_bias, dilations = z_q_29_dilations_0, groups = z_q_29_groups_0, pad = z_q_29_pad_0, pad_type = z_q_29_pad_type_0, strides = z_q_29_strides_0, weight = quantizers_14_out_proj_weight, x = input_303)[name = tensor<string, []>("z_q_29")]; |
| tensor<fp32, [1, 512, ?]> input_305 = sub(x = input_295, y = z_q_29)[name = tensor<string, []>("input_305")]; |
| tensor<int32, [2]> var_3413 = const()[name = tensor<string, []>("op_3413"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_3414 = reshape(shape = var_3413, x = input_301)[name = tensor<string, []>("op_3414")]; |
| tensor<string, []> z_e_pad_type_0 = const()[name = tensor<string, []>("z_e_pad_type_0"), val = tensor<string, []>("valid")]; |
| tensor<int32, [1]> z_e_strides_0 = const()[name = tensor<string, []>("z_e_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [2]> z_e_pad_0 = const()[name = tensor<string, []>("z_e_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> z_e_dilations_0 = const()[name = tensor<string, []>("z_e_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, []> z_e_groups_0 = const()[name = tensor<string, []>("z_e_groups_0"), val = tensor<int32, []>(1)]; |
| tensor<fp32, [1, 8, ?]> z_e = conv(bias = quantizers_15_in_proj_bias, dilations = z_e_dilations_0, groups = z_e_groups_0, pad = z_e_pad_0, pad_type = z_e_pad_type_0, strides = z_e_strides_0, weight = quantizers_15_in_proj_weight, x = input_305)[name = tensor<string, []>("z_e")]; |
| tensor<int32, [3]> var_3428_perm_0 = const()[name = tensor<string, []>("op_3428_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<int32, [2]> var_3434 = const()[name = tensor<string, []>("op_3434"), val = tensor<int32, [2]>([-1, 8])]; |
| tensor<fp32, [1, ?, 8]> var_3428 = transpose(perm = var_3428_perm_0, x = z_e)[name = tensor<string, []>("transpose_16")]; |
| tensor<fp32, [?, 8]> input_307 = reshape(shape = var_3434, x = var_3428)[name = tensor<string, []>("input_307")]; |
| tensor<int32, [1]> var_3438 = const()[name = tensor<string, []>("op_3438"), val = tensor<int32, [1]>([-1])]; |
| tensor<bool, []> var_3439 = const()[name = tensor<string, []>("op_3439"), val = tensor<bool, []>(true)]; |
| tensor<fp32, [?, 1]> var_3441 = reduce_l2_norm(axes = var_3438, keep_dims = var_3439, x = input_307)[name = tensor<string, []>("op_3441")]; |
| tensor<fp32, []> var_3442 = const()[name = tensor<string, []>("op_3442"), val = tensor<fp32, []>(0x1.197998p-40)]; |
| tensor<fp32, [?, 1]> var_3443 = maximum(x = var_3441, y = var_3442)[name = tensor<string, []>("op_3443")]; |
| tensor<int32, [2]> shape_45 = shape(x = input_307)[name = tensor<string, []>("shape_45")]; |
| tensor<int32, [2]> shape_46 = shape(x = var_3443)[name = tensor<string, []>("shape_46")]; |
| tensor<int32, []> equal_15_y_0 = const()[name = tensor<string, []>("equal_15_y_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, [2]> equal_15 = equal(x = shape_45, y = equal_15_y_0)[name = tensor<string, []>("equal_15")]; |
| tensor<int32, [2]> select_15 = select(a = shape_46, b = shape_45, cond = equal_15)[name = tensor<string, []>("select_15")]; |
| tensor<int32, [2]> real_div_15 = real_div(x = select_15, y = shape_46)[name = tensor<string, []>("real_div_15")]; |
| tensor<fp32, [?, ?]> denom_61 = tile(reps = real_div_15, x = var_3443)[name = tensor<string, []>("denom_61")]; |
| tensor<fp32, [?, 8]> enc = real_div(x = input_307, y = denom_61)[name = tensor<string, []>("enc")]; |
| tensor<fp32, [1024, 8]> cb = const()[name = tensor<string, []>("cb"), val = tensor<fp32, [1024, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44926720)))]; |
| tensor<fp32, [?, 1024]> sim = linear(bias = linear_3_bias_0, weight = cb, x = enc)[name = tensor<string, []>("sim")]; |
| tensor<int32, []> input_311_axis_0 = const()[name = tensor<string, []>("input_311_axis_0"), val = tensor<int32, []>(-1)]; |
| tensor<bool, []> input_311_keep_dims_0 = const()[name = tensor<string, []>("input_311_keep_dims_0"), val = tensor<bool, []>(false)]; |
| tensor<string, []> input_311_output_dtype_0 = const()[name = tensor<string, []>("input_311_output_dtype_0"), val = tensor<string, []>("int32")]; |
| tensor<int32, [?]> input_311 = reduce_argmax(axis = input_311_axis_0, keep_dims = input_311_keep_dims_0, output_dtype = input_311_output_dtype_0, x = sim)[name = tensor<string, []>("input_311")]; |
| tensor<int32, [2]> var_3466 = const()[name = tensor<string, []>("op_3466"), val = tensor<int32, [2]>([1, -1])]; |
| tensor<int32, [1, ?]> var_3467 = reshape(shape = var_3466, x = input_311)[name = tensor<string, []>("op_3467")]; |
| tensor<int32, []> var_3470_axis_0 = const()[name = tensor<string, []>("op_3470_axis_0"), val = tensor<int32, []>(0)]; |
| tensor<int32, [16, 1, ?]> codes = stack(axis = var_3470_axis_0, values = (var_2238, var_2322, var_2406, var_2490, var_2574, var_2658, var_2742, var_2826, var_2910, var_2994, var_3078, var_3162, var_3246, var_3330, var_3414, var_3467))[name = tensor<string, []>("op_3470")]; |
| } -> (codes); |
| } |
