program(1.0) [buildInfo = dict, tensor>({{"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(tensor audio) [FlexibleShapeInformation = tuple, dict, tensor>>, tuple, dict, list, ?>>>>((("DefaultShapes", {{"audio", [1, 2, 380160]}}), ("RangeDims", {{"audio", [[1, 1], [2, 2], [3840, 721920]]}})))] { tensor quantizers_14_codebook_weight = const()[name = tensor("quantizers_14_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; tensor quantizers_13_codebook_weight = const()[name = tensor("quantizers_13_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; tensor quantizers_12_codebook_weight = const()[name = tensor("quantizers_12_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65728)))]; tensor quantizers_11_codebook_weight = const()[name = tensor("quantizers_11_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(98560)))]; tensor quantizers_10_codebook_weight = const()[name = tensor("quantizers_10_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131392)))]; tensor quantizers_9_codebook_weight = const()[name = tensor("quantizers_9_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164224)))]; tensor quantizers_8_codebook_weight = const()[name = tensor("quantizers_8_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197056)))]; tensor quantizers_7_codebook_weight = const()[name = tensor("quantizers_7_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229888)))]; tensor quantizers_6_codebook_weight = const()[name = tensor("quantizers_6_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(262720)))]; tensor quantizers_5_codebook_weight = const()[name = tensor("quantizers_5_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(295552)))]; tensor quantizers_4_codebook_weight = const()[name = tensor("quantizers_4_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328384)))]; tensor quantizers_3_codebook_weight = const()[name = tensor("quantizers_3_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361216)))]; tensor quantizers_2_codebook_weight = const()[name = tensor("quantizers_2_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(394048)))]; tensor quantizers_1_codebook_weight = const()[name = tensor("quantizers_1_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(426880)))]; tensor quantizers_0_codebook_weight = const()[name = tensor("quantizers_0_codebook_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(459712)))]; tensor encoder_1_input_proj_weight = const()[name = tensor("encoder_1_input_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(492544)))]; tensor encoder_1_transformer_layers_0_norm1_bias = const()[name = tensor("encoder_1_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(738368)))]; tensor encoder_1_transformer_layers_0_norm1_weight = const()[name = tensor("encoder_1_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(739456)))]; tensor encoder_1_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("encoder_1_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(740544)))]; tensor encoder_1_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("encoder_1_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1527040)))]; tensor encoder_1_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("encoder_1_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1789248)))]; tensor encoder_1_transformer_layers_0_norm2_bias = const()[name = tensor("encoder_1_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1790336)))]; tensor encoder_1_transformer_layers_0_norm2_weight = const()[name = tensor("encoder_1_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1791424)))]; tensor encoder_1_transformer_layers_0_ffn_0_weight = const()[name = tensor("encoder_1_transformer_layers_0_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1792512)))]; tensor encoder_1_transformer_layers_0_ffn_2_weight = const()[name = tensor("encoder_1_transformer_layers_0_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2841152)))]; tensor encoder_1_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("encoder_1_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3889792)))]; tensor encoder_1_transformer_layers_1_norm1_bias = const()[name = tensor("encoder_1_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3890880)))]; tensor encoder_1_transformer_layers_1_norm1_weight = const()[name = tensor("encoder_1_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3891968)))]; tensor encoder_1_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("encoder_1_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3893056)))]; tensor encoder_1_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("encoder_1_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4679552)))]; tensor encoder_1_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("encoder_1_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4941760)))]; tensor encoder_1_transformer_layers_1_norm2_bias = const()[name = tensor("encoder_1_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4942848)))]; tensor encoder_1_transformer_layers_1_norm2_weight = const()[name = tensor("encoder_1_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4943936)))]; tensor encoder_1_transformer_layers_1_ffn_0_weight = const()[name = tensor("encoder_1_transformer_layers_1_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4945024)))]; tensor encoder_1_transformer_layers_1_ffn_2_weight = const()[name = tensor("encoder_1_transformer_layers_1_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5993664)))]; tensor encoder_1_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("encoder_1_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7042304)))]; tensor encoder_1_transformer_layers_2_norm1_bias = const()[name = tensor("encoder_1_transformer_layers_2_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7043392)))]; tensor encoder_1_transformer_layers_2_norm1_weight = const()[name = tensor("encoder_1_transformer_layers_2_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7044480)))]; tensor encoder_1_transformer_layers_2_self_attn_in_proj_weight = const()[name = tensor("encoder_1_transformer_layers_2_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7045568)))]; tensor encoder_1_transformer_layers_2_self_attn_out_proj_weight = const()[name = tensor("encoder_1_transformer_layers_2_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7832064)))]; tensor encoder_1_transformer_layers_2_layer_scale_1_scale = const()[name = tensor("encoder_1_transformer_layers_2_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8094272)))]; tensor encoder_1_transformer_layers_2_norm2_bias = const()[name = tensor("encoder_1_transformer_layers_2_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8095360)))]; tensor encoder_1_transformer_layers_2_norm2_weight = const()[name = tensor("encoder_1_transformer_layers_2_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8096448)))]; tensor encoder_1_transformer_layers_2_ffn_0_weight = const()[name = tensor("encoder_1_transformer_layers_2_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8097536)))]; tensor encoder_1_transformer_layers_2_ffn_2_weight = const()[name = tensor("encoder_1_transformer_layers_2_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9146176)))]; tensor encoder_1_transformer_layers_2_layer_scale_2_scale = const()[name = tensor("encoder_1_transformer_layers_2_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10194816)))]; tensor encoder_1_transformer_layers_3_norm1_bias = const()[name = tensor("encoder_1_transformer_layers_3_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10195904)))]; tensor encoder_1_transformer_layers_3_norm1_weight = const()[name = tensor("encoder_1_transformer_layers_3_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10196992)))]; tensor encoder_1_transformer_layers_3_self_attn_in_proj_weight = const()[name = tensor("encoder_1_transformer_layers_3_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10198080)))]; tensor encoder_1_transformer_layers_3_self_attn_out_proj_weight = const()[name = tensor("encoder_1_transformer_layers_3_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10984576)))]; tensor encoder_1_transformer_layers_3_layer_scale_1_scale = const()[name = tensor("encoder_1_transformer_layers_3_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11246784)))]; tensor encoder_1_transformer_layers_3_norm2_bias = const()[name = tensor("encoder_1_transformer_layers_3_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11247872)))]; tensor encoder_1_transformer_layers_3_norm2_weight = const()[name = tensor("encoder_1_transformer_layers_3_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11248960)))]; tensor encoder_1_transformer_layers_3_ffn_0_weight = const()[name = tensor("encoder_1_transformer_layers_3_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11250048)))]; tensor encoder_1_transformer_layers_3_ffn_2_weight = const()[name = tensor("encoder_1_transformer_layers_3_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12298688)))]; tensor encoder_1_transformer_layers_3_layer_scale_2_scale = const()[name = tensor("encoder_1_transformer_layers_3_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13347328)))]; tensor encoder_1_output_proj_weight = const()[name = tensor("encoder_1_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13348416)))]; tensor encoder_3_input_proj_weight = const()[name = tensor("encoder_3_input_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13741696)))]; tensor encoder_3_transformer_layers_0_norm1_bias = const()[name = tensor("encoder_3_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14528192)))]; tensor encoder_3_transformer_layers_0_norm1_weight = const()[name = tensor("encoder_3_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14529280)))]; tensor encoder_3_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("encoder_3_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14530368)))]; tensor encoder_3_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("encoder_3_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15316864)))]; tensor encoder_3_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("encoder_3_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15579072)))]; tensor encoder_3_transformer_layers_0_norm2_bias = const()[name = tensor("encoder_3_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15580160)))]; tensor encoder_3_transformer_layers_0_norm2_weight = const()[name = tensor("encoder_3_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15581248)))]; tensor encoder_3_transformer_layers_0_ffn_0_weight = const()[name = tensor("encoder_3_transformer_layers_0_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15582336)))]; tensor encoder_3_transformer_layers_0_ffn_2_weight = const()[name = tensor("encoder_3_transformer_layers_0_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16630976)))]; tensor encoder_3_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("encoder_3_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17679616)))]; tensor encoder_3_transformer_layers_1_norm1_bias = const()[name = tensor("encoder_3_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17680704)))]; tensor encoder_3_transformer_layers_1_norm1_weight = const()[name = tensor("encoder_3_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17681792)))]; tensor encoder_3_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("encoder_3_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17682880)))]; tensor encoder_3_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("encoder_3_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18469376)))]; tensor encoder_3_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("encoder_3_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18731584)))]; tensor encoder_3_transformer_layers_1_norm2_bias = const()[name = tensor("encoder_3_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18732672)))]; tensor encoder_3_transformer_layers_1_norm2_weight = const()[name = tensor("encoder_3_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18733760)))]; tensor encoder_3_transformer_layers_1_ffn_0_weight = const()[name = tensor("encoder_3_transformer_layers_1_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18734848)))]; tensor encoder_3_transformer_layers_1_ffn_2_weight = const()[name = tensor("encoder_3_transformer_layers_1_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19783488)))]; tensor encoder_3_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("encoder_3_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20832128)))]; tensor encoder_3_output_proj_weight = const()[name = tensor("encoder_3_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20833216)))]; tensor encoder_5_input_proj_weight = const()[name = tensor("encoder_5_input_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21226496)))]; tensor encoder_5_transformer_layers_0_norm1_bias = const()[name = tensor("encoder_5_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22012992)))]; tensor encoder_5_transformer_layers_0_norm1_weight = const()[name = tensor("encoder_5_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22014080)))]; tensor encoder_5_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("encoder_5_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22015168)))]; tensor encoder_5_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("encoder_5_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22801664)))]; tensor encoder_5_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("encoder_5_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23063872)))]; tensor encoder_5_transformer_layers_0_norm2_bias = const()[name = tensor("encoder_5_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23064960)))]; tensor encoder_5_transformer_layers_0_norm2_weight = const()[name = tensor("encoder_5_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23066048)))]; tensor encoder_5_transformer_layers_0_ffn_0_weight = const()[name = tensor("encoder_5_transformer_layers_0_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23067136)))]; tensor encoder_5_transformer_layers_0_ffn_2_weight = const()[name = tensor("encoder_5_transformer_layers_0_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24115776)))]; tensor encoder_5_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("encoder_5_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25164416)))]; tensor encoder_5_transformer_layers_1_norm1_bias = const()[name = tensor("encoder_5_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25165504)))]; tensor encoder_5_transformer_layers_1_norm1_weight = const()[name = tensor("encoder_5_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25166592)))]; tensor encoder_5_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("encoder_5_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25167680)))]; tensor encoder_5_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("encoder_5_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25954176)))]; tensor encoder_5_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("encoder_5_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26216384)))]; tensor encoder_5_transformer_layers_1_norm2_bias = const()[name = tensor("encoder_5_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26217472)))]; tensor encoder_5_transformer_layers_1_norm2_weight = const()[name = tensor("encoder_5_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26218560)))]; tensor encoder_5_transformer_layers_1_ffn_0_weight = const()[name = tensor("encoder_5_transformer_layers_1_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26219648)))]; tensor encoder_5_transformer_layers_1_ffn_2_weight = const()[name = tensor("encoder_5_transformer_layers_1_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27268288)))]; tensor encoder_5_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("encoder_5_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28316928)))]; tensor encoder_5_output_proj_weight = const()[name = tensor("encoder_5_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28318016)))]; tensor encoder_7_input_proj_weight = const()[name = tensor("encoder_7_input_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28711296)))]; tensor encoder_7_transformer_layers_0_norm1_bias = const()[name = tensor("encoder_7_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29497792)))]; tensor encoder_7_transformer_layers_0_norm1_weight = const()[name = tensor("encoder_7_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29498880)))]; tensor encoder_7_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("encoder_7_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29499968)))]; tensor encoder_7_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("encoder_7_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30286464)))]; tensor encoder_7_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("encoder_7_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30548672)))]; tensor encoder_7_transformer_layers_0_norm2_bias = const()[name = tensor("encoder_7_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30549760)))]; tensor encoder_7_transformer_layers_0_norm2_weight = const()[name = tensor("encoder_7_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30550848)))]; tensor encoder_7_transformer_layers_0_ffn_0_weight = const()[name = tensor("encoder_7_transformer_layers_0_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30551936)))]; tensor encoder_7_transformer_layers_0_ffn_2_weight = const()[name = tensor("encoder_7_transformer_layers_0_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31600576)))]; tensor encoder_7_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("encoder_7_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32649216)))]; tensor encoder_7_transformer_layers_1_norm1_bias = const()[name = tensor("encoder_7_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32650304)))]; tensor encoder_7_transformer_layers_1_norm1_weight = const()[name = tensor("encoder_7_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32651392)))]; tensor encoder_7_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("encoder_7_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32652480)))]; tensor encoder_7_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("encoder_7_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33438976)))]; tensor encoder_7_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("encoder_7_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33701184)))]; tensor encoder_7_transformer_layers_1_norm2_bias = const()[name = tensor("encoder_7_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33702272)))]; tensor encoder_7_transformer_layers_1_norm2_weight = const()[name = tensor("encoder_7_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33703360)))]; tensor encoder_7_transformer_layers_1_ffn_0_weight = const()[name = tensor("encoder_7_transformer_layers_1_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33704448)))]; tensor encoder_7_transformer_layers_1_ffn_2_weight = const()[name = tensor("encoder_7_transformer_layers_1_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34753088)))]; tensor encoder_7_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("encoder_7_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35801728)))]; tensor encoder_7_transformer_layers_2_norm1_bias = const()[name = tensor("encoder_7_transformer_layers_2_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35802816)))]; tensor encoder_7_transformer_layers_2_norm1_weight = const()[name = tensor("encoder_7_transformer_layers_2_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35803904)))]; tensor encoder_7_transformer_layers_2_self_attn_in_proj_weight = const()[name = tensor("encoder_7_transformer_layers_2_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35804992)))]; tensor encoder_7_transformer_layers_2_self_attn_out_proj_weight = const()[name = tensor("encoder_7_transformer_layers_2_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36591488)))]; tensor encoder_7_transformer_layers_2_layer_scale_1_scale = const()[name = tensor("encoder_7_transformer_layers_2_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36853696)))]; tensor encoder_7_transformer_layers_2_norm2_bias = const()[name = tensor("encoder_7_transformer_layers_2_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36854784)))]; tensor encoder_7_transformer_layers_2_norm2_weight = const()[name = tensor("encoder_7_transformer_layers_2_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36855872)))]; tensor encoder_7_transformer_layers_2_ffn_0_weight = const()[name = tensor("encoder_7_transformer_layers_2_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(36856960)))]; tensor encoder_7_transformer_layers_2_ffn_2_weight = const()[name = tensor("encoder_7_transformer_layers_2_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37905600)))]; tensor encoder_7_transformer_layers_2_layer_scale_2_scale = const()[name = tensor("encoder_7_transformer_layers_2_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38954240)))]; tensor encoder_7_transformer_layers_3_norm1_bias = const()[name = tensor("encoder_7_transformer_layers_3_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38955328)))]; tensor encoder_7_transformer_layers_3_norm1_weight = const()[name = tensor("encoder_7_transformer_layers_3_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38956416)))]; tensor encoder_7_transformer_layers_3_self_attn_in_proj_weight = const()[name = tensor("encoder_7_transformer_layers_3_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38957504)))]; tensor encoder_7_transformer_layers_3_self_attn_out_proj_weight = const()[name = tensor("encoder_7_transformer_layers_3_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39744000)))]; tensor encoder_7_transformer_layers_3_layer_scale_1_scale = const()[name = tensor("encoder_7_transformer_layers_3_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40006208)))]; tensor encoder_7_transformer_layers_3_norm2_bias = const()[name = tensor("encoder_7_transformer_layers_3_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40007296)))]; tensor encoder_7_transformer_layers_3_norm2_weight = const()[name = tensor("encoder_7_transformer_layers_3_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40008384)))]; tensor encoder_7_transformer_layers_3_ffn_0_weight = const()[name = tensor("encoder_7_transformer_layers_3_ffn_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40009472)))]; tensor encoder_7_transformer_layers_3_ffn_2_weight = const()[name = tensor("encoder_7_transformer_layers_3_ffn_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41058112)))]; tensor encoder_7_transformer_layers_3_layer_scale_2_scale = const()[name = tensor("encoder_7_transformer_layers_3_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42106752)))]; tensor encoder_7_output_proj_weight = const()[name = tensor("encoder_7_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42107840)))]; tensor input_proj_bias = const()[name = tensor("input_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42304512)))]; tensor input_proj_weight = const()[name = tensor("input_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42306624)))]; tensor quantizers_0_in_proj_bias = const()[name = tensor("quantizers_0_in_proj_bias"), val = tensor([-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 quantizers_0_in_proj_weight = const()[name = tensor("quantizers_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43879552)))]; tensor quantizers_0_out_proj_bias = const()[name = tensor("quantizers_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43896000)))]; tensor quantizers_0_out_proj_weight = const()[name = tensor("quantizers_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43898112)))]; tensor quantizers_1_in_proj_bias = const()[name = tensor("quantizers_1_in_proj_bias"), val = tensor([-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 quantizers_1_in_proj_weight = const()[name = tensor("quantizers_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43914560)))]; tensor quantizers_1_out_proj_bias = const()[name = tensor("quantizers_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43931008)))]; tensor quantizers_1_out_proj_weight = const()[name = tensor("quantizers_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43933120)))]; tensor quantizers_2_in_proj_bias = const()[name = tensor("quantizers_2_in_proj_bias"), val = tensor([-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 quantizers_2_in_proj_weight = const()[name = tensor("quantizers_2_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43949568)))]; tensor quantizers_2_out_proj_bias = const()[name = tensor("quantizers_2_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43966016)))]; tensor quantizers_2_out_proj_weight = const()[name = tensor("quantizers_2_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43968128)))]; tensor quantizers_3_in_proj_bias = const()[name = tensor("quantizers_3_in_proj_bias"), val = tensor([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 quantizers_3_in_proj_weight = const()[name = tensor("quantizers_3_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43984576)))]; tensor quantizers_3_out_proj_bias = const()[name = tensor("quantizers_3_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44001024)))]; tensor quantizers_3_out_proj_weight = const()[name = tensor("quantizers_3_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44003136)))]; tensor quantizers_4_in_proj_bias = const()[name = tensor("quantizers_4_in_proj_bias"), val = tensor([-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 quantizers_4_in_proj_weight = const()[name = tensor("quantizers_4_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44019584)))]; tensor quantizers_4_out_proj_bias = const()[name = tensor("quantizers_4_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44036032)))]; tensor quantizers_4_out_proj_weight = const()[name = tensor("quantizers_4_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44038144)))]; tensor quantizers_5_in_proj_bias = const()[name = tensor("quantizers_5_in_proj_bias"), val = tensor([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 quantizers_5_in_proj_weight = const()[name = tensor("quantizers_5_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44054592)))]; tensor quantizers_5_out_proj_bias = const()[name = tensor("quantizers_5_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44071040)))]; tensor quantizers_5_out_proj_weight = const()[name = tensor("quantizers_5_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44073152)))]; tensor quantizers_6_in_proj_bias = const()[name = tensor("quantizers_6_in_proj_bias"), val = tensor([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 quantizers_6_in_proj_weight = const()[name = tensor("quantizers_6_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44089600)))]; tensor quantizers_6_out_proj_bias = const()[name = tensor("quantizers_6_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44106048)))]; tensor quantizers_6_out_proj_weight = const()[name = tensor("quantizers_6_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44108160)))]; tensor quantizers_7_in_proj_bias = const()[name = tensor("quantizers_7_in_proj_bias"), val = tensor([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 quantizers_7_in_proj_weight = const()[name = tensor("quantizers_7_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44124608)))]; tensor quantizers_7_out_proj_bias = const()[name = tensor("quantizers_7_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44141056)))]; tensor quantizers_7_out_proj_weight = const()[name = tensor("quantizers_7_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44143168)))]; tensor quantizers_8_in_proj_bias = const()[name = tensor("quantizers_8_in_proj_bias"), val = tensor([-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 quantizers_8_in_proj_weight = const()[name = tensor("quantizers_8_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44159616)))]; tensor quantizers_8_out_proj_bias = const()[name = tensor("quantizers_8_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44176064)))]; tensor quantizers_8_out_proj_weight = const()[name = tensor("quantizers_8_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44178176)))]; tensor quantizers_9_in_proj_bias = const()[name = tensor("quantizers_9_in_proj_bias"), val = tensor([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 quantizers_9_in_proj_weight = const()[name = tensor("quantizers_9_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44194624)))]; tensor quantizers_9_out_proj_bias = const()[name = tensor("quantizers_9_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44211072)))]; tensor quantizers_9_out_proj_weight = const()[name = tensor("quantizers_9_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44213184)))]; tensor quantizers_10_in_proj_bias = const()[name = tensor("quantizers_10_in_proj_bias"), val = tensor([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 quantizers_10_in_proj_weight = const()[name = tensor("quantizers_10_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44229632)))]; tensor quantizers_10_out_proj_bias = const()[name = tensor("quantizers_10_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44246080)))]; tensor quantizers_10_out_proj_weight = const()[name = tensor("quantizers_10_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44248192)))]; tensor quantizers_11_in_proj_bias = const()[name = tensor("quantizers_11_in_proj_bias"), val = tensor([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 quantizers_11_in_proj_weight = const()[name = tensor("quantizers_11_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44264640)))]; tensor quantizers_11_out_proj_bias = const()[name = tensor("quantizers_11_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44281088)))]; tensor quantizers_11_out_proj_weight = const()[name = tensor("quantizers_11_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44283200)))]; tensor quantizers_12_in_proj_bias = const()[name = tensor("quantizers_12_in_proj_bias"), val = tensor([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 quantizers_12_in_proj_weight = const()[name = tensor("quantizers_12_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44299648)))]; tensor quantizers_12_out_proj_bias = const()[name = tensor("quantizers_12_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44316096)))]; tensor quantizers_12_out_proj_weight = const()[name = tensor("quantizers_12_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44318208)))]; tensor quantizers_13_in_proj_bias = const()[name = tensor("quantizers_13_in_proj_bias"), val = tensor([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 quantizers_13_in_proj_weight = const()[name = tensor("quantizers_13_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44334656)))]; tensor quantizers_13_out_proj_bias = const()[name = tensor("quantizers_13_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44351104)))]; tensor quantizers_13_out_proj_weight = const()[name = tensor("quantizers_13_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44353216)))]; tensor quantizers_14_in_proj_bias = const()[name = tensor("quantizers_14_in_proj_bias"), val = tensor([-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 quantizers_14_in_proj_weight = const()[name = tensor("quantizers_14_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44369664)))]; tensor quantizers_14_out_proj_bias = const()[name = tensor("quantizers_14_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44386112)))]; tensor quantizers_14_out_proj_weight = const()[name = tensor("quantizers_14_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44388224)))]; tensor quantizers_15_in_proj_bias = const()[name = tensor("quantizers_15_in_proj_bias"), val = tensor([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 quantizers_15_in_proj_weight = const()[name = tensor("quantizers_15_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44404672)))]; tensor var_215_perm_0 = const()[name = tensor("op_215_perm_0"), val = tensor([0, 2, 1])]; tensor var_222 = const()[name = tensor("op_222"), val = tensor([1, 1, -1])]; tensor var_215 = transpose(perm = var_215_perm_0, x = audio)[name = tensor("transpose_84")]; tensor x_1 = reshape(shape = var_222, x = var_215)[name = tensor("x_1")]; tensor var_231_begin_0 = const()[name = tensor("op_231_begin_0"), val = tensor([0, 0, 0])]; tensor var_231_end_0 = const()[name = tensor("op_231_end_0"), val = tensor([1, 1, 0])]; tensor var_231_end_mask_0 = const()[name = tensor("op_231_end_mask_0"), val = tensor([true, false, true])]; tensor var_231_squeeze_mask_0 = const()[name = tensor("op_231_squeeze_mask_0"), val = tensor([false, true, false])]; tensor 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("op_231")]; tensor fill_like_0_value_0 = const()[name = tensor("fill_like_0_value_0"), val = tensor(0x1p+0)]; tensor fill_like_0 = fill_like(ref_tensor = var_231, value = fill_like_0_value_0)[name = tensor("fill_like_0")]; tensor var_247_axes_0 = const()[name = tensor("op_247_axes_0"), val = tensor([-1])]; tensor var_247_keep_dims_0 = const()[name = tensor("op_247_keep_dims_0"), val = tensor(false)]; tensor var_247 = reduce_sum(axes = var_247_axes_0, keep_dims = var_247_keep_dims_0, x = fill_like_0)[name = tensor("op_247")]; tensor var_252_dtype_0 = const()[name = tensor("op_252_dtype_0"), val = tensor("int32")]; tensor var_253 = const()[name = tensor("op_253"), val = tensor(240)]; tensor var_263 = const()[name = tensor("op_263"), val = tensor([1, 1, -1, 240])]; tensor var_264 = reshape(shape = var_263, x = x_1)[name = tensor("op_264")]; tensor var_265 = const()[name = tensor("op_265"), val = tensor([0, 1, 3, 2])]; tensor var_270 = const()[name = tensor("op_270"), val = tensor([1, 240, -1])]; tensor var_266 = transpose(perm = var_265, x = var_264)[name = tensor("transpose_83")]; tensor x_3 = reshape(shape = var_270, x = var_266)[name = tensor("x_3")]; tensor var_252 = cast(dtype = var_252_dtype_0, x = var_247)[name = tensor("cast_478")]; tensor floor_div_0 = floor_div(x = var_252, y = var_253)[name = tensor("floor_div_0")]; tensor input_lengths_1_dtype_0 = const()[name = tensor("input_lengths_1_dtype_0"), val = tensor("fp32")]; tensor var_277 = const()[name = tensor("op_277"), val = tensor(0)]; tensor var_284 = const()[name = tensor("op_284"), val = tensor(1600)]; tensor var_291 = const()[name = tensor("op_291"), val = tensor(0x1.4f8b58p-17)]; tensor var_295 = const()[name = tensor("op_295"), val = tensor(1)]; tensor input_1_perm_0 = const()[name = tensor("input_1_perm_0"), val = tensor([0, 2, 1])]; tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44421120)))]; tensor input_1 = transpose(perm = input_1_perm_0, x = x_3)[name = tensor("transpose_82")]; tensor x_5 = linear(bias = linear_0_bias_0, weight = encoder_1_input_proj_weight, x = input_1)[name = tensor("linear_0")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; tensor 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("x_7")]; tensor var_323_shape = shape(x = x_7)[name = tensor("op_323_shape")]; tensor gather_4_batch_dims_0 = const()[name = tensor("gather_4_batch_dims_0"), val = tensor(0)]; tensor gather_4_validate_indices_0 = const()[name = tensor("gather_4_validate_indices_0"), val = tensor(false)]; tensor select_16 = const()[name = tensor("select_16"), val = tensor(1)]; tensor gather_4_axis_1 = const()[name = tensor("gather_4_axis_1"), val = tensor(0)]; tensor 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("gather_4")]; tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44422208)))]; tensor projected_1 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_0_self_attn_in_proj_weight, x = x_7)[name = tensor("linear_1")]; tensor concat_0x = const()[name = tensor("concat_0x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_330 = reshape(shape = concat_0x, x = projected_1)[name = tensor("op_330")]; tensor var_331 = const()[name = tensor("op_331"), val = tensor([2, 0, 3, 1, 4])]; tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_3 = transpose(perm = var_331, x = var_330)[name = tensor("transpose_81")]; tensor 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("q_1")]; tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_1")]; tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_1")]; tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44425344)))]; tensor var_343_shape = shape(x = q_1)[name = tensor("op_343_shape")]; tensor gather_8_batch_dims_0 = const()[name = tensor("gather_8_batch_dims_0"), val = tensor(0)]; tensor gather_8_validate_indices_0 = const()[name = tensor("gather_8_validate_indices_0"), val = tensor(false)]; tensor select_17 = const()[name = tensor("select_17"), val = tensor(2)]; tensor gather_8_axis_1 = const()[name = tensor("gather_8_axis_1"), val = tensor(0)]; tensor 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("gather_8")]; tensor var_346 = const()[name = tensor("op_346"), val = tensor([[0x0p+0]])]; tensor const_3 = const()[name = tensor("const_3"), val = tensor(0)]; tensor const_4 = const()[name = tensor("const_4"), val = tensor(1)]; tensor var_347 = range_1d(end = gather_8, start = const_3, step = const_4)[name = tensor("op_347")]; tensor var_347_promoted_dtype_0 = const()[name = tensor("op_347_promoted_dtype_0"), val = tensor("fp32")]; tensor var_347_promoted = cast(dtype = var_347_promoted_dtype_0, x = var_347)[name = tensor("cast_476")]; tensor ts_1 = add(x = var_346, y = var_347_promoted)[name = tensor("ts_1")]; tensor var_350 = const()[name = tensor("op_350"), val = tensor([1, 1, -1, 1])]; tensor ts_3 = reshape(shape = var_350, x = ts_1)[name = tensor("ts_3")]; tensor concat_1x = const()[name = tensor("concat_1x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_1 = reshape(shape = concat_1x, x = q_1)[name = tensor("q2_1")]; tensor concat_2x = const()[name = tensor("concat_2x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_1 = reshape(shape = concat_2x, x = k_1)[name = tensor("k2_1")]; tensor qr_1_begin_0 = const()[name = tensor("qr_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_1_end_0 = const()[name = tensor("qr_1_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_1_end_mask_0 = const()[name = tensor("qr_1_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_1_squeeze_mask_0 = const()[name = tensor("qr_1_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_1")]; tensor qi_1_begin_0 = const()[name = tensor("qi_1_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_1_end_0 = const()[name = tensor("qi_1_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_1_end_mask_0 = const()[name = tensor("qi_1_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_1_squeeze_mask_0 = const()[name = tensor("qi_1_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_1")]; tensor kr_1_begin_0 = const()[name = tensor("kr_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_1_end_0 = const()[name = tensor("kr_1_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_1_end_mask_0 = const()[name = tensor("kr_1_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_1_squeeze_mask_0 = const()[name = tensor("kr_1_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_1")]; tensor ki_1_begin_0 = const()[name = tensor("ki_1_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_1_end_0 = const()[name = tensor("ki_1_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_1_end_mask_0 = const()[name = tensor("ki_1_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_1_squeeze_mask_0 = const()[name = tensor("ki_1_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_1")]; tensor var_363 = mul(x = freqs_1, y = ts_3)[name = tensor("op_363")]; tensor rotr_1 = cos(x = var_363)[name = tensor("rotr_1")]; tensor roti_1 = sin(x = var_363)[name = tensor("roti_1")]; tensor var_367 = mul(x = qr_1, y = rotr_1)[name = tensor("op_367")]; tensor var_368 = mul(x = qi_1, y = roti_1)[name = tensor("op_368")]; tensor var_369 = sub(x = var_367, y = var_368)[name = tensor("op_369")]; tensor var_370 = mul(x = qr_1, y = roti_1)[name = tensor("op_370")]; tensor var_371 = mul(x = qi_1, y = rotr_1)[name = tensor("op_371")]; tensor var_372 = add(x = var_370, y = var_371)[name = tensor("op_372")]; tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; tensor qo_1 = stack(axis = qo_1_axis_0, values = (var_369, var_372))[name = tensor("qo_1")]; tensor var_375 = mul(x = kr_1, y = rotr_1)[name = tensor("op_375")]; tensor var_376 = mul(x = ki_1, y = roti_1)[name = tensor("op_376")]; tensor var_377 = sub(x = var_375, y = var_376)[name = tensor("op_377")]; tensor var_378 = mul(x = kr_1, y = roti_1)[name = tensor("op_378")]; tensor var_379 = mul(x = ki_1, y = rotr_1)[name = tensor("op_379")]; tensor var_380 = add(x = var_378, y = var_379)[name = tensor("op_380")]; tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; tensor ko_1 = stack(axis = ko_1_axis_0, values = (var_377, var_380))[name = tensor("ko_1")]; tensor concat_3x = const()[name = tensor("concat_3x"), val = tensor([1, 4, -1, 64])]; tensor q_3 = reshape(shape = concat_3x, x = qo_1)[name = tensor("q_3")]; tensor concat_4x = const()[name = tensor("concat_4x"), val = tensor([1, 4, -1, 64])]; tensor k_3 = reshape(shape = concat_4x, x = ko_1)[name = tensor("k_3")]; tensor const_5 = const()[name = tensor("const_5"), val = tensor(0)]; tensor const_6 = const()[name = tensor("const_6"), val = tensor(1)]; tensor positions_1 = range_1d(end = gather_4, start = const_5, step = const_6)[name = tensor("positions_1")]; tensor concat_5x = const()[name = tensor("concat_5x"), val = tensor([1, 1, -1])]; tensor var_392 = reshape(shape = concat_5x, x = positions_1)[name = tensor("op_392")]; tensor var_393 = const()[name = tensor("op_393"), val = tensor([-1, 1, 1])]; tensor input_lengths_1 = cast(dtype = input_lengths_1_dtype_0, x = floor_div_0)[name = tensor("cast_477")]; tensor var_394 = reshape(shape = var_393, x = input_lengths_1)[name = tensor("op_394")]; tensor var_392_promoted_dtype_0 = const()[name = tensor("op_392_promoted_dtype_0"), val = tensor("fp32")]; tensor var_392_promoted = cast(dtype = var_392_promoted_dtype_0, x = var_392)[name = tensor("cast_475")]; tensor valid_k_1 = less(x = var_392_promoted, y = var_394)[name = tensor("valid_k_1")]; tensor concat_6x = const()[name = tensor("concat_6x"), val = tensor([1, -1, 1])]; tensor var_397 = reshape(shape = concat_6x, x = positions_1)[name = tensor("op_397")]; tensor delta_1 = sub(x = var_397, y = var_392)[name = tensor("delta_1")]; tensor concat_8_axis_0 = const()[name = tensor("concat_8_axis_0"), val = tensor(0)]; tensor concat_8_interleave_0 = const()[name = tensor("concat_8_interleave_0"), val = tensor(false)]; tensor concat_8 = concat(axis = concat_8_axis_0, interleave = concat_8_interleave_0, values = (var_295, gather_4, gather_4))[name = tensor("concat_8")]; tensor fill_0_value_0 = const()[name = tensor("fill_0_value_0"), val = tensor(0x1p+0)]; tensor fill_0 = fill(shape = concat_8, value = fill_0_value_0)[name = tensor("fill_0")]; tensor attn_bias_1_dtype_0 = const()[name = tensor("attn_bias_1_dtype_0"), val = tensor("bool")]; tensor var_403 = greater_equal(x = delta_1, y = var_277)[name = tensor("op_403")]; tensor attn_bias_1 = cast(dtype = attn_bias_1_dtype_0, x = fill_0)[name = tensor("cast_474")]; tensor attn_bias_3 = logical_and(x = attn_bias_1, y = var_403)[name = tensor("attn_bias_3")]; tensor var_405 = less(x = delta_1, y = var_284)[name = tensor("op_405")]; tensor attn_bias_5 = logical_and(x = attn_bias_3, y = var_405)[name = tensor("attn_bias_5")]; tensor var_407 = logical_and(x = attn_bias_5, y = valid_k_1)[name = tensor("op_407")]; tensor var_409_axes_0 = const()[name = tensor("op_409_axes_0"), val = tensor([1])]; tensor var_409 = expand_dims(axes = var_409_axes_0, x = var_407)[name = tensor("op_409")]; tensor cast_20_dtype_0 = const()[name = tensor("cast_20_dtype_0"), val = tensor("fp32")]; tensor sub_0_x_0 = const()[name = tensor("sub_0_x_0"), val = tensor(0x1p+0)]; tensor cast_20 = cast(dtype = cast_20_dtype_0, x = var_409)[name = tensor("cast_473")]; tensor sub_0 = sub(x = sub_0_x_0, y = cast_20)[name = tensor("sub_0")]; tensor mul_0_x_0 = const()[name = tensor("mul_0_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_0 = mul(x = mul_0_x_0, y = sub_0)[name = tensor("mul_0")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_3, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_0")]; tensor add_0 = add(x = matmul_0, y = mul_0)[name = tensor("add_0")]; tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = add_0)[name = tensor("softmax_0")]; tensor out_1_transpose_x_0 = const()[name = tensor("out_1_transpose_x_0"), val = tensor(false)]; tensor out_1_transpose_y_0 = const()[name = tensor("out_1_transpose_y_0"), val = tensor(false)]; tensor 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("out_1")]; tensor concat_9x = const()[name = tensor("concat_9x"), val = tensor([1, -1])]; tensor var_415 = reshape(shape = concat_9x, x = positions_1)[name = tensor("op_415")]; tensor var_416 = const()[name = tensor("op_416"), val = tensor([-1, 1])]; tensor var_417 = reshape(shape = var_416, x = input_lengths_1)[name = tensor("op_417")]; tensor var_415_promoted_dtype_0 = const()[name = tensor("op_415_promoted_dtype_0"), val = tensor("fp32")]; tensor var_415_promoted = cast(dtype = var_415_promoted_dtype_0, x = var_415)[name = tensor("cast_472")]; tensor var_418 = less(x = var_415_promoted, y = var_417)[name = tensor("op_418")]; tensor concat_10x = const()[name = tensor("concat_10x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_1 = reshape(shape = concat_10x, x = var_418)[name = tensor("valid_q_1")]; tensor var_422 = const()[name = tensor("op_422"), val = tensor(0x0p+0)]; tensor out_3 = select(a = out_1, b = var_422, cond = valid_q_1)[name = tensor("out_3")]; tensor var_424_perm_0 = const()[name = tensor("op_424_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_11x = const()[name = tensor("concat_11x"), val = tensor([1, -1, 256])]; tensor var_424 = transpose(perm = var_424_perm_0, x = out_3)[name = tensor("transpose_80")]; tensor input_3 = reshape(shape = concat_11x, x = var_424)[name = tensor("input_3")]; tensor x_9 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_0_self_attn_out_proj_weight, x = input_3)[name = tensor("linear_2")]; tensor var_430 = mul(x = encoder_1_transformer_layers_0_layer_scale_1_scale, y = x_9)[name = tensor("op_430")]; tensor input_5 = add(x = x_5, y = var_430)[name = tensor("input_5")]; tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; tensor 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("input_7")]; tensor linear_3_bias_0 = const()[name = tensor("linear_3_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44425536)))]; tensor input_9 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_0_ffn_0_weight, x = input_7)[name = tensor("linear_3")]; tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; tensor input_11 = gelu(mode = input_11_mode_0, x = input_9)[name = tensor("input_11")]; tensor x_11 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_0_ffn_2_weight, x = input_11)[name = tensor("linear_4")]; tensor var_445 = mul(x = encoder_1_transformer_layers_0_layer_scale_2_scale, y = x_11)[name = tensor("op_445")]; tensor input_13 = add(x = input_5, y = var_445)[name = tensor("input_13")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; tensor 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("x_13")]; tensor var_460_shape = shape(x = x_13)[name = tensor("op_460_shape")]; tensor gather_14_batch_dims_0 = const()[name = tensor("gather_14_batch_dims_0"), val = tensor(0)]; tensor gather_14_validate_indices_0 = const()[name = tensor("gather_14_validate_indices_0"), val = tensor(false)]; tensor select_18 = const()[name = tensor("select_18"), val = tensor(1)]; tensor gather_14_axis_1 = const()[name = tensor("gather_14_axis_1"), val = tensor(0)]; tensor 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("gather_14")]; tensor projected_5 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_1_self_attn_in_proj_weight, x = x_13)[name = tensor("linear_5")]; tensor concat_12x = const()[name = tensor("concat_12x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_467 = reshape(shape = concat_12x, x = projected_5)[name = tensor("op_467")]; tensor var_468 = const()[name = tensor("op_468"), val = tensor([2, 0, 3, 1, 4])]; tensor q_5_begin_0 = const()[name = tensor("q_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_5_end_0 = const()[name = tensor("q_5_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_5_end_mask_0 = const()[name = tensor("q_5_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_5_squeeze_mask_0 = const()[name = tensor("q_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_7 = transpose(perm = var_468, x = var_467)[name = tensor("transpose_79")]; tensor 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("q_5")]; tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_5")]; tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_3")]; tensor freqs_3 = const()[name = tensor("freqs_3"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44429696)))]; tensor var_480_shape = shape(x = q_5)[name = tensor("op_480_shape")]; tensor gather_18_batch_dims_0 = const()[name = tensor("gather_18_batch_dims_0"), val = tensor(0)]; tensor gather_18_validate_indices_0 = const()[name = tensor("gather_18_validate_indices_0"), val = tensor(false)]; tensor select_19 = const()[name = tensor("select_19"), val = tensor(2)]; tensor gather_18_axis_1 = const()[name = tensor("gather_18_axis_1"), val = tensor(0)]; tensor 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("gather_18")]; tensor var_483 = const()[name = tensor("op_483"), val = tensor([[0x0p+0]])]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0)]; tensor const_13 = const()[name = tensor("const_13"), val = tensor(1)]; tensor var_484 = range_1d(end = gather_18, start = const_12, step = const_13)[name = tensor("op_484")]; tensor var_484_promoted_dtype_0 = const()[name = tensor("op_484_promoted_dtype_0"), val = tensor("fp32")]; tensor var_484_promoted = cast(dtype = var_484_promoted_dtype_0, x = var_484)[name = tensor("cast_471")]; tensor ts_5 = add(x = var_483, y = var_484_promoted)[name = tensor("ts_5")]; tensor var_487 = const()[name = tensor("op_487"), val = tensor([1, 1, -1, 1])]; tensor ts_7 = reshape(shape = var_487, x = ts_5)[name = tensor("ts_7")]; tensor concat_13x = const()[name = tensor("concat_13x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_3 = reshape(shape = concat_13x, x = q_5)[name = tensor("q2_3")]; tensor concat_14x = const()[name = tensor("concat_14x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_3 = reshape(shape = concat_14x, x = k_5)[name = tensor("k2_3")]; tensor qr_3_begin_0 = const()[name = tensor("qr_3_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_3_end_0 = const()[name = tensor("qr_3_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_3_end_mask_0 = const()[name = tensor("qr_3_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_3_squeeze_mask_0 = const()[name = tensor("qr_3_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_3")]; tensor qi_3_begin_0 = const()[name = tensor("qi_3_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_3_end_0 = const()[name = tensor("qi_3_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_3_end_mask_0 = const()[name = tensor("qi_3_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_3_squeeze_mask_0 = const()[name = tensor("qi_3_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_3")]; tensor kr_3_begin_0 = const()[name = tensor("kr_3_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_3_end_0 = const()[name = tensor("kr_3_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_3_end_mask_0 = const()[name = tensor("kr_3_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_3_squeeze_mask_0 = const()[name = tensor("kr_3_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_3")]; tensor ki_3_begin_0 = const()[name = tensor("ki_3_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_3_end_0 = const()[name = tensor("ki_3_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_3_end_mask_0 = const()[name = tensor("ki_3_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_3_squeeze_mask_0 = const()[name = tensor("ki_3_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_3")]; tensor var_500 = mul(x = freqs_3, y = ts_7)[name = tensor("op_500")]; tensor rotr_3 = cos(x = var_500)[name = tensor("rotr_3")]; tensor roti_3 = sin(x = var_500)[name = tensor("roti_3")]; tensor var_504 = mul(x = qr_3, y = rotr_3)[name = tensor("op_504")]; tensor var_505 = mul(x = qi_3, y = roti_3)[name = tensor("op_505")]; tensor var_506 = sub(x = var_504, y = var_505)[name = tensor("op_506")]; tensor var_507 = mul(x = qr_3, y = roti_3)[name = tensor("op_507")]; tensor var_508 = mul(x = qi_3, y = rotr_3)[name = tensor("op_508")]; tensor var_509 = add(x = var_507, y = var_508)[name = tensor("op_509")]; tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; tensor qo_3 = stack(axis = qo_3_axis_0, values = (var_506, var_509))[name = tensor("qo_3")]; tensor var_512 = mul(x = kr_3, y = rotr_3)[name = tensor("op_512")]; tensor var_513 = mul(x = ki_3, y = roti_3)[name = tensor("op_513")]; tensor var_514 = sub(x = var_512, y = var_513)[name = tensor("op_514")]; tensor var_515 = mul(x = kr_3, y = roti_3)[name = tensor("op_515")]; tensor var_516 = mul(x = ki_3, y = rotr_3)[name = tensor("op_516")]; tensor var_517 = add(x = var_515, y = var_516)[name = tensor("op_517")]; tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; tensor ko_3 = stack(axis = ko_3_axis_0, values = (var_514, var_517))[name = tensor("ko_3")]; tensor concat_15x = const()[name = tensor("concat_15x"), val = tensor([1, 4, -1, 64])]; tensor q_7 = reshape(shape = concat_15x, x = qo_3)[name = tensor("q_7")]; tensor concat_16x = const()[name = tensor("concat_16x"), val = tensor([1, 4, -1, 64])]; tensor k_7 = reshape(shape = concat_16x, x = ko_3)[name = tensor("k_7")]; tensor const_14 = const()[name = tensor("const_14"), val = tensor(0)]; tensor const_15 = const()[name = tensor("const_15"), val = tensor(1)]; tensor positions_3 = range_1d(end = gather_14, start = const_14, step = const_15)[name = tensor("positions_3")]; tensor concat_17x = const()[name = tensor("concat_17x"), val = tensor([1, 1, -1])]; tensor var_529 = reshape(shape = concat_17x, x = positions_3)[name = tensor("op_529")]; tensor var_529_promoted_dtype_0 = const()[name = tensor("op_529_promoted_dtype_0"), val = tensor("fp32")]; tensor var_529_promoted = cast(dtype = var_529_promoted_dtype_0, x = var_529)[name = tensor("cast_470")]; tensor valid_k_3 = less(x = var_529_promoted, y = var_394)[name = tensor("valid_k_3")]; tensor concat_18x = const()[name = tensor("concat_18x"), val = tensor([1, -1, 1])]; tensor var_534 = reshape(shape = concat_18x, x = positions_3)[name = tensor("op_534")]; tensor delta_3 = sub(x = var_534, y = var_529)[name = tensor("delta_3")]; tensor concat_20_axis_0 = const()[name = tensor("concat_20_axis_0"), val = tensor(0)]; tensor concat_20_interleave_0 = const()[name = tensor("concat_20_interleave_0"), val = tensor(false)]; tensor concat_20 = concat(axis = concat_20_axis_0, interleave = concat_20_interleave_0, values = (var_295, gather_14, gather_14))[name = tensor("concat_20")]; tensor fill_1_value_0 = const()[name = tensor("fill_1_value_0"), val = tensor(0x1p+0)]; tensor fill_1 = fill(shape = concat_20, value = fill_1_value_0)[name = tensor("fill_1")]; tensor attn_bias_9_dtype_0 = const()[name = tensor("attn_bias_9_dtype_0"), val = tensor("bool")]; tensor var_540 = greater_equal(x = delta_3, y = var_277)[name = tensor("op_540")]; tensor attn_bias_9 = cast(dtype = attn_bias_9_dtype_0, x = fill_1)[name = tensor("cast_469")]; tensor attn_bias_11 = logical_and(x = attn_bias_9, y = var_540)[name = tensor("attn_bias_11")]; tensor var_542 = less(x = delta_3, y = var_284)[name = tensor("op_542")]; tensor attn_bias_13 = logical_and(x = attn_bias_11, y = var_542)[name = tensor("attn_bias_13")]; tensor var_544 = logical_and(x = attn_bias_13, y = valid_k_3)[name = tensor("op_544")]; tensor var_546_axes_0 = const()[name = tensor("op_546_axes_0"), val = tensor([1])]; tensor var_546 = expand_dims(axes = var_546_axes_0, x = var_544)[name = tensor("op_546")]; tensor cast_42_dtype_0 = const()[name = tensor("cast_42_dtype_0"), val = tensor("fp32")]; tensor sub_1_x_0 = const()[name = tensor("sub_1_x_0"), val = tensor(0x1p+0)]; tensor cast_42 = cast(dtype = cast_42_dtype_0, x = var_546)[name = tensor("cast_468")]; tensor sub_1 = sub(x = sub_1_x_0, y = cast_42)[name = tensor("sub_1")]; tensor mul_2_x_0 = const()[name = tensor("mul_2_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_2 = mul(x = mul_2_x_0, y = sub_1)[name = tensor("mul_2")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q_7, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_1")]; tensor add_1 = add(x = matmul_1, y = mul_2)[name = tensor("add_1")]; tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = add_1)[name = tensor("softmax_1")]; tensor out_5_transpose_x_0 = const()[name = tensor("out_5_transpose_x_0"), val = tensor(false)]; tensor out_5_transpose_y_0 = const()[name = tensor("out_5_transpose_y_0"), val = tensor(false)]; tensor 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("out_5")]; tensor concat_21x = const()[name = tensor("concat_21x"), val = tensor([1, -1])]; tensor var_552 = reshape(shape = concat_21x, x = positions_3)[name = tensor("op_552")]; tensor var_552_promoted_dtype_0 = const()[name = tensor("op_552_promoted_dtype_0"), val = tensor("fp32")]; tensor var_552_promoted = cast(dtype = var_552_promoted_dtype_0, x = var_552)[name = tensor("cast_467")]; tensor var_555 = less(x = var_552_promoted, y = var_417)[name = tensor("op_555")]; tensor concat_22x = const()[name = tensor("concat_22x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_3 = reshape(shape = concat_22x, x = var_555)[name = tensor("valid_q_3")]; tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x0p+0)]; tensor out_7 = select(a = out_5, b = var_559, cond = valid_q_3)[name = tensor("out_7")]; tensor var_561_perm_0 = const()[name = tensor("op_561_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_23x = const()[name = tensor("concat_23x"), val = tensor([1, -1, 256])]; tensor var_561 = transpose(perm = var_561_perm_0, x = out_7)[name = tensor("transpose_78")]; tensor input_15 = reshape(shape = concat_23x, x = var_561)[name = tensor("input_15")]; tensor x_15 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_6")]; tensor var_567 = mul(x = encoder_1_transformer_layers_1_layer_scale_1_scale, y = x_15)[name = tensor("op_567")]; tensor input_17 = add(x = input_13, y = var_567)[name = tensor("input_17")]; tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; tensor 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("input_19")]; tensor input_21 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_1_ffn_0_weight, x = input_19)[name = tensor("linear_7")]; tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; tensor input_23 = gelu(mode = input_23_mode_0, x = input_21)[name = tensor("input_23")]; tensor x_17 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_1_ffn_2_weight, x = input_23)[name = tensor("linear_8")]; tensor var_582 = mul(x = encoder_1_transformer_layers_1_layer_scale_2_scale, y = x_17)[name = tensor("op_582")]; tensor input_25 = add(x = input_17, y = var_582)[name = tensor("input_25")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; tensor 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("x_19")]; tensor var_597_shape = shape(x = x_19)[name = tensor("op_597_shape")]; tensor gather_24_batch_dims_0 = const()[name = tensor("gather_24_batch_dims_0"), val = tensor(0)]; tensor gather_24_validate_indices_0 = const()[name = tensor("gather_24_validate_indices_0"), val = tensor(false)]; tensor select_20 = const()[name = tensor("select_20"), val = tensor(1)]; tensor gather_24_axis_1 = const()[name = tensor("gather_24_axis_1"), val = tensor(0)]; tensor 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("gather_24")]; tensor projected_9 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_2_self_attn_in_proj_weight, x = x_19)[name = tensor("linear_9")]; tensor concat_24x = const()[name = tensor("concat_24x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_604 = reshape(shape = concat_24x, x = projected_9)[name = tensor("op_604")]; tensor var_605 = const()[name = tensor("op_605"), val = tensor([2, 0, 3, 1, 4])]; tensor q_9_begin_0 = const()[name = tensor("q_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_9_end_0 = const()[name = tensor("q_9_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_9_end_mask_0 = const()[name = tensor("q_9_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_9_squeeze_mask_0 = const()[name = tensor("q_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_11 = transpose(perm = var_605, x = var_604)[name = tensor("transpose_77")]; tensor 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("q_9")]; tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_9")]; tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_5")]; tensor freqs_5 = const()[name = tensor("freqs_5"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44429888)))]; tensor var_617_shape = shape(x = q_9)[name = tensor("op_617_shape")]; tensor gather_28_batch_dims_0 = const()[name = tensor("gather_28_batch_dims_0"), val = tensor(0)]; tensor gather_28_validate_indices_0 = const()[name = tensor("gather_28_validate_indices_0"), val = tensor(false)]; tensor select_21 = const()[name = tensor("select_21"), val = tensor(2)]; tensor gather_28_axis_1 = const()[name = tensor("gather_28_axis_1"), val = tensor(0)]; tensor 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("gather_28")]; tensor var_620 = const()[name = tensor("op_620"), val = tensor([[0x0p+0]])]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0)]; tensor const_22 = const()[name = tensor("const_22"), val = tensor(1)]; tensor var_621 = range_1d(end = gather_28, start = const_21, step = const_22)[name = tensor("op_621")]; tensor var_621_promoted_dtype_0 = const()[name = tensor("op_621_promoted_dtype_0"), val = tensor("fp32")]; tensor var_621_promoted = cast(dtype = var_621_promoted_dtype_0, x = var_621)[name = tensor("cast_466")]; tensor ts_9 = add(x = var_620, y = var_621_promoted)[name = tensor("ts_9")]; tensor var_624 = const()[name = tensor("op_624"), val = tensor([1, 1, -1, 1])]; tensor ts_11 = reshape(shape = var_624, x = ts_9)[name = tensor("ts_11")]; tensor concat_25x = const()[name = tensor("concat_25x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_5 = reshape(shape = concat_25x, x = q_9)[name = tensor("q2_5")]; tensor concat_26x = const()[name = tensor("concat_26x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_5 = reshape(shape = concat_26x, x = k_9)[name = tensor("k2_5")]; tensor qr_5_begin_0 = const()[name = tensor("qr_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_5_end_0 = const()[name = tensor("qr_5_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_5_end_mask_0 = const()[name = tensor("qr_5_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_5_squeeze_mask_0 = const()[name = tensor("qr_5_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_5")]; tensor qi_5_begin_0 = const()[name = tensor("qi_5_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_5_end_0 = const()[name = tensor("qi_5_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_5_end_mask_0 = const()[name = tensor("qi_5_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_5_squeeze_mask_0 = const()[name = tensor("qi_5_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_5")]; tensor kr_5_begin_0 = const()[name = tensor("kr_5_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_5_end_0 = const()[name = tensor("kr_5_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_5_end_mask_0 = const()[name = tensor("kr_5_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_5_squeeze_mask_0 = const()[name = tensor("kr_5_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_5")]; tensor ki_5_begin_0 = const()[name = tensor("ki_5_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_5_end_0 = const()[name = tensor("ki_5_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_5_end_mask_0 = const()[name = tensor("ki_5_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_5_squeeze_mask_0 = const()[name = tensor("ki_5_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_5")]; tensor var_637 = mul(x = freqs_5, y = ts_11)[name = tensor("op_637")]; tensor rotr_5 = cos(x = var_637)[name = tensor("rotr_5")]; tensor roti_5 = sin(x = var_637)[name = tensor("roti_5")]; tensor var_641 = mul(x = qr_5, y = rotr_5)[name = tensor("op_641")]; tensor var_642 = mul(x = qi_5, y = roti_5)[name = tensor("op_642")]; tensor var_643 = sub(x = var_641, y = var_642)[name = tensor("op_643")]; tensor var_644 = mul(x = qr_5, y = roti_5)[name = tensor("op_644")]; tensor var_645 = mul(x = qi_5, y = rotr_5)[name = tensor("op_645")]; tensor var_646 = add(x = var_644, y = var_645)[name = tensor("op_646")]; tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; tensor qo_5 = stack(axis = qo_5_axis_0, values = (var_643, var_646))[name = tensor("qo_5")]; tensor var_649 = mul(x = kr_5, y = rotr_5)[name = tensor("op_649")]; tensor var_650 = mul(x = ki_5, y = roti_5)[name = tensor("op_650")]; tensor var_651 = sub(x = var_649, y = var_650)[name = tensor("op_651")]; tensor var_652 = mul(x = kr_5, y = roti_5)[name = tensor("op_652")]; tensor var_653 = mul(x = ki_5, y = rotr_5)[name = tensor("op_653")]; tensor var_654 = add(x = var_652, y = var_653)[name = tensor("op_654")]; tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; tensor ko_5 = stack(axis = ko_5_axis_0, values = (var_651, var_654))[name = tensor("ko_5")]; tensor concat_27x = const()[name = tensor("concat_27x"), val = tensor([1, 4, -1, 64])]; tensor q_11 = reshape(shape = concat_27x, x = qo_5)[name = tensor("q_11")]; tensor concat_28x = const()[name = tensor("concat_28x"), val = tensor([1, 4, -1, 64])]; tensor k_11 = reshape(shape = concat_28x, x = ko_5)[name = tensor("k_11")]; tensor const_23 = const()[name = tensor("const_23"), val = tensor(0)]; tensor const_24 = const()[name = tensor("const_24"), val = tensor(1)]; tensor positions_5 = range_1d(end = gather_24, start = const_23, step = const_24)[name = tensor("positions_5")]; tensor concat_29x = const()[name = tensor("concat_29x"), val = tensor([1, 1, -1])]; tensor var_666 = reshape(shape = concat_29x, x = positions_5)[name = tensor("op_666")]; tensor var_666_promoted_dtype_0 = const()[name = tensor("op_666_promoted_dtype_0"), val = tensor("fp32")]; tensor var_666_promoted = cast(dtype = var_666_promoted_dtype_0, x = var_666)[name = tensor("cast_465")]; tensor valid_k_5 = less(x = var_666_promoted, y = var_394)[name = tensor("valid_k_5")]; tensor concat_30x = const()[name = tensor("concat_30x"), val = tensor([1, -1, 1])]; tensor var_671 = reshape(shape = concat_30x, x = positions_5)[name = tensor("op_671")]; tensor delta_5 = sub(x = var_671, y = var_666)[name = tensor("delta_5")]; tensor concat_32_axis_0 = const()[name = tensor("concat_32_axis_0"), val = tensor(0)]; tensor concat_32_interleave_0 = const()[name = tensor("concat_32_interleave_0"), val = tensor(false)]; tensor concat_32 = concat(axis = concat_32_axis_0, interleave = concat_32_interleave_0, values = (var_295, gather_24, gather_24))[name = tensor("concat_32")]; tensor fill_2_value_0 = const()[name = tensor("fill_2_value_0"), val = tensor(0x1p+0)]; tensor fill_2 = fill(shape = concat_32, value = fill_2_value_0)[name = tensor("fill_2")]; tensor attn_bias_17_dtype_0 = const()[name = tensor("attn_bias_17_dtype_0"), val = tensor("bool")]; tensor var_677 = greater_equal(x = delta_5, y = var_277)[name = tensor("op_677")]; tensor attn_bias_17 = cast(dtype = attn_bias_17_dtype_0, x = fill_2)[name = tensor("cast_464")]; tensor attn_bias_19 = logical_and(x = attn_bias_17, y = var_677)[name = tensor("attn_bias_19")]; tensor var_679 = less(x = delta_5, y = var_284)[name = tensor("op_679")]; tensor attn_bias_21 = logical_and(x = attn_bias_19, y = var_679)[name = tensor("attn_bias_21")]; tensor var_681 = logical_and(x = attn_bias_21, y = valid_k_5)[name = tensor("op_681")]; tensor var_683_axes_0 = const()[name = tensor("op_683_axes_0"), val = tensor([1])]; tensor var_683 = expand_dims(axes = var_683_axes_0, x = var_681)[name = tensor("op_683")]; tensor cast_64_dtype_0 = const()[name = tensor("cast_64_dtype_0"), val = tensor("fp32")]; tensor sub_2_x_0 = const()[name = tensor("sub_2_x_0"), val = tensor(0x1p+0)]; tensor cast_64 = cast(dtype = cast_64_dtype_0, x = var_683)[name = tensor("cast_463")]; tensor sub_2 = sub(x = sub_2_x_0, y = cast_64)[name = tensor("sub_2")]; tensor mul_4_x_0 = const()[name = tensor("mul_4_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_4 = mul(x = mul_4_x_0, y = sub_2)[name = tensor("mul_4")]; tensor mul_5_y_0 = const()[name = tensor("mul_5_y_0"), val = tensor(0x1p-3)]; tensor mul_5 = mul(x = q_11, y = mul_5_y_0)[name = tensor("mul_5")]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(true)]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_2")]; tensor add_2 = add(x = matmul_2, y = mul_4)[name = tensor("add_2")]; tensor softmax_2_axis_0 = const()[name = tensor("softmax_2_axis_0"), val = tensor(-1)]; tensor softmax_2 = softmax(axis = softmax_2_axis_0, x = add_2)[name = tensor("softmax_2")]; tensor out_9_transpose_x_0 = const()[name = tensor("out_9_transpose_x_0"), val = tensor(false)]; tensor out_9_transpose_y_0 = const()[name = tensor("out_9_transpose_y_0"), val = tensor(false)]; tensor 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("out_9")]; tensor concat_33x = const()[name = tensor("concat_33x"), val = tensor([1, -1])]; tensor var_689 = reshape(shape = concat_33x, x = positions_5)[name = tensor("op_689")]; tensor var_689_promoted_dtype_0 = const()[name = tensor("op_689_promoted_dtype_0"), val = tensor("fp32")]; tensor var_689_promoted = cast(dtype = var_689_promoted_dtype_0, x = var_689)[name = tensor("cast_462")]; tensor var_692 = less(x = var_689_promoted, y = var_417)[name = tensor("op_692")]; tensor concat_34x = const()[name = tensor("concat_34x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_5 = reshape(shape = concat_34x, x = var_692)[name = tensor("valid_q_5")]; tensor var_696 = const()[name = tensor("op_696"), val = tensor(0x0p+0)]; tensor out_11 = select(a = out_9, b = var_696, cond = valid_q_5)[name = tensor("out_11")]; tensor var_698_perm_0 = const()[name = tensor("op_698_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_35x = const()[name = tensor("concat_35x"), val = tensor([1, -1, 256])]; tensor var_698 = transpose(perm = var_698_perm_0, x = out_11)[name = tensor("transpose_76")]; tensor input_27 = reshape(shape = concat_35x, x = var_698)[name = tensor("input_27")]; tensor x_21 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_2_self_attn_out_proj_weight, x = input_27)[name = tensor("linear_10")]; tensor var_704 = mul(x = encoder_1_transformer_layers_2_layer_scale_1_scale, y = x_21)[name = tensor("op_704")]; tensor input_29 = add(x = input_25, y = var_704)[name = tensor("input_29")]; tensor input_31_axes_0 = const()[name = tensor("input_31_axes_0"), val = tensor([-1])]; tensor 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("input_31")]; tensor input_33 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_2_ffn_0_weight, x = input_31)[name = tensor("linear_11")]; tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; tensor input_35 = gelu(mode = input_35_mode_0, x = input_33)[name = tensor("input_35")]; tensor x_23 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_2_ffn_2_weight, x = input_35)[name = tensor("linear_12")]; tensor var_719 = mul(x = encoder_1_transformer_layers_2_layer_scale_2_scale, y = x_23)[name = tensor("op_719")]; tensor input_37 = add(x = input_29, y = var_719)[name = tensor("input_37")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; tensor 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("x_25")]; tensor var_734_shape = shape(x = x_25)[name = tensor("op_734_shape")]; tensor gather_34_batch_dims_0 = const()[name = tensor("gather_34_batch_dims_0"), val = tensor(0)]; tensor gather_34_validate_indices_0 = const()[name = tensor("gather_34_validate_indices_0"), val = tensor(false)]; tensor select_22 = const()[name = tensor("select_22"), val = tensor(1)]; tensor gather_34_axis_1 = const()[name = tensor("gather_34_axis_1"), val = tensor(0)]; tensor 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("gather_34")]; tensor projected_13 = linear(bias = linear_1_bias_0, weight = encoder_1_transformer_layers_3_self_attn_in_proj_weight, x = x_25)[name = tensor("linear_13")]; tensor concat_36x = const()[name = tensor("concat_36x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_741 = reshape(shape = concat_36x, x = projected_13)[name = tensor("op_741")]; tensor var_742 = const()[name = tensor("op_742"), val = tensor([2, 0, 3, 1, 4])]; tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_15 = transpose(perm = var_742, x = var_741)[name = tensor("transpose_75")]; tensor 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("q_13")]; tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_13")]; tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_7")]; tensor freqs_7 = const()[name = tensor("freqs_7"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44430080)))]; tensor var_754_shape = shape(x = q_13)[name = tensor("op_754_shape")]; tensor gather_38_batch_dims_0 = const()[name = tensor("gather_38_batch_dims_0"), val = tensor(0)]; tensor gather_38_validate_indices_0 = const()[name = tensor("gather_38_validate_indices_0"), val = tensor(false)]; tensor select_23 = const()[name = tensor("select_23"), val = tensor(2)]; tensor gather_38_axis_1 = const()[name = tensor("gather_38_axis_1"), val = tensor(0)]; tensor 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("gather_38")]; tensor var_757 = const()[name = tensor("op_757"), val = tensor([[0x0p+0]])]; tensor const_30 = const()[name = tensor("const_30"), val = tensor(0)]; tensor const_31 = const()[name = tensor("const_31"), val = tensor(1)]; tensor var_758 = range_1d(end = gather_38, start = const_30, step = const_31)[name = tensor("op_758")]; tensor var_758_promoted_dtype_0 = const()[name = tensor("op_758_promoted_dtype_0"), val = tensor("fp32")]; tensor var_758_promoted = cast(dtype = var_758_promoted_dtype_0, x = var_758)[name = tensor("cast_461")]; tensor ts_13 = add(x = var_757, y = var_758_promoted)[name = tensor("ts_13")]; tensor var_761 = const()[name = tensor("op_761"), val = tensor([1, 1, -1, 1])]; tensor ts_15 = reshape(shape = var_761, x = ts_13)[name = tensor("ts_15")]; tensor concat_37x = const()[name = tensor("concat_37x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_7 = reshape(shape = concat_37x, x = q_13)[name = tensor("q2_7")]; tensor concat_38x = const()[name = tensor("concat_38x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_7 = reshape(shape = concat_38x, x = k_13)[name = tensor("k2_7")]; tensor qr_7_begin_0 = const()[name = tensor("qr_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_7_end_0 = const()[name = tensor("qr_7_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_7_end_mask_0 = const()[name = tensor("qr_7_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_7_squeeze_mask_0 = const()[name = tensor("qr_7_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_7")]; tensor qi_7_begin_0 = const()[name = tensor("qi_7_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_7_end_0 = const()[name = tensor("qi_7_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_7_end_mask_0 = const()[name = tensor("qi_7_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_7_squeeze_mask_0 = const()[name = tensor("qi_7_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_7")]; tensor kr_7_begin_0 = const()[name = tensor("kr_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_7_end_0 = const()[name = tensor("kr_7_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_7_end_mask_0 = const()[name = tensor("kr_7_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_7_squeeze_mask_0 = const()[name = tensor("kr_7_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_7")]; tensor ki_7_begin_0 = const()[name = tensor("ki_7_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_7_end_0 = const()[name = tensor("ki_7_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_7_end_mask_0 = const()[name = tensor("ki_7_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_7_squeeze_mask_0 = const()[name = tensor("ki_7_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_7")]; tensor var_774 = mul(x = freqs_7, y = ts_15)[name = tensor("op_774")]; tensor rotr_7 = cos(x = var_774)[name = tensor("rotr_7")]; tensor roti_7 = sin(x = var_774)[name = tensor("roti_7")]; tensor var_778 = mul(x = qr_7, y = rotr_7)[name = tensor("op_778")]; tensor var_779 = mul(x = qi_7, y = roti_7)[name = tensor("op_779")]; tensor var_780 = sub(x = var_778, y = var_779)[name = tensor("op_780")]; tensor var_781 = mul(x = qr_7, y = roti_7)[name = tensor("op_781")]; tensor var_782 = mul(x = qi_7, y = rotr_7)[name = tensor("op_782")]; tensor var_783 = add(x = var_781, y = var_782)[name = tensor("op_783")]; tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; tensor qo_7 = stack(axis = qo_7_axis_0, values = (var_780, var_783))[name = tensor("qo_7")]; tensor var_786 = mul(x = kr_7, y = rotr_7)[name = tensor("op_786")]; tensor var_787 = mul(x = ki_7, y = roti_7)[name = tensor("op_787")]; tensor var_788 = sub(x = var_786, y = var_787)[name = tensor("op_788")]; tensor var_789 = mul(x = kr_7, y = roti_7)[name = tensor("op_789")]; tensor var_790 = mul(x = ki_7, y = rotr_7)[name = tensor("op_790")]; tensor var_791 = add(x = var_789, y = var_790)[name = tensor("op_791")]; tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; tensor ko_7 = stack(axis = ko_7_axis_0, values = (var_788, var_791))[name = tensor("ko_7")]; tensor concat_39x = const()[name = tensor("concat_39x"), val = tensor([1, 4, -1, 64])]; tensor q_15 = reshape(shape = concat_39x, x = qo_7)[name = tensor("q_15")]; tensor concat_40x = const()[name = tensor("concat_40x"), val = tensor([1, 4, -1, 64])]; tensor k_15 = reshape(shape = concat_40x, x = ko_7)[name = tensor("k_15")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0)]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(1)]; tensor positions_7 = range_1d(end = gather_34, start = const_32, step = const_33)[name = tensor("positions_7")]; tensor concat_41x = const()[name = tensor("concat_41x"), val = tensor([1, 1, -1])]; tensor var_803 = reshape(shape = concat_41x, x = positions_7)[name = tensor("op_803")]; tensor var_803_promoted_dtype_0 = const()[name = tensor("op_803_promoted_dtype_0"), val = tensor("fp32")]; tensor var_803_promoted = cast(dtype = var_803_promoted_dtype_0, x = var_803)[name = tensor("cast_460")]; tensor valid_k_7 = less(x = var_803_promoted, y = var_394)[name = tensor("valid_k_7")]; tensor concat_42x = const()[name = tensor("concat_42x"), val = tensor([1, -1, 1])]; tensor var_808 = reshape(shape = concat_42x, x = positions_7)[name = tensor("op_808")]; tensor delta_7 = sub(x = var_808, y = var_803)[name = tensor("delta_7")]; tensor concat_44_axis_0 = const()[name = tensor("concat_44_axis_0"), val = tensor(0)]; tensor concat_44_interleave_0 = const()[name = tensor("concat_44_interleave_0"), val = tensor(false)]; tensor concat_44 = concat(axis = concat_44_axis_0, interleave = concat_44_interleave_0, values = (var_295, gather_34, gather_34))[name = tensor("concat_44")]; tensor fill_3_value_0 = const()[name = tensor("fill_3_value_0"), val = tensor(0x1p+0)]; tensor fill_3 = fill(shape = concat_44, value = fill_3_value_0)[name = tensor("fill_3")]; tensor attn_bias_25_dtype_0 = const()[name = tensor("attn_bias_25_dtype_0"), val = tensor("bool")]; tensor var_814 = greater_equal(x = delta_7, y = var_277)[name = tensor("op_814")]; tensor attn_bias_25 = cast(dtype = attn_bias_25_dtype_0, x = fill_3)[name = tensor("cast_459")]; tensor attn_bias_27 = logical_and(x = attn_bias_25, y = var_814)[name = tensor("attn_bias_27")]; tensor var_816 = less(x = delta_7, y = var_284)[name = tensor("op_816")]; tensor attn_bias_29 = logical_and(x = attn_bias_27, y = var_816)[name = tensor("attn_bias_29")]; tensor var_818 = logical_and(x = attn_bias_29, y = valid_k_7)[name = tensor("op_818")]; tensor var_820_axes_0 = const()[name = tensor("op_820_axes_0"), val = tensor([1])]; tensor var_820 = expand_dims(axes = var_820_axes_0, x = var_818)[name = tensor("op_820")]; tensor cast_86_dtype_0 = const()[name = tensor("cast_86_dtype_0"), val = tensor("fp32")]; tensor sub_3_x_0 = const()[name = tensor("sub_3_x_0"), val = tensor(0x1p+0)]; tensor cast_86 = cast(dtype = cast_86_dtype_0, x = var_820)[name = tensor("cast_458")]; tensor sub_3 = sub(x = sub_3_x_0, y = cast_86)[name = tensor("sub_3")]; tensor mul_6_x_0 = const()[name = tensor("mul_6_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_6 = mul(x = mul_6_x_0, y = sub_3)[name = tensor("mul_6")]; tensor mul_7_y_0 = const()[name = tensor("mul_7_y_0"), val = tensor(0x1p-3)]; tensor mul_7 = mul(x = q_15, y = mul_7_y_0)[name = tensor("mul_7")]; tensor matmul_3_transpose_y_0 = const()[name = tensor("matmul_3_transpose_y_0"), val = tensor(true)]; tensor matmul_3_transpose_x_0 = const()[name = tensor("matmul_3_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_3")]; tensor add_3 = add(x = matmul_3, y = mul_6)[name = tensor("add_3")]; tensor softmax_3_axis_0 = const()[name = tensor("softmax_3_axis_0"), val = tensor(-1)]; tensor softmax_3 = softmax(axis = softmax_3_axis_0, x = add_3)[name = tensor("softmax_3")]; tensor out_13_transpose_x_0 = const()[name = tensor("out_13_transpose_x_0"), val = tensor(false)]; tensor out_13_transpose_y_0 = const()[name = tensor("out_13_transpose_y_0"), val = tensor(false)]; tensor 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("out_13")]; tensor concat_45x = const()[name = tensor("concat_45x"), val = tensor([1, -1])]; tensor var_826 = reshape(shape = concat_45x, x = positions_7)[name = tensor("op_826")]; tensor var_826_promoted_dtype_0 = const()[name = tensor("op_826_promoted_dtype_0"), val = tensor("fp32")]; tensor var_826_promoted = cast(dtype = var_826_promoted_dtype_0, x = var_826)[name = tensor("cast_457")]; tensor var_829 = less(x = var_826_promoted, y = var_417)[name = tensor("op_829")]; tensor concat_46x = const()[name = tensor("concat_46x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_7 = reshape(shape = concat_46x, x = var_829)[name = tensor("valid_q_7")]; tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x0p+0)]; tensor out_15 = select(a = out_13, b = var_833, cond = valid_q_7)[name = tensor("out_15")]; tensor var_835_perm_0 = const()[name = tensor("op_835_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_47x = const()[name = tensor("concat_47x"), val = tensor([1, -1, 256])]; tensor var_835 = transpose(perm = var_835_perm_0, x = out_15)[name = tensor("transpose_74")]; tensor input_39 = reshape(shape = concat_47x, x = var_835)[name = tensor("input_39")]; tensor x_27 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_3_self_attn_out_proj_weight, x = input_39)[name = tensor("linear_14")]; tensor var_841 = mul(x = encoder_1_transformer_layers_3_layer_scale_1_scale, y = x_27)[name = tensor("op_841")]; tensor input_41 = add(x = input_37, y = var_841)[name = tensor("input_41")]; tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; tensor 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("input_43")]; tensor input_45 = linear(bias = linear_3_bias_0, weight = encoder_1_transformer_layers_3_ffn_0_weight, x = input_43)[name = tensor("linear_15")]; tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; tensor input_47 = gelu(mode = input_47_mode_0, x = input_45)[name = tensor("input_47")]; tensor x_29 = linear(bias = linear_0_bias_0, weight = encoder_1_transformer_layers_3_ffn_2_weight, x = input_47)[name = tensor("linear_16")]; tensor var_856 = mul(x = encoder_1_transformer_layers_3_layer_scale_2_scale, y = x_29)[name = tensor("op_856")]; tensor input_49 = add(x = input_41, y = var_856)[name = tensor("input_49")]; tensor linear_17_bias_0 = const()[name = tensor("linear_17_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44430272)))]; tensor var_859 = linear(bias = linear_17_bias_0, weight = encoder_1_output_proj_weight, x = input_49)[name = tensor("linear_17")]; tensor x_31_perm_0 = const()[name = tensor("x_31_perm_0"), val = tensor([0, 2, 1])]; tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 384, -1, 2])]; tensor x_31 = transpose(perm = x_31_perm_0, x = var_859)[name = tensor("transpose_73")]; tensor var_871 = reshape(shape = var_870, x = x_31)[name = tensor("op_871")]; tensor var_872 = const()[name = tensor("op_872"), val = tensor([0, 1, 3, 2])]; tensor var_877 = const()[name = tensor("op_877"), val = tensor([1, 768, -1])]; tensor var_873 = transpose(perm = var_872, x = var_871)[name = tensor("transpose_72")]; tensor x_33 = reshape(shape = var_877, x = var_873)[name = tensor("x_33")]; tensor var_861_promoted = const()[name = tensor("op_861_promoted"), val = tensor(0x1p+1)]; tensor floor_div_1 = floor_div(x = input_lengths_1, y = var_861_promoted)[name = tensor("floor_div_1")]; tensor var_884 = const()[name = tensor("op_884"), val = tensor(0)]; tensor var_891 = const()[name = tensor("op_891"), val = tensor(1200)]; tensor var_898 = const()[name = tensor("op_898"), val = tensor(0x1.4f8b58p-17)]; tensor var_902 = const()[name = tensor("op_902"), val = tensor(1)]; tensor input_51_perm_0 = const()[name = tensor("input_51_perm_0"), val = tensor([0, 2, 1])]; tensor input_51 = transpose(perm = input_51_perm_0, x = x_33)[name = tensor("transpose_71")]; tensor x_35 = linear(bias = linear_0_bias_0, weight = encoder_3_input_proj_weight, x = input_51)[name = tensor("linear_18")]; tensor x_37_axes_0 = const()[name = tensor("x_37_axes_0"), val = tensor([-1])]; tensor 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("x_37")]; tensor var_926_shape = shape(x = x_37)[name = tensor("op_926_shape")]; tensor gather_46_batch_dims_0 = const()[name = tensor("gather_46_batch_dims_0"), val = tensor(0)]; tensor gather_46_validate_indices_0 = const()[name = tensor("gather_46_validate_indices_0"), val = tensor(false)]; tensor select_24 = const()[name = tensor("select_24"), val = tensor(1)]; tensor gather_46_axis_1 = const()[name = tensor("gather_46_axis_1"), val = tensor(0)]; tensor 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("gather_46")]; tensor projected_17 = linear(bias = linear_1_bias_0, weight = encoder_3_transformer_layers_0_self_attn_in_proj_weight, x = x_37)[name = tensor("linear_19")]; tensor concat_48x = const()[name = tensor("concat_48x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_933 = reshape(shape = concat_48x, x = projected_17)[name = tensor("op_933")]; tensor var_934 = const()[name = tensor("op_934"), val = tensor([2, 0, 3, 1, 4])]; tensor q_17_begin_0 = const()[name = tensor("q_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_17_end_0 = const()[name = tensor("q_17_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_17_end_mask_0 = const()[name = tensor("q_17_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_17_squeeze_mask_0 = const()[name = tensor("q_17_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_19 = transpose(perm = var_934, x = var_933)[name = tensor("transpose_70")]; tensor 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("q_17")]; tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_17")]; tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_9")]; tensor freqs_9 = const()[name = tensor("freqs_9"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44431872)))]; tensor var_946_shape = shape(x = q_17)[name = tensor("op_946_shape")]; tensor gather_50_batch_dims_0 = const()[name = tensor("gather_50_batch_dims_0"), val = tensor(0)]; tensor gather_50_validate_indices_0 = const()[name = tensor("gather_50_validate_indices_0"), val = tensor(false)]; tensor select_25 = const()[name = tensor("select_25"), val = tensor(2)]; tensor gather_50_axis_1 = const()[name = tensor("gather_50_axis_1"), val = tensor(0)]; tensor 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("gather_50")]; tensor var_949 = const()[name = tensor("op_949"), val = tensor([[0x0p+0]])]; tensor const_39 = const()[name = tensor("const_39"), val = tensor(0)]; tensor const_40 = const()[name = tensor("const_40"), val = tensor(1)]; tensor var_950 = range_1d(end = gather_50, start = const_39, step = const_40)[name = tensor("op_950")]; tensor var_950_promoted_dtype_0 = const()[name = tensor("op_950_promoted_dtype_0"), val = tensor("fp32")]; tensor var_950_promoted = cast(dtype = var_950_promoted_dtype_0, x = var_950)[name = tensor("cast_456")]; tensor ts_17 = add(x = var_949, y = var_950_promoted)[name = tensor("ts_17")]; tensor var_953 = const()[name = tensor("op_953"), val = tensor([1, 1, -1, 1])]; tensor ts_19 = reshape(shape = var_953, x = ts_17)[name = tensor("ts_19")]; tensor concat_49x = const()[name = tensor("concat_49x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_9 = reshape(shape = concat_49x, x = q_17)[name = tensor("q2_9")]; tensor concat_50x = const()[name = tensor("concat_50x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_9 = reshape(shape = concat_50x, x = k_17)[name = tensor("k2_9")]; tensor qr_9_begin_0 = const()[name = tensor("qr_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_9_end_0 = const()[name = tensor("qr_9_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_9_end_mask_0 = const()[name = tensor("qr_9_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_9_squeeze_mask_0 = const()[name = tensor("qr_9_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_9")]; tensor qi_9_begin_0 = const()[name = tensor("qi_9_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_9_end_0 = const()[name = tensor("qi_9_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_9_end_mask_0 = const()[name = tensor("qi_9_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_9_squeeze_mask_0 = const()[name = tensor("qi_9_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_9")]; tensor kr_9_begin_0 = const()[name = tensor("kr_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_9_end_0 = const()[name = tensor("kr_9_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_9_end_mask_0 = const()[name = tensor("kr_9_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_9_squeeze_mask_0 = const()[name = tensor("kr_9_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_9")]; tensor ki_9_begin_0 = const()[name = tensor("ki_9_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_9_end_0 = const()[name = tensor("ki_9_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_9_end_mask_0 = const()[name = tensor("ki_9_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_9_squeeze_mask_0 = const()[name = tensor("ki_9_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_9")]; tensor var_966 = mul(x = freqs_9, y = ts_19)[name = tensor("op_966")]; tensor rotr_9 = cos(x = var_966)[name = tensor("rotr_9")]; tensor roti_9 = sin(x = var_966)[name = tensor("roti_9")]; tensor var_970 = mul(x = qr_9, y = rotr_9)[name = tensor("op_970")]; tensor var_971 = mul(x = qi_9, y = roti_9)[name = tensor("op_971")]; tensor var_972 = sub(x = var_970, y = var_971)[name = tensor("op_972")]; tensor var_973 = mul(x = qr_9, y = roti_9)[name = tensor("op_973")]; tensor var_974 = mul(x = qi_9, y = rotr_9)[name = tensor("op_974")]; tensor var_975 = add(x = var_973, y = var_974)[name = tensor("op_975")]; tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; tensor qo_9 = stack(axis = qo_9_axis_0, values = (var_972, var_975))[name = tensor("qo_9")]; tensor var_978 = mul(x = kr_9, y = rotr_9)[name = tensor("op_978")]; tensor var_979 = mul(x = ki_9, y = roti_9)[name = tensor("op_979")]; tensor var_980 = sub(x = var_978, y = var_979)[name = tensor("op_980")]; tensor var_981 = mul(x = kr_9, y = roti_9)[name = tensor("op_981")]; tensor var_982 = mul(x = ki_9, y = rotr_9)[name = tensor("op_982")]; tensor var_983 = add(x = var_981, y = var_982)[name = tensor("op_983")]; tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; tensor ko_9 = stack(axis = ko_9_axis_0, values = (var_980, var_983))[name = tensor("ko_9")]; tensor concat_51x = const()[name = tensor("concat_51x"), val = tensor([1, 4, -1, 64])]; tensor q_19 = reshape(shape = concat_51x, x = qo_9)[name = tensor("q_19")]; tensor concat_52x = const()[name = tensor("concat_52x"), val = tensor([1, 4, -1, 64])]; tensor k_19 = reshape(shape = concat_52x, x = ko_9)[name = tensor("k_19")]; tensor const_41 = const()[name = tensor("const_41"), val = tensor(0)]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(1)]; tensor positions_9 = range_1d(end = gather_46, start = const_41, step = const_42)[name = tensor("positions_9")]; tensor concat_53x = const()[name = tensor("concat_53x"), val = tensor([1, 1, -1])]; tensor var_995 = reshape(shape = concat_53x, x = positions_9)[name = tensor("op_995")]; tensor var_996 = const()[name = tensor("op_996"), val = tensor([-1, 1, 1])]; tensor var_997 = reshape(shape = var_996, x = floor_div_1)[name = tensor("op_997")]; tensor var_995_promoted_dtype_0 = const()[name = tensor("op_995_promoted_dtype_0"), val = tensor("fp32")]; tensor var_995_promoted = cast(dtype = var_995_promoted_dtype_0, x = var_995)[name = tensor("cast_455")]; tensor valid_k_9 = less(x = var_995_promoted, y = var_997)[name = tensor("valid_k_9")]; tensor concat_54x = const()[name = tensor("concat_54x"), val = tensor([1, -1, 1])]; tensor var_1000 = reshape(shape = concat_54x, x = positions_9)[name = tensor("op_1000")]; tensor delta_9 = sub(x = var_1000, y = var_995)[name = tensor("delta_9")]; tensor concat_56_axis_0 = const()[name = tensor("concat_56_axis_0"), val = tensor(0)]; tensor concat_56_interleave_0 = const()[name = tensor("concat_56_interleave_0"), val = tensor(false)]; tensor concat_56 = concat(axis = concat_56_axis_0, interleave = concat_56_interleave_0, values = (var_902, gather_46, gather_46))[name = tensor("concat_56")]; tensor fill_4_value_0 = const()[name = tensor("fill_4_value_0"), val = tensor(0x1p+0)]; tensor fill_4 = fill(shape = concat_56, value = fill_4_value_0)[name = tensor("fill_4")]; tensor attn_bias_33_dtype_0 = const()[name = tensor("attn_bias_33_dtype_0"), val = tensor("bool")]; tensor var_1006 = greater_equal(x = delta_9, y = var_884)[name = tensor("op_1006")]; tensor attn_bias_33 = cast(dtype = attn_bias_33_dtype_0, x = fill_4)[name = tensor("cast_454")]; tensor attn_bias_35 = logical_and(x = attn_bias_33, y = var_1006)[name = tensor("attn_bias_35")]; tensor var_1008 = less(x = delta_9, y = var_891)[name = tensor("op_1008")]; tensor attn_bias_37 = logical_and(x = attn_bias_35, y = var_1008)[name = tensor("attn_bias_37")]; tensor var_1010 = logical_and(x = attn_bias_37, y = valid_k_9)[name = tensor("op_1010")]; tensor var_1012_axes_0 = const()[name = tensor("op_1012_axes_0"), val = tensor([1])]; tensor var_1012 = expand_dims(axes = var_1012_axes_0, x = var_1010)[name = tensor("op_1012")]; tensor cast_110_dtype_0 = const()[name = tensor("cast_110_dtype_0"), val = tensor("fp32")]; tensor sub_4_x_0 = const()[name = tensor("sub_4_x_0"), val = tensor(0x1p+0)]; tensor cast_110 = cast(dtype = cast_110_dtype_0, x = var_1012)[name = tensor("cast_453")]; tensor sub_4 = sub(x = sub_4_x_0, y = cast_110)[name = tensor("sub_4")]; tensor mul_8_x_0 = const()[name = tensor("mul_8_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_8 = mul(x = mul_8_x_0, y = sub_4)[name = tensor("mul_8")]; tensor mul_9_y_0 = const()[name = tensor("mul_9_y_0"), val = tensor(0x1p-3)]; tensor mul_9 = mul(x = q_19, y = mul_9_y_0)[name = tensor("mul_9")]; tensor matmul_4_transpose_y_0 = const()[name = tensor("matmul_4_transpose_y_0"), val = tensor(true)]; tensor matmul_4_transpose_x_0 = const()[name = tensor("matmul_4_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_4")]; tensor add_4 = add(x = matmul_4, y = mul_8)[name = tensor("add_4")]; tensor softmax_4_axis_0 = const()[name = tensor("softmax_4_axis_0"), val = tensor(-1)]; tensor softmax_4 = softmax(axis = softmax_4_axis_0, x = add_4)[name = tensor("softmax_4")]; tensor out_17_transpose_x_0 = const()[name = tensor("out_17_transpose_x_0"), val = tensor(false)]; tensor out_17_transpose_y_0 = const()[name = tensor("out_17_transpose_y_0"), val = tensor(false)]; tensor 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("out_17")]; tensor concat_57x = const()[name = tensor("concat_57x"), val = tensor([1, -1])]; tensor var_1018 = reshape(shape = concat_57x, x = positions_9)[name = tensor("op_1018")]; tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([-1, 1])]; tensor var_1020 = reshape(shape = var_1019, x = floor_div_1)[name = tensor("op_1020")]; tensor var_1018_promoted_dtype_0 = const()[name = tensor("op_1018_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1018_promoted = cast(dtype = var_1018_promoted_dtype_0, x = var_1018)[name = tensor("cast_452")]; tensor var_1021 = less(x = var_1018_promoted, y = var_1020)[name = tensor("op_1021")]; tensor concat_58x = const()[name = tensor("concat_58x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_9 = reshape(shape = concat_58x, x = var_1021)[name = tensor("valid_q_9")]; tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x0p+0)]; tensor out_19 = select(a = out_17, b = var_1025, cond = valid_q_9)[name = tensor("out_19")]; tensor var_1027_perm_0 = const()[name = tensor("op_1027_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_59x = const()[name = tensor("concat_59x"), val = tensor([1, -1, 256])]; tensor var_1027 = transpose(perm = var_1027_perm_0, x = out_19)[name = tensor("transpose_69")]; tensor input_53 = reshape(shape = concat_59x, x = var_1027)[name = tensor("input_53")]; tensor x_39 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_0_self_attn_out_proj_weight, x = input_53)[name = tensor("linear_20")]; tensor var_1033 = mul(x = encoder_3_transformer_layers_0_layer_scale_1_scale, y = x_39)[name = tensor("op_1033")]; tensor input_55 = add(x = x_35, y = var_1033)[name = tensor("input_55")]; tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; tensor 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("input_57")]; tensor input_59 = linear(bias = linear_3_bias_0, weight = encoder_3_transformer_layers_0_ffn_0_weight, x = input_57)[name = tensor("linear_21")]; tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; tensor input_61 = gelu(mode = input_61_mode_0, x = input_59)[name = tensor("input_61")]; tensor x_41 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_0_ffn_2_weight, x = input_61)[name = tensor("linear_22")]; tensor var_1048 = mul(x = encoder_3_transformer_layers_0_layer_scale_2_scale, y = x_41)[name = tensor("op_1048")]; tensor input_63 = add(x = input_55, y = var_1048)[name = tensor("input_63")]; tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; tensor 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("x_43")]; tensor var_1063_shape = shape(x = x_43)[name = tensor("op_1063_shape")]; tensor gather_56_batch_dims_0 = const()[name = tensor("gather_56_batch_dims_0"), val = tensor(0)]; tensor gather_56_validate_indices_0 = const()[name = tensor("gather_56_validate_indices_0"), val = tensor(false)]; tensor select_26 = const()[name = tensor("select_26"), val = tensor(1)]; tensor gather_56_axis_1 = const()[name = tensor("gather_56_axis_1"), val = tensor(0)]; tensor 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("gather_56")]; tensor projected_21 = linear(bias = linear_1_bias_0, weight = encoder_3_transformer_layers_1_self_attn_in_proj_weight, x = x_43)[name = tensor("linear_23")]; tensor concat_60x = const()[name = tensor("concat_60x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_1070 = reshape(shape = concat_60x, x = projected_21)[name = tensor("op_1070")]; tensor var_1071 = const()[name = tensor("op_1071"), val = tensor([2, 0, 3, 1, 4])]; tensor q_21_begin_0 = const()[name = tensor("q_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_21_end_0 = const()[name = tensor("q_21_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_21_end_mask_0 = const()[name = tensor("q_21_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_21_squeeze_mask_0 = const()[name = tensor("q_21_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_23 = transpose(perm = var_1071, x = var_1070)[name = tensor("transpose_68")]; tensor 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("q_21")]; tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_21")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_11")]; tensor freqs_11 = const()[name = tensor("freqs_11"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44432064)))]; tensor var_1083_shape = shape(x = q_21)[name = tensor("op_1083_shape")]; tensor gather_60_batch_dims_0 = const()[name = tensor("gather_60_batch_dims_0"), val = tensor(0)]; tensor gather_60_validate_indices_0 = const()[name = tensor("gather_60_validate_indices_0"), val = tensor(false)]; tensor select_27 = const()[name = tensor("select_27"), val = tensor(2)]; tensor gather_60_axis_1 = const()[name = tensor("gather_60_axis_1"), val = tensor(0)]; tensor 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("gather_60")]; tensor var_1086 = const()[name = tensor("op_1086"), val = tensor([[0x0p+0]])]; tensor const_48 = const()[name = tensor("const_48"), val = tensor(0)]; tensor const_49 = const()[name = tensor("const_49"), val = tensor(1)]; tensor var_1087 = range_1d(end = gather_60, start = const_48, step = const_49)[name = tensor("op_1087")]; tensor var_1087_promoted_dtype_0 = const()[name = tensor("op_1087_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1087_promoted = cast(dtype = var_1087_promoted_dtype_0, x = var_1087)[name = tensor("cast_451")]; tensor ts_21 = add(x = var_1086, y = var_1087_promoted)[name = tensor("ts_21")]; tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 1, -1, 1])]; tensor ts_23 = reshape(shape = var_1090, x = ts_21)[name = tensor("ts_23")]; tensor concat_61x = const()[name = tensor("concat_61x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_11 = reshape(shape = concat_61x, x = q_21)[name = tensor("q2_11")]; tensor concat_62x = const()[name = tensor("concat_62x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_11 = reshape(shape = concat_62x, x = k_21)[name = tensor("k2_11")]; tensor qr_11_begin_0 = const()[name = tensor("qr_11_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_11_end_0 = const()[name = tensor("qr_11_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_11_end_mask_0 = const()[name = tensor("qr_11_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_11_squeeze_mask_0 = const()[name = tensor("qr_11_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_11")]; tensor qi_11_begin_0 = const()[name = tensor("qi_11_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_11_end_0 = const()[name = tensor("qi_11_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_11_end_mask_0 = const()[name = tensor("qi_11_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_11_squeeze_mask_0 = const()[name = tensor("qi_11_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_11")]; tensor kr_11_begin_0 = const()[name = tensor("kr_11_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_11_end_0 = const()[name = tensor("kr_11_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_11_end_mask_0 = const()[name = tensor("kr_11_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_11_squeeze_mask_0 = const()[name = tensor("kr_11_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_11")]; tensor ki_11_begin_0 = const()[name = tensor("ki_11_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_11_end_0 = const()[name = tensor("ki_11_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_11_end_mask_0 = const()[name = tensor("ki_11_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_11_squeeze_mask_0 = const()[name = tensor("ki_11_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_11")]; tensor var_1103 = mul(x = freqs_11, y = ts_23)[name = tensor("op_1103")]; tensor rotr_11 = cos(x = var_1103)[name = tensor("rotr_11")]; tensor roti_11 = sin(x = var_1103)[name = tensor("roti_11")]; tensor var_1107 = mul(x = qr_11, y = rotr_11)[name = tensor("op_1107")]; tensor var_1108 = mul(x = qi_11, y = roti_11)[name = tensor("op_1108")]; tensor var_1109 = sub(x = var_1107, y = var_1108)[name = tensor("op_1109")]; tensor var_1110 = mul(x = qr_11, y = roti_11)[name = tensor("op_1110")]; tensor var_1111 = mul(x = qi_11, y = rotr_11)[name = tensor("op_1111")]; tensor var_1112 = add(x = var_1110, y = var_1111)[name = tensor("op_1112")]; tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; tensor qo_11 = stack(axis = qo_11_axis_0, values = (var_1109, var_1112))[name = tensor("qo_11")]; tensor var_1115 = mul(x = kr_11, y = rotr_11)[name = tensor("op_1115")]; tensor var_1116 = mul(x = ki_11, y = roti_11)[name = tensor("op_1116")]; tensor var_1117 = sub(x = var_1115, y = var_1116)[name = tensor("op_1117")]; tensor var_1118 = mul(x = kr_11, y = roti_11)[name = tensor("op_1118")]; tensor var_1119 = mul(x = ki_11, y = rotr_11)[name = tensor("op_1119")]; tensor var_1120 = add(x = var_1118, y = var_1119)[name = tensor("op_1120")]; tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; tensor ko_11 = stack(axis = ko_11_axis_0, values = (var_1117, var_1120))[name = tensor("ko_11")]; tensor concat_63x = const()[name = tensor("concat_63x"), val = tensor([1, 4, -1, 64])]; tensor q_23 = reshape(shape = concat_63x, x = qo_11)[name = tensor("q_23")]; tensor concat_64x = const()[name = tensor("concat_64x"), val = tensor([1, 4, -1, 64])]; tensor k_23 = reshape(shape = concat_64x, x = ko_11)[name = tensor("k_23")]; tensor const_50 = const()[name = tensor("const_50"), val = tensor(0)]; tensor const_51 = const()[name = tensor("const_51"), val = tensor(1)]; tensor positions_11 = range_1d(end = gather_56, start = const_50, step = const_51)[name = tensor("positions_11")]; tensor concat_65x = const()[name = tensor("concat_65x"), val = tensor([1, 1, -1])]; tensor var_1132 = reshape(shape = concat_65x, x = positions_11)[name = tensor("op_1132")]; tensor var_1132_promoted_dtype_0 = const()[name = tensor("op_1132_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1132_promoted = cast(dtype = var_1132_promoted_dtype_0, x = var_1132)[name = tensor("cast_450")]; tensor valid_k_11 = less(x = var_1132_promoted, y = var_997)[name = tensor("valid_k_11")]; tensor concat_66x = const()[name = tensor("concat_66x"), val = tensor([1, -1, 1])]; tensor var_1137 = reshape(shape = concat_66x, x = positions_11)[name = tensor("op_1137")]; tensor delta_11 = sub(x = var_1137, y = var_1132)[name = tensor("delta_11")]; tensor concat_68_axis_0 = const()[name = tensor("concat_68_axis_0"), val = tensor(0)]; tensor concat_68_interleave_0 = const()[name = tensor("concat_68_interleave_0"), val = tensor(false)]; tensor concat_68 = concat(axis = concat_68_axis_0, interleave = concat_68_interleave_0, values = (var_902, gather_56, gather_56))[name = tensor("concat_68")]; tensor fill_5_value_0 = const()[name = tensor("fill_5_value_0"), val = tensor(0x1p+0)]; tensor fill_5 = fill(shape = concat_68, value = fill_5_value_0)[name = tensor("fill_5")]; tensor attn_bias_41_dtype_0 = const()[name = tensor("attn_bias_41_dtype_0"), val = tensor("bool")]; tensor var_1143 = greater_equal(x = delta_11, y = var_884)[name = tensor("op_1143")]; tensor attn_bias_41 = cast(dtype = attn_bias_41_dtype_0, x = fill_5)[name = tensor("cast_449")]; tensor attn_bias_43 = logical_and(x = attn_bias_41, y = var_1143)[name = tensor("attn_bias_43")]; tensor var_1145 = less(x = delta_11, y = var_891)[name = tensor("op_1145")]; tensor attn_bias_45 = logical_and(x = attn_bias_43, y = var_1145)[name = tensor("attn_bias_45")]; tensor var_1147 = logical_and(x = attn_bias_45, y = valid_k_11)[name = tensor("op_1147")]; tensor var_1149_axes_0 = const()[name = tensor("op_1149_axes_0"), val = tensor([1])]; tensor var_1149 = expand_dims(axes = var_1149_axes_0, x = var_1147)[name = tensor("op_1149")]; tensor cast_132_dtype_0 = const()[name = tensor("cast_132_dtype_0"), val = tensor("fp32")]; tensor sub_5_x_0 = const()[name = tensor("sub_5_x_0"), val = tensor(0x1p+0)]; tensor cast_132 = cast(dtype = cast_132_dtype_0, x = var_1149)[name = tensor("cast_448")]; tensor sub_5 = sub(x = sub_5_x_0, y = cast_132)[name = tensor("sub_5")]; tensor mul_10_x_0 = const()[name = tensor("mul_10_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_10 = mul(x = mul_10_x_0, y = sub_5)[name = tensor("mul_10")]; tensor mul_11_y_0 = const()[name = tensor("mul_11_y_0"), val = tensor(0x1p-3)]; tensor mul_11 = mul(x = q_23, y = mul_11_y_0)[name = tensor("mul_11")]; tensor matmul_5_transpose_y_0 = const()[name = tensor("matmul_5_transpose_y_0"), val = tensor(true)]; tensor matmul_5_transpose_x_0 = const()[name = tensor("matmul_5_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_5")]; tensor add_5 = add(x = matmul_5, y = mul_10)[name = tensor("add_5")]; tensor softmax_5_axis_0 = const()[name = tensor("softmax_5_axis_0"), val = tensor(-1)]; tensor softmax_5 = softmax(axis = softmax_5_axis_0, x = add_5)[name = tensor("softmax_5")]; tensor out_21_transpose_x_0 = const()[name = tensor("out_21_transpose_x_0"), val = tensor(false)]; tensor out_21_transpose_y_0 = const()[name = tensor("out_21_transpose_y_0"), val = tensor(false)]; tensor 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("out_21")]; tensor concat_69x = const()[name = tensor("concat_69x"), val = tensor([1, -1])]; tensor var_1155 = reshape(shape = concat_69x, x = positions_11)[name = tensor("op_1155")]; tensor var_1155_promoted_dtype_0 = const()[name = tensor("op_1155_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1155_promoted = cast(dtype = var_1155_promoted_dtype_0, x = var_1155)[name = tensor("cast_447")]; tensor var_1158 = less(x = var_1155_promoted, y = var_1020)[name = tensor("op_1158")]; tensor concat_70x = const()[name = tensor("concat_70x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_11 = reshape(shape = concat_70x, x = var_1158)[name = tensor("valid_q_11")]; tensor var_1162 = const()[name = tensor("op_1162"), val = tensor(0x0p+0)]; tensor out_23 = select(a = out_21, b = var_1162, cond = valid_q_11)[name = tensor("out_23")]; tensor var_1164_perm_0 = const()[name = tensor("op_1164_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_71x = const()[name = tensor("concat_71x"), val = tensor([1, -1, 256])]; tensor var_1164 = transpose(perm = var_1164_perm_0, x = out_23)[name = tensor("transpose_67")]; tensor input_65 = reshape(shape = concat_71x, x = var_1164)[name = tensor("input_65")]; tensor x_45 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_1_self_attn_out_proj_weight, x = input_65)[name = tensor("linear_24")]; tensor var_1170 = mul(x = encoder_3_transformer_layers_1_layer_scale_1_scale, y = x_45)[name = tensor("op_1170")]; tensor input_67 = add(x = input_63, y = var_1170)[name = tensor("input_67")]; tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; tensor 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("input_69")]; tensor input_71 = linear(bias = linear_3_bias_0, weight = encoder_3_transformer_layers_1_ffn_0_weight, x = input_69)[name = tensor("linear_25")]; tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; tensor input_73 = gelu(mode = input_73_mode_0, x = input_71)[name = tensor("input_73")]; tensor x_47 = linear(bias = linear_0_bias_0, weight = encoder_3_transformer_layers_1_ffn_2_weight, x = input_73)[name = tensor("linear_26")]; tensor var_1185 = mul(x = encoder_3_transformer_layers_1_layer_scale_2_scale, y = x_47)[name = tensor("op_1185")]; tensor input_75 = add(x = input_67, y = var_1185)[name = tensor("input_75")]; tensor var_1188 = linear(bias = linear_17_bias_0, weight = encoder_3_output_proj_weight, x = input_75)[name = tensor("linear_27")]; tensor x_49_perm_0 = const()[name = tensor("x_49_perm_0"), val = tensor([0, 2, 1])]; tensor var_1199 = const()[name = tensor("op_1199"), val = tensor([1, 384, -1, 2])]; tensor x_49 = transpose(perm = x_49_perm_0, x = var_1188)[name = tensor("transpose_66")]; tensor var_1200 = reshape(shape = var_1199, x = x_49)[name = tensor("op_1200")]; tensor var_1201 = const()[name = tensor("op_1201"), val = tensor([0, 1, 3, 2])]; tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([1, 768, -1])]; tensor var_1202 = transpose(perm = var_1201, x = var_1200)[name = tensor("transpose_65")]; tensor x_51 = reshape(shape = var_1206, x = var_1202)[name = tensor("x_51")]; tensor var_1190_promoted = const()[name = tensor("op_1190_promoted"), val = tensor(0x1p+1)]; tensor floor_div_2 = floor_div(x = floor_div_1, y = var_1190_promoted)[name = tensor("floor_div_2")]; tensor var_1213 = const()[name = tensor("op_1213"), val = tensor(0)]; tensor var_1220 = const()[name = tensor("op_1220"), val = tensor(800)]; tensor var_1227 = const()[name = tensor("op_1227"), val = tensor(0x1.4f8b58p-17)]; tensor var_1231 = const()[name = tensor("op_1231"), val = tensor(1)]; tensor input_77_perm_0 = const()[name = tensor("input_77_perm_0"), val = tensor([0, 2, 1])]; tensor input_77 = transpose(perm = input_77_perm_0, x = x_51)[name = tensor("transpose_64")]; tensor x_53 = linear(bias = linear_0_bias_0, weight = encoder_5_input_proj_weight, x = input_77)[name = tensor("linear_28")]; tensor x_55_axes_0 = const()[name = tensor("x_55_axes_0"), val = tensor([-1])]; tensor 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("x_55")]; tensor var_1255_shape = shape(x = x_55)[name = tensor("op_1255_shape")]; tensor gather_68_batch_dims_0 = const()[name = tensor("gather_68_batch_dims_0"), val = tensor(0)]; tensor gather_68_validate_indices_0 = const()[name = tensor("gather_68_validate_indices_0"), val = tensor(false)]; tensor select_28 = const()[name = tensor("select_28"), val = tensor(1)]; tensor gather_68_axis_1 = const()[name = tensor("gather_68_axis_1"), val = tensor(0)]; tensor 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("gather_68")]; tensor projected_25 = linear(bias = linear_1_bias_0, weight = encoder_5_transformer_layers_0_self_attn_in_proj_weight, x = x_55)[name = tensor("linear_29")]; tensor concat_72x = const()[name = tensor("concat_72x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_1262 = reshape(shape = concat_72x, x = projected_25)[name = tensor("op_1262")]; tensor var_1263 = const()[name = tensor("op_1263"), val = tensor([2, 0, 3, 1, 4])]; tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_27 = transpose(perm = var_1263, x = var_1262)[name = tensor("transpose_63")]; tensor 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("q_25")]; tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_25")]; tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_13")]; tensor freqs_13 = const()[name = tensor("freqs_13"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44432256)))]; tensor var_1275_shape = shape(x = q_25)[name = tensor("op_1275_shape")]; tensor gather_72_batch_dims_0 = const()[name = tensor("gather_72_batch_dims_0"), val = tensor(0)]; tensor gather_72_validate_indices_0 = const()[name = tensor("gather_72_validate_indices_0"), val = tensor(false)]; tensor select_29 = const()[name = tensor("select_29"), val = tensor(2)]; tensor gather_72_axis_1 = const()[name = tensor("gather_72_axis_1"), val = tensor(0)]; tensor 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("gather_72")]; tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([[0x0p+0]])]; tensor const_57 = const()[name = tensor("const_57"), val = tensor(0)]; tensor const_58 = const()[name = tensor("const_58"), val = tensor(1)]; tensor var_1279 = range_1d(end = gather_72, start = const_57, step = const_58)[name = tensor("op_1279")]; tensor var_1279_promoted_dtype_0 = const()[name = tensor("op_1279_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1279_promoted = cast(dtype = var_1279_promoted_dtype_0, x = var_1279)[name = tensor("cast_446")]; tensor ts_25 = add(x = var_1278, y = var_1279_promoted)[name = tensor("ts_25")]; tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([1, 1, -1, 1])]; tensor ts_27 = reshape(shape = var_1282, x = ts_25)[name = tensor("ts_27")]; tensor concat_73x = const()[name = tensor("concat_73x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_13 = reshape(shape = concat_73x, x = q_25)[name = tensor("q2_13")]; tensor concat_74x = const()[name = tensor("concat_74x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_13 = reshape(shape = concat_74x, x = k_25)[name = tensor("k2_13")]; tensor qr_13_begin_0 = const()[name = tensor("qr_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_13_end_0 = const()[name = tensor("qr_13_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_13_end_mask_0 = const()[name = tensor("qr_13_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_13_squeeze_mask_0 = const()[name = tensor("qr_13_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_13")]; tensor qi_13_begin_0 = const()[name = tensor("qi_13_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_13_end_0 = const()[name = tensor("qi_13_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_13_end_mask_0 = const()[name = tensor("qi_13_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_13_squeeze_mask_0 = const()[name = tensor("qi_13_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_13")]; tensor kr_13_begin_0 = const()[name = tensor("kr_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_13_end_0 = const()[name = tensor("kr_13_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_13_end_mask_0 = const()[name = tensor("kr_13_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_13_squeeze_mask_0 = const()[name = tensor("kr_13_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_13")]; tensor ki_13_begin_0 = const()[name = tensor("ki_13_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_13_end_0 = const()[name = tensor("ki_13_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_13_end_mask_0 = const()[name = tensor("ki_13_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_13_squeeze_mask_0 = const()[name = tensor("ki_13_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_13")]; tensor var_1295 = mul(x = freqs_13, y = ts_27)[name = tensor("op_1295")]; tensor rotr_13 = cos(x = var_1295)[name = tensor("rotr_13")]; tensor roti_13 = sin(x = var_1295)[name = tensor("roti_13")]; tensor var_1299 = mul(x = qr_13, y = rotr_13)[name = tensor("op_1299")]; tensor var_1300 = mul(x = qi_13, y = roti_13)[name = tensor("op_1300")]; tensor var_1301 = sub(x = var_1299, y = var_1300)[name = tensor("op_1301")]; tensor var_1302 = mul(x = qr_13, y = roti_13)[name = tensor("op_1302")]; tensor var_1303 = mul(x = qi_13, y = rotr_13)[name = tensor("op_1303")]; tensor var_1304 = add(x = var_1302, y = var_1303)[name = tensor("op_1304")]; tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; tensor qo_13 = stack(axis = qo_13_axis_0, values = (var_1301, var_1304))[name = tensor("qo_13")]; tensor var_1307 = mul(x = kr_13, y = rotr_13)[name = tensor("op_1307")]; tensor var_1308 = mul(x = ki_13, y = roti_13)[name = tensor("op_1308")]; tensor var_1309 = sub(x = var_1307, y = var_1308)[name = tensor("op_1309")]; tensor var_1310 = mul(x = kr_13, y = roti_13)[name = tensor("op_1310")]; tensor var_1311 = mul(x = ki_13, y = rotr_13)[name = tensor("op_1311")]; tensor var_1312 = add(x = var_1310, y = var_1311)[name = tensor("op_1312")]; tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; tensor ko_13 = stack(axis = ko_13_axis_0, values = (var_1309, var_1312))[name = tensor("ko_13")]; tensor concat_75x = const()[name = tensor("concat_75x"), val = tensor([1, 4, -1, 64])]; tensor q_27 = reshape(shape = concat_75x, x = qo_13)[name = tensor("q_27")]; tensor concat_76x = const()[name = tensor("concat_76x"), val = tensor([1, 4, -1, 64])]; tensor k_27 = reshape(shape = concat_76x, x = ko_13)[name = tensor("k_27")]; tensor const_59 = const()[name = tensor("const_59"), val = tensor(0)]; tensor const_60 = const()[name = tensor("const_60"), val = tensor(1)]; tensor positions_13 = range_1d(end = gather_68, start = const_59, step = const_60)[name = tensor("positions_13")]; tensor concat_77x = const()[name = tensor("concat_77x"), val = tensor([1, 1, -1])]; tensor var_1324 = reshape(shape = concat_77x, x = positions_13)[name = tensor("op_1324")]; tensor var_1325 = const()[name = tensor("op_1325"), val = tensor([-1, 1, 1])]; tensor var_1326 = reshape(shape = var_1325, x = floor_div_2)[name = tensor("op_1326")]; tensor var_1324_promoted_dtype_0 = const()[name = tensor("op_1324_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1324_promoted = cast(dtype = var_1324_promoted_dtype_0, x = var_1324)[name = tensor("cast_445")]; tensor valid_k_13 = less(x = var_1324_promoted, y = var_1326)[name = tensor("valid_k_13")]; tensor concat_78x = const()[name = tensor("concat_78x"), val = tensor([1, -1, 1])]; tensor var_1329 = reshape(shape = concat_78x, x = positions_13)[name = tensor("op_1329")]; tensor delta_13 = sub(x = var_1329, y = var_1324)[name = tensor("delta_13")]; tensor concat_80_axis_0 = const()[name = tensor("concat_80_axis_0"), val = tensor(0)]; tensor concat_80_interleave_0 = const()[name = tensor("concat_80_interleave_0"), val = tensor(false)]; tensor concat_80 = concat(axis = concat_80_axis_0, interleave = concat_80_interleave_0, values = (var_1231, gather_68, gather_68))[name = tensor("concat_80")]; tensor fill_6_value_0 = const()[name = tensor("fill_6_value_0"), val = tensor(0x1p+0)]; tensor fill_6 = fill(shape = concat_80, value = fill_6_value_0)[name = tensor("fill_6")]; tensor attn_bias_49_dtype_0 = const()[name = tensor("attn_bias_49_dtype_0"), val = tensor("bool")]; tensor var_1335 = greater_equal(x = delta_13, y = var_1213)[name = tensor("op_1335")]; tensor attn_bias_49 = cast(dtype = attn_bias_49_dtype_0, x = fill_6)[name = tensor("cast_444")]; tensor attn_bias_51 = logical_and(x = attn_bias_49, y = var_1335)[name = tensor("attn_bias_51")]; tensor var_1337 = less(x = delta_13, y = var_1220)[name = tensor("op_1337")]; tensor attn_bias_53 = logical_and(x = attn_bias_51, y = var_1337)[name = tensor("attn_bias_53")]; tensor var_1339 = logical_and(x = attn_bias_53, y = valid_k_13)[name = tensor("op_1339")]; tensor var_1341_axes_0 = const()[name = tensor("op_1341_axes_0"), val = tensor([1])]; tensor var_1341 = expand_dims(axes = var_1341_axes_0, x = var_1339)[name = tensor("op_1341")]; tensor cast_156_dtype_0 = const()[name = tensor("cast_156_dtype_0"), val = tensor("fp32")]; tensor sub_6_x_0 = const()[name = tensor("sub_6_x_0"), val = tensor(0x1p+0)]; tensor cast_156 = cast(dtype = cast_156_dtype_0, x = var_1341)[name = tensor("cast_443")]; tensor sub_6 = sub(x = sub_6_x_0, y = cast_156)[name = tensor("sub_6")]; tensor mul_12_x_0 = const()[name = tensor("mul_12_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_12 = mul(x = mul_12_x_0, y = sub_6)[name = tensor("mul_12")]; tensor mul_13_y_0 = const()[name = tensor("mul_13_y_0"), val = tensor(0x1p-3)]; tensor mul_13 = mul(x = q_27, y = mul_13_y_0)[name = tensor("mul_13")]; tensor matmul_6_transpose_y_0 = const()[name = tensor("matmul_6_transpose_y_0"), val = tensor(true)]; tensor matmul_6_transpose_x_0 = const()[name = tensor("matmul_6_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_6")]; tensor add_6 = add(x = matmul_6, y = mul_12)[name = tensor("add_6")]; tensor softmax_6_axis_0 = const()[name = tensor("softmax_6_axis_0"), val = tensor(-1)]; tensor softmax_6 = softmax(axis = softmax_6_axis_0, x = add_6)[name = tensor("softmax_6")]; tensor out_25_transpose_x_0 = const()[name = tensor("out_25_transpose_x_0"), val = tensor(false)]; tensor out_25_transpose_y_0 = const()[name = tensor("out_25_transpose_y_0"), val = tensor(false)]; tensor 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("out_25")]; tensor concat_81x = const()[name = tensor("concat_81x"), val = tensor([1, -1])]; tensor var_1347 = reshape(shape = concat_81x, x = positions_13)[name = tensor("op_1347")]; tensor var_1348 = const()[name = tensor("op_1348"), val = tensor([-1, 1])]; tensor var_1349 = reshape(shape = var_1348, x = floor_div_2)[name = tensor("op_1349")]; tensor var_1347_promoted_dtype_0 = const()[name = tensor("op_1347_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1347_promoted = cast(dtype = var_1347_promoted_dtype_0, x = var_1347)[name = tensor("cast_442")]; tensor var_1350 = less(x = var_1347_promoted, y = var_1349)[name = tensor("op_1350")]; tensor concat_82x = const()[name = tensor("concat_82x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_13 = reshape(shape = concat_82x, x = var_1350)[name = tensor("valid_q_13")]; tensor var_1354 = const()[name = tensor("op_1354"), val = tensor(0x0p+0)]; tensor out_27 = select(a = out_25, b = var_1354, cond = valid_q_13)[name = tensor("out_27")]; tensor var_1356_perm_0 = const()[name = tensor("op_1356_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_83x = const()[name = tensor("concat_83x"), val = tensor([1, -1, 256])]; tensor var_1356 = transpose(perm = var_1356_perm_0, x = out_27)[name = tensor("transpose_62")]; tensor input_79 = reshape(shape = concat_83x, x = var_1356)[name = tensor("input_79")]; tensor x_57 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_0_self_attn_out_proj_weight, x = input_79)[name = tensor("linear_30")]; tensor var_1362 = mul(x = encoder_5_transformer_layers_0_layer_scale_1_scale, y = x_57)[name = tensor("op_1362")]; tensor input_81 = add(x = x_53, y = var_1362)[name = tensor("input_81")]; tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; tensor 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("input_83")]; tensor input_85 = linear(bias = linear_3_bias_0, weight = encoder_5_transformer_layers_0_ffn_0_weight, x = input_83)[name = tensor("linear_31")]; tensor input_87_mode_0 = const()[name = tensor("input_87_mode_0"), val = tensor("EXACT")]; tensor input_87 = gelu(mode = input_87_mode_0, x = input_85)[name = tensor("input_87")]; tensor x_59 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_0_ffn_2_weight, x = input_87)[name = tensor("linear_32")]; tensor var_1377 = mul(x = encoder_5_transformer_layers_0_layer_scale_2_scale, y = x_59)[name = tensor("op_1377")]; tensor input_89 = add(x = input_81, y = var_1377)[name = tensor("input_89")]; tensor x_61_axes_0 = const()[name = tensor("x_61_axes_0"), val = tensor([-1])]; tensor 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("x_61")]; tensor var_1392_shape = shape(x = x_61)[name = tensor("op_1392_shape")]; tensor gather_78_batch_dims_0 = const()[name = tensor("gather_78_batch_dims_0"), val = tensor(0)]; tensor gather_78_validate_indices_0 = const()[name = tensor("gather_78_validate_indices_0"), val = tensor(false)]; tensor select_30 = const()[name = tensor("select_30"), val = tensor(1)]; tensor gather_78_axis_1 = const()[name = tensor("gather_78_axis_1"), val = tensor(0)]; tensor 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("gather_78")]; tensor projected_29 = linear(bias = linear_1_bias_0, weight = encoder_5_transformer_layers_1_self_attn_in_proj_weight, x = x_61)[name = tensor("linear_33")]; tensor concat_84x = const()[name = tensor("concat_84x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_1399 = reshape(shape = concat_84x, x = projected_29)[name = tensor("op_1399")]; tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([2, 0, 3, 1, 4])]; tensor q_29_begin_0 = const()[name = tensor("q_29_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_29_end_0 = const()[name = tensor("q_29_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_29_end_mask_0 = const()[name = tensor("q_29_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_29_squeeze_mask_0 = const()[name = tensor("q_29_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_31 = transpose(perm = var_1400, x = var_1399)[name = tensor("transpose_61")]; tensor 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("q_29")]; tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_29")]; tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_15")]; tensor freqs_15 = const()[name = tensor("freqs_15"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44432448)))]; tensor var_1412_shape = shape(x = q_29)[name = tensor("op_1412_shape")]; tensor gather_82_batch_dims_0 = const()[name = tensor("gather_82_batch_dims_0"), val = tensor(0)]; tensor gather_82_validate_indices_0 = const()[name = tensor("gather_82_validate_indices_0"), val = tensor(false)]; tensor select_31 = const()[name = tensor("select_31"), val = tensor(2)]; tensor gather_82_axis_1 = const()[name = tensor("gather_82_axis_1"), val = tensor(0)]; tensor 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("gather_82")]; tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([[0x0p+0]])]; tensor const_66 = const()[name = tensor("const_66"), val = tensor(0)]; tensor const_67 = const()[name = tensor("const_67"), val = tensor(1)]; tensor var_1416 = range_1d(end = gather_82, start = const_66, step = const_67)[name = tensor("op_1416")]; tensor var_1416_promoted_dtype_0 = const()[name = tensor("op_1416_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1416_promoted = cast(dtype = var_1416_promoted_dtype_0, x = var_1416)[name = tensor("cast_441")]; tensor ts_29 = add(x = var_1415, y = var_1416_promoted)[name = tensor("ts_29")]; tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([1, 1, -1, 1])]; tensor ts_31 = reshape(shape = var_1419, x = ts_29)[name = tensor("ts_31")]; tensor concat_85x = const()[name = tensor("concat_85x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_15 = reshape(shape = concat_85x, x = q_29)[name = tensor("q2_15")]; tensor concat_86x = const()[name = tensor("concat_86x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_15 = reshape(shape = concat_86x, x = k_29)[name = tensor("k2_15")]; tensor qr_15_begin_0 = const()[name = tensor("qr_15_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_15_end_0 = const()[name = tensor("qr_15_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_15_end_mask_0 = const()[name = tensor("qr_15_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_15_squeeze_mask_0 = const()[name = tensor("qr_15_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_15")]; tensor qi_15_begin_0 = const()[name = tensor("qi_15_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_15_end_0 = const()[name = tensor("qi_15_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_15_end_mask_0 = const()[name = tensor("qi_15_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_15_squeeze_mask_0 = const()[name = tensor("qi_15_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_15")]; tensor kr_15_begin_0 = const()[name = tensor("kr_15_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_15_end_0 = const()[name = tensor("kr_15_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_15_end_mask_0 = const()[name = tensor("kr_15_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_15_squeeze_mask_0 = const()[name = tensor("kr_15_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_15")]; tensor ki_15_begin_0 = const()[name = tensor("ki_15_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_15_end_0 = const()[name = tensor("ki_15_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_15_end_mask_0 = const()[name = tensor("ki_15_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_15_squeeze_mask_0 = const()[name = tensor("ki_15_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_15")]; tensor var_1432 = mul(x = freqs_15, y = ts_31)[name = tensor("op_1432")]; tensor rotr_15 = cos(x = var_1432)[name = tensor("rotr_15")]; tensor roti_15 = sin(x = var_1432)[name = tensor("roti_15")]; tensor var_1436 = mul(x = qr_15, y = rotr_15)[name = tensor("op_1436")]; tensor var_1437 = mul(x = qi_15, y = roti_15)[name = tensor("op_1437")]; tensor var_1438 = sub(x = var_1436, y = var_1437)[name = tensor("op_1438")]; tensor var_1439 = mul(x = qr_15, y = roti_15)[name = tensor("op_1439")]; tensor var_1440 = mul(x = qi_15, y = rotr_15)[name = tensor("op_1440")]; tensor var_1441 = add(x = var_1439, y = var_1440)[name = tensor("op_1441")]; tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; tensor qo_15 = stack(axis = qo_15_axis_0, values = (var_1438, var_1441))[name = tensor("qo_15")]; tensor var_1444 = mul(x = kr_15, y = rotr_15)[name = tensor("op_1444")]; tensor var_1445 = mul(x = ki_15, y = roti_15)[name = tensor("op_1445")]; tensor var_1446 = sub(x = var_1444, y = var_1445)[name = tensor("op_1446")]; tensor var_1447 = mul(x = kr_15, y = roti_15)[name = tensor("op_1447")]; tensor var_1448 = mul(x = ki_15, y = rotr_15)[name = tensor("op_1448")]; tensor var_1449 = add(x = var_1447, y = var_1448)[name = tensor("op_1449")]; tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; tensor ko_15 = stack(axis = ko_15_axis_0, values = (var_1446, var_1449))[name = tensor("ko_15")]; tensor concat_87x = const()[name = tensor("concat_87x"), val = tensor([1, 4, -1, 64])]; tensor q_31 = reshape(shape = concat_87x, x = qo_15)[name = tensor("q_31")]; tensor concat_88x = const()[name = tensor("concat_88x"), val = tensor([1, 4, -1, 64])]; tensor k_31 = reshape(shape = concat_88x, x = ko_15)[name = tensor("k_31")]; tensor const_68 = const()[name = tensor("const_68"), val = tensor(0)]; tensor const_69 = const()[name = tensor("const_69"), val = tensor(1)]; tensor positions_15 = range_1d(end = gather_78, start = const_68, step = const_69)[name = tensor("positions_15")]; tensor concat_89x = const()[name = tensor("concat_89x"), val = tensor([1, 1, -1])]; tensor var_1461 = reshape(shape = concat_89x, x = positions_15)[name = tensor("op_1461")]; tensor var_1461_promoted_dtype_0 = const()[name = tensor("op_1461_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1461_promoted = cast(dtype = var_1461_promoted_dtype_0, x = var_1461)[name = tensor("cast_440")]; tensor valid_k_15 = less(x = var_1461_promoted, y = var_1326)[name = tensor("valid_k_15")]; tensor concat_90x = const()[name = tensor("concat_90x"), val = tensor([1, -1, 1])]; tensor var_1466 = reshape(shape = concat_90x, x = positions_15)[name = tensor("op_1466")]; tensor delta_15 = sub(x = var_1466, y = var_1461)[name = tensor("delta_15")]; tensor concat_92_axis_0 = const()[name = tensor("concat_92_axis_0"), val = tensor(0)]; tensor concat_92_interleave_0 = const()[name = tensor("concat_92_interleave_0"), val = tensor(false)]; tensor concat_92 = concat(axis = concat_92_axis_0, interleave = concat_92_interleave_0, values = (var_1231, gather_78, gather_78))[name = tensor("concat_92")]; tensor fill_7_value_0 = const()[name = tensor("fill_7_value_0"), val = tensor(0x1p+0)]; tensor fill_7 = fill(shape = concat_92, value = fill_7_value_0)[name = tensor("fill_7")]; tensor attn_bias_57_dtype_0 = const()[name = tensor("attn_bias_57_dtype_0"), val = tensor("bool")]; tensor var_1472 = greater_equal(x = delta_15, y = var_1213)[name = tensor("op_1472")]; tensor attn_bias_57 = cast(dtype = attn_bias_57_dtype_0, x = fill_7)[name = tensor("cast_439")]; tensor attn_bias_59 = logical_and(x = attn_bias_57, y = var_1472)[name = tensor("attn_bias_59")]; tensor var_1474 = less(x = delta_15, y = var_1220)[name = tensor("op_1474")]; tensor attn_bias_61 = logical_and(x = attn_bias_59, y = var_1474)[name = tensor("attn_bias_61")]; tensor var_1476 = logical_and(x = attn_bias_61, y = valid_k_15)[name = tensor("op_1476")]; tensor var_1478_axes_0 = const()[name = tensor("op_1478_axes_0"), val = tensor([1])]; tensor var_1478 = expand_dims(axes = var_1478_axes_0, x = var_1476)[name = tensor("op_1478")]; tensor cast_178_dtype_0 = const()[name = tensor("cast_178_dtype_0"), val = tensor("fp32")]; tensor sub_7_x_0 = const()[name = tensor("sub_7_x_0"), val = tensor(0x1p+0)]; tensor cast_178 = cast(dtype = cast_178_dtype_0, x = var_1478)[name = tensor("cast_438")]; tensor sub_7 = sub(x = sub_7_x_0, y = cast_178)[name = tensor("sub_7")]; tensor mul_14_x_0 = const()[name = tensor("mul_14_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_14 = mul(x = mul_14_x_0, y = sub_7)[name = tensor("mul_14")]; tensor mul_15_y_0 = const()[name = tensor("mul_15_y_0"), val = tensor(0x1p-3)]; tensor mul_15 = mul(x = q_31, y = mul_15_y_0)[name = tensor("mul_15")]; tensor matmul_7_transpose_y_0 = const()[name = tensor("matmul_7_transpose_y_0"), val = tensor(true)]; tensor matmul_7_transpose_x_0 = const()[name = tensor("matmul_7_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_7")]; tensor add_7 = add(x = matmul_7, y = mul_14)[name = tensor("add_7")]; tensor softmax_7_axis_0 = const()[name = tensor("softmax_7_axis_0"), val = tensor(-1)]; tensor softmax_7 = softmax(axis = softmax_7_axis_0, x = add_7)[name = tensor("softmax_7")]; tensor out_29_transpose_x_0 = const()[name = tensor("out_29_transpose_x_0"), val = tensor(false)]; tensor out_29_transpose_y_0 = const()[name = tensor("out_29_transpose_y_0"), val = tensor(false)]; tensor 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("out_29")]; tensor concat_93x = const()[name = tensor("concat_93x"), val = tensor([1, -1])]; tensor var_1484 = reshape(shape = concat_93x, x = positions_15)[name = tensor("op_1484")]; tensor var_1484_promoted_dtype_0 = const()[name = tensor("op_1484_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1484_promoted = cast(dtype = var_1484_promoted_dtype_0, x = var_1484)[name = tensor("cast_437")]; tensor var_1487 = less(x = var_1484_promoted, y = var_1349)[name = tensor("op_1487")]; tensor concat_94x = const()[name = tensor("concat_94x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_15 = reshape(shape = concat_94x, x = var_1487)[name = tensor("valid_q_15")]; tensor var_1491 = const()[name = tensor("op_1491"), val = tensor(0x0p+0)]; tensor out_31 = select(a = out_29, b = var_1491, cond = valid_q_15)[name = tensor("out_31")]; tensor var_1493_perm_0 = const()[name = tensor("op_1493_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_95x = const()[name = tensor("concat_95x"), val = tensor([1, -1, 256])]; tensor var_1493 = transpose(perm = var_1493_perm_0, x = out_31)[name = tensor("transpose_60")]; tensor input_91 = reshape(shape = concat_95x, x = var_1493)[name = tensor("input_91")]; tensor x_63 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_1_self_attn_out_proj_weight, x = input_91)[name = tensor("linear_34")]; tensor var_1499 = mul(x = encoder_5_transformer_layers_1_layer_scale_1_scale, y = x_63)[name = tensor("op_1499")]; tensor input_93 = add(x = input_89, y = var_1499)[name = tensor("input_93")]; tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; tensor 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("input_95")]; tensor input_97 = linear(bias = linear_3_bias_0, weight = encoder_5_transformer_layers_1_ffn_0_weight, x = input_95)[name = tensor("linear_35")]; tensor input_99_mode_0 = const()[name = tensor("input_99_mode_0"), val = tensor("EXACT")]; tensor input_99 = gelu(mode = input_99_mode_0, x = input_97)[name = tensor("input_99")]; tensor x_65 = linear(bias = linear_0_bias_0, weight = encoder_5_transformer_layers_1_ffn_2_weight, x = input_99)[name = tensor("linear_36")]; tensor var_1514 = mul(x = encoder_5_transformer_layers_1_layer_scale_2_scale, y = x_65)[name = tensor("op_1514")]; tensor input_101 = add(x = input_93, y = var_1514)[name = tensor("input_101")]; tensor var_1517 = linear(bias = linear_17_bias_0, weight = encoder_5_output_proj_weight, x = input_101)[name = tensor("linear_37")]; tensor x_67_perm_0 = const()[name = tensor("x_67_perm_0"), val = tensor([0, 2, 1])]; tensor var_1528 = const()[name = tensor("op_1528"), val = tensor([1, 384, -1, 2])]; tensor x_67 = transpose(perm = x_67_perm_0, x = var_1517)[name = tensor("transpose_59")]; tensor var_1529 = reshape(shape = var_1528, x = x_67)[name = tensor("op_1529")]; tensor var_1530 = const()[name = tensor("op_1530"), val = tensor([0, 1, 3, 2])]; tensor var_1535 = const()[name = tensor("op_1535"), val = tensor([1, 768, -1])]; tensor var_1531 = transpose(perm = var_1530, x = var_1529)[name = tensor("transpose_58")]; tensor x_69 = reshape(shape = var_1535, x = var_1531)[name = tensor("x_69")]; tensor var_1519_promoted = const()[name = tensor("op_1519_promoted"), val = tensor(0x1p+1)]; tensor floor_div_3 = floor_div(x = floor_div_2, y = var_1519_promoted)[name = tensor("floor_div_3")]; tensor var_1542 = const()[name = tensor("op_1542"), val = tensor(0)]; tensor var_1549 = const()[name = tensor("op_1549"), val = tensor(500)]; tensor var_1556 = const()[name = tensor("op_1556"), val = tensor(0x1.4f8b58p-17)]; tensor var_1560 = const()[name = tensor("op_1560"), val = tensor(1)]; tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; tensor input_103 = transpose(perm = input_103_perm_0, x = x_69)[name = tensor("transpose_57")]; tensor x_71 = linear(bias = linear_0_bias_0, weight = encoder_7_input_proj_weight, x = input_103)[name = tensor("linear_38")]; tensor x_73_axes_0 = const()[name = tensor("x_73_axes_0"), val = tensor([-1])]; tensor 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("x_73")]; tensor var_1588_shape = shape(x = x_73)[name = tensor("op_1588_shape")]; tensor gather_90_batch_dims_0 = const()[name = tensor("gather_90_batch_dims_0"), val = tensor(0)]; tensor gather_90_validate_indices_0 = const()[name = tensor("gather_90_validate_indices_0"), val = tensor(false)]; tensor select_32 = const()[name = tensor("select_32"), val = tensor(1)]; tensor gather_90_axis_1 = const()[name = tensor("gather_90_axis_1"), val = tensor(0)]; tensor 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("gather_90")]; tensor projected_33 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_0_self_attn_in_proj_weight, x = x_73)[name = tensor("linear_39")]; tensor concat_96x = const()[name = tensor("concat_96x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_1595 = reshape(shape = concat_96x, x = projected_33)[name = tensor("op_1595")]; tensor var_1596 = const()[name = tensor("op_1596"), val = tensor([2, 0, 3, 1, 4])]; tensor q_33_begin_0 = const()[name = tensor("q_33_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_33_end_0 = const()[name = tensor("q_33_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_33_end_mask_0 = const()[name = tensor("q_33_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_33_squeeze_mask_0 = const()[name = tensor("q_33_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_35 = transpose(perm = var_1596, x = var_1595)[name = tensor("transpose_56")]; tensor 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("q_33")]; tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_33")]; tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_17")]; tensor freqs_17 = const()[name = tensor("freqs_17"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44432640)))]; tensor var_1608_shape = shape(x = q_33)[name = tensor("op_1608_shape")]; tensor gather_94_batch_dims_0 = const()[name = tensor("gather_94_batch_dims_0"), val = tensor(0)]; tensor gather_94_validate_indices_0 = const()[name = tensor("gather_94_validate_indices_0"), val = tensor(false)]; tensor select_33 = const()[name = tensor("select_33"), val = tensor(2)]; tensor gather_94_axis_1 = const()[name = tensor("gather_94_axis_1"), val = tensor(0)]; tensor 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("gather_94")]; tensor var_1611 = const()[name = tensor("op_1611"), val = tensor([[0x0p+0]])]; tensor const_75 = const()[name = tensor("const_75"), val = tensor(0)]; tensor const_76 = const()[name = tensor("const_76"), val = tensor(1)]; tensor var_1612 = range_1d(end = gather_94, start = const_75, step = const_76)[name = tensor("op_1612")]; tensor var_1612_promoted_dtype_0 = const()[name = tensor("op_1612_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1612_promoted = cast(dtype = var_1612_promoted_dtype_0, x = var_1612)[name = tensor("cast_436")]; tensor ts_33 = add(x = var_1611, y = var_1612_promoted)[name = tensor("ts_33")]; tensor var_1615 = const()[name = tensor("op_1615"), val = tensor([1, 1, -1, 1])]; tensor ts_35 = reshape(shape = var_1615, x = ts_33)[name = tensor("ts_35")]; tensor concat_97x = const()[name = tensor("concat_97x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_17 = reshape(shape = concat_97x, x = q_33)[name = tensor("q2_17")]; tensor concat_98x = const()[name = tensor("concat_98x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_17 = reshape(shape = concat_98x, x = k_33)[name = tensor("k2_17")]; tensor qr_17_begin_0 = const()[name = tensor("qr_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_17_end_0 = const()[name = tensor("qr_17_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_17_end_mask_0 = const()[name = tensor("qr_17_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_17_squeeze_mask_0 = const()[name = tensor("qr_17_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_17")]; tensor qi_17_begin_0 = const()[name = tensor("qi_17_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_17_end_0 = const()[name = tensor("qi_17_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_17_end_mask_0 = const()[name = tensor("qi_17_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_17_squeeze_mask_0 = const()[name = tensor("qi_17_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_17")]; tensor kr_17_begin_0 = const()[name = tensor("kr_17_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_17_end_0 = const()[name = tensor("kr_17_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_17_end_mask_0 = const()[name = tensor("kr_17_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_17_squeeze_mask_0 = const()[name = tensor("kr_17_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_17")]; tensor ki_17_begin_0 = const()[name = tensor("ki_17_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_17_end_0 = const()[name = tensor("ki_17_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_17_end_mask_0 = const()[name = tensor("ki_17_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_17_squeeze_mask_0 = const()[name = tensor("ki_17_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_17")]; tensor var_1628 = mul(x = freqs_17, y = ts_35)[name = tensor("op_1628")]; tensor rotr_17 = cos(x = var_1628)[name = tensor("rotr_17")]; tensor roti_17 = sin(x = var_1628)[name = tensor("roti_17")]; tensor var_1632 = mul(x = qr_17, y = rotr_17)[name = tensor("op_1632")]; tensor var_1633 = mul(x = qi_17, y = roti_17)[name = tensor("op_1633")]; tensor var_1634 = sub(x = var_1632, y = var_1633)[name = tensor("op_1634")]; tensor var_1635 = mul(x = qr_17, y = roti_17)[name = tensor("op_1635")]; tensor var_1636 = mul(x = qi_17, y = rotr_17)[name = tensor("op_1636")]; tensor var_1637 = add(x = var_1635, y = var_1636)[name = tensor("op_1637")]; tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; tensor qo_17 = stack(axis = qo_17_axis_0, values = (var_1634, var_1637))[name = tensor("qo_17")]; tensor var_1640 = mul(x = kr_17, y = rotr_17)[name = tensor("op_1640")]; tensor var_1641 = mul(x = ki_17, y = roti_17)[name = tensor("op_1641")]; tensor var_1642 = sub(x = var_1640, y = var_1641)[name = tensor("op_1642")]; tensor var_1643 = mul(x = kr_17, y = roti_17)[name = tensor("op_1643")]; tensor var_1644 = mul(x = ki_17, y = rotr_17)[name = tensor("op_1644")]; tensor var_1645 = add(x = var_1643, y = var_1644)[name = tensor("op_1645")]; tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; tensor ko_17 = stack(axis = ko_17_axis_0, values = (var_1642, var_1645))[name = tensor("ko_17")]; tensor concat_99x = const()[name = tensor("concat_99x"), val = tensor([1, 4, -1, 64])]; tensor q_35 = reshape(shape = concat_99x, x = qo_17)[name = tensor("q_35")]; tensor concat_100x = const()[name = tensor("concat_100x"), val = tensor([1, 4, -1, 64])]; tensor k_35 = reshape(shape = concat_100x, x = ko_17)[name = tensor("k_35")]; tensor const_77 = const()[name = tensor("const_77"), val = tensor(0)]; tensor const_78 = const()[name = tensor("const_78"), val = tensor(1)]; tensor positions_17 = range_1d(end = gather_90, start = const_77, step = const_78)[name = tensor("positions_17")]; tensor concat_101x = const()[name = tensor("concat_101x"), val = tensor([1, 1, -1])]; tensor var_1657 = reshape(shape = concat_101x, x = positions_17)[name = tensor("op_1657")]; tensor var_1658 = const()[name = tensor("op_1658"), val = tensor([-1, 1, 1])]; tensor var_1659 = reshape(shape = var_1658, x = floor_div_3)[name = tensor("op_1659")]; tensor var_1657_promoted_dtype_0 = const()[name = tensor("op_1657_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1657_promoted = cast(dtype = var_1657_promoted_dtype_0, x = var_1657)[name = tensor("cast_435")]; tensor valid_k_17 = less(x = var_1657_promoted, y = var_1659)[name = tensor("valid_k_17")]; tensor concat_102x = const()[name = tensor("concat_102x"), val = tensor([1, -1, 1])]; tensor var_1662 = reshape(shape = concat_102x, x = positions_17)[name = tensor("op_1662")]; tensor delta_17 = sub(x = var_1662, y = var_1657)[name = tensor("delta_17")]; tensor concat_104_axis_0 = const()[name = tensor("concat_104_axis_0"), val = tensor(0)]; tensor concat_104_interleave_0 = const()[name = tensor("concat_104_interleave_0"), val = tensor(false)]; tensor concat_104 = concat(axis = concat_104_axis_0, interleave = concat_104_interleave_0, values = (var_1560, gather_90, gather_90))[name = tensor("concat_104")]; tensor fill_8_value_0 = const()[name = tensor("fill_8_value_0"), val = tensor(0x1p+0)]; tensor fill_8 = fill(shape = concat_104, value = fill_8_value_0)[name = tensor("fill_8")]; tensor attn_bias_65_dtype_0 = const()[name = tensor("attn_bias_65_dtype_0"), val = tensor("bool")]; tensor var_1668 = greater_equal(x = delta_17, y = var_1542)[name = tensor("op_1668")]; tensor attn_bias_65 = cast(dtype = attn_bias_65_dtype_0, x = fill_8)[name = tensor("cast_434")]; tensor attn_bias_67 = logical_and(x = attn_bias_65, y = var_1668)[name = tensor("attn_bias_67")]; tensor var_1670 = less(x = delta_17, y = var_1549)[name = tensor("op_1670")]; tensor attn_bias_69 = logical_and(x = attn_bias_67, y = var_1670)[name = tensor("attn_bias_69")]; tensor var_1672 = logical_and(x = attn_bias_69, y = valid_k_17)[name = tensor("op_1672")]; tensor var_1674_axes_0 = const()[name = tensor("op_1674_axes_0"), val = tensor([1])]; tensor var_1674 = expand_dims(axes = var_1674_axes_0, x = var_1672)[name = tensor("op_1674")]; tensor cast_202_dtype_0 = const()[name = tensor("cast_202_dtype_0"), val = tensor("fp32")]; tensor sub_8_x_0 = const()[name = tensor("sub_8_x_0"), val = tensor(0x1p+0)]; tensor cast_202 = cast(dtype = cast_202_dtype_0, x = var_1674)[name = tensor("cast_433")]; tensor sub_8 = sub(x = sub_8_x_0, y = cast_202)[name = tensor("sub_8")]; tensor mul_16_x_0 = const()[name = tensor("mul_16_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_16 = mul(x = mul_16_x_0, y = sub_8)[name = tensor("mul_16")]; tensor mul_17_y_0 = const()[name = tensor("mul_17_y_0"), val = tensor(0x1p-3)]; tensor mul_17 = mul(x = q_35, y = mul_17_y_0)[name = tensor("mul_17")]; tensor matmul_8_transpose_y_0 = const()[name = tensor("matmul_8_transpose_y_0"), val = tensor(true)]; tensor matmul_8_transpose_x_0 = const()[name = tensor("matmul_8_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_8")]; tensor add_8 = add(x = matmul_8, y = mul_16)[name = tensor("add_8")]; tensor softmax_8_axis_0 = const()[name = tensor("softmax_8_axis_0"), val = tensor(-1)]; tensor softmax_8 = softmax(axis = softmax_8_axis_0, x = add_8)[name = tensor("softmax_8")]; tensor out_33_transpose_x_0 = const()[name = tensor("out_33_transpose_x_0"), val = tensor(false)]; tensor out_33_transpose_y_0 = const()[name = tensor("out_33_transpose_y_0"), val = tensor(false)]; tensor 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("out_33")]; tensor concat_105x = const()[name = tensor("concat_105x"), val = tensor([1, -1])]; tensor var_1680 = reshape(shape = concat_105x, x = positions_17)[name = tensor("op_1680")]; tensor var_1681 = const()[name = tensor("op_1681"), val = tensor([-1, 1])]; tensor var_1682 = reshape(shape = var_1681, x = floor_div_3)[name = tensor("op_1682")]; tensor var_1680_promoted_dtype_0 = const()[name = tensor("op_1680_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1680_promoted = cast(dtype = var_1680_promoted_dtype_0, x = var_1680)[name = tensor("cast_432")]; tensor var_1683 = less(x = var_1680_promoted, y = var_1682)[name = tensor("op_1683")]; tensor concat_106x = const()[name = tensor("concat_106x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_17 = reshape(shape = concat_106x, x = var_1683)[name = tensor("valid_q_17")]; tensor var_1687 = const()[name = tensor("op_1687"), val = tensor(0x0p+0)]; tensor out_35 = select(a = out_33, b = var_1687, cond = valid_q_17)[name = tensor("out_35")]; tensor var_1689_perm_0 = const()[name = tensor("op_1689_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_107x = const()[name = tensor("concat_107x"), val = tensor([1, -1, 256])]; tensor var_1689 = transpose(perm = var_1689_perm_0, x = out_35)[name = tensor("transpose_55")]; tensor input_105 = reshape(shape = concat_107x, x = var_1689)[name = tensor("input_105")]; tensor x_75 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_0_self_attn_out_proj_weight, x = input_105)[name = tensor("linear_40")]; tensor var_1695 = mul(x = encoder_7_transformer_layers_0_layer_scale_1_scale, y = x_75)[name = tensor("op_1695")]; tensor input_107 = add(x = x_71, y = var_1695)[name = tensor("input_107")]; tensor input_109_axes_0 = const()[name = tensor("input_109_axes_0"), val = tensor([-1])]; tensor 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("input_109")]; tensor input_111 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_0_ffn_0_weight, x = input_109)[name = tensor("linear_41")]; tensor input_113_mode_0 = const()[name = tensor("input_113_mode_0"), val = tensor("EXACT")]; tensor input_113 = gelu(mode = input_113_mode_0, x = input_111)[name = tensor("input_113")]; tensor x_77 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_0_ffn_2_weight, x = input_113)[name = tensor("linear_42")]; tensor var_1710 = mul(x = encoder_7_transformer_layers_0_layer_scale_2_scale, y = x_77)[name = tensor("op_1710")]; tensor input_115 = add(x = input_107, y = var_1710)[name = tensor("input_115")]; tensor x_79_axes_0 = const()[name = tensor("x_79_axes_0"), val = tensor([-1])]; tensor 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("x_79")]; tensor var_1725_shape = shape(x = x_79)[name = tensor("op_1725_shape")]; tensor gather_100_batch_dims_0 = const()[name = tensor("gather_100_batch_dims_0"), val = tensor(0)]; tensor gather_100_validate_indices_0 = const()[name = tensor("gather_100_validate_indices_0"), val = tensor(false)]; tensor select_34 = const()[name = tensor("select_34"), val = tensor(1)]; tensor gather_100_axis_1 = const()[name = tensor("gather_100_axis_1"), val = tensor(0)]; tensor 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("gather_100")]; tensor projected_37 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_1_self_attn_in_proj_weight, x = x_79)[name = tensor("linear_43")]; tensor concat_108x = const()[name = tensor("concat_108x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_1732 = reshape(shape = concat_108x, x = projected_37)[name = tensor("op_1732")]; tensor var_1733 = const()[name = tensor("op_1733"), val = tensor([2, 0, 3, 1, 4])]; tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_39 = transpose(perm = var_1733, x = var_1732)[name = tensor("transpose_54")]; tensor 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("q_37")]; tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_37")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_19")]; tensor freqs_19 = const()[name = tensor("freqs_19"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44432832)))]; tensor var_1745_shape = shape(x = q_37)[name = tensor("op_1745_shape")]; tensor gather_104_batch_dims_0 = const()[name = tensor("gather_104_batch_dims_0"), val = tensor(0)]; tensor gather_104_validate_indices_0 = const()[name = tensor("gather_104_validate_indices_0"), val = tensor(false)]; tensor select_35 = const()[name = tensor("select_35"), val = tensor(2)]; tensor gather_104_axis_1 = const()[name = tensor("gather_104_axis_1"), val = tensor(0)]; tensor 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("gather_104")]; tensor var_1748 = const()[name = tensor("op_1748"), val = tensor([[0x0p+0]])]; tensor const_84 = const()[name = tensor("const_84"), val = tensor(0)]; tensor const_85 = const()[name = tensor("const_85"), val = tensor(1)]; tensor var_1749 = range_1d(end = gather_104, start = const_84, step = const_85)[name = tensor("op_1749")]; tensor var_1749_promoted_dtype_0 = const()[name = tensor("op_1749_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1749_promoted = cast(dtype = var_1749_promoted_dtype_0, x = var_1749)[name = tensor("cast_431")]; tensor ts_37 = add(x = var_1748, y = var_1749_promoted)[name = tensor("ts_37")]; tensor var_1752 = const()[name = tensor("op_1752"), val = tensor([1, 1, -1, 1])]; tensor ts_39 = reshape(shape = var_1752, x = ts_37)[name = tensor("ts_39")]; tensor concat_109x = const()[name = tensor("concat_109x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_19 = reshape(shape = concat_109x, x = q_37)[name = tensor("q2_19")]; tensor concat_110x = const()[name = tensor("concat_110x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_19 = reshape(shape = concat_110x, x = k_37)[name = tensor("k2_19")]; tensor qr_19_begin_0 = const()[name = tensor("qr_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_19_end_0 = const()[name = tensor("qr_19_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_19_end_mask_0 = const()[name = tensor("qr_19_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_19_squeeze_mask_0 = const()[name = tensor("qr_19_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_19")]; tensor qi_19_begin_0 = const()[name = tensor("qi_19_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_19_end_0 = const()[name = tensor("qi_19_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_19_end_mask_0 = const()[name = tensor("qi_19_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_19_squeeze_mask_0 = const()[name = tensor("qi_19_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_19")]; tensor kr_19_begin_0 = const()[name = tensor("kr_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_19_end_0 = const()[name = tensor("kr_19_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_19_end_mask_0 = const()[name = tensor("kr_19_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_19_squeeze_mask_0 = const()[name = tensor("kr_19_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_19")]; tensor ki_19_begin_0 = const()[name = tensor("ki_19_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_19_end_0 = const()[name = tensor("ki_19_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_19_end_mask_0 = const()[name = tensor("ki_19_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_19_squeeze_mask_0 = const()[name = tensor("ki_19_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_19")]; tensor var_1765 = mul(x = freqs_19, y = ts_39)[name = tensor("op_1765")]; tensor rotr_19 = cos(x = var_1765)[name = tensor("rotr_19")]; tensor roti_19 = sin(x = var_1765)[name = tensor("roti_19")]; tensor var_1769 = mul(x = qr_19, y = rotr_19)[name = tensor("op_1769")]; tensor var_1770 = mul(x = qi_19, y = roti_19)[name = tensor("op_1770")]; tensor var_1771 = sub(x = var_1769, y = var_1770)[name = tensor("op_1771")]; tensor var_1772 = mul(x = qr_19, y = roti_19)[name = tensor("op_1772")]; tensor var_1773 = mul(x = qi_19, y = rotr_19)[name = tensor("op_1773")]; tensor var_1774 = add(x = var_1772, y = var_1773)[name = tensor("op_1774")]; tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; tensor qo_19 = stack(axis = qo_19_axis_0, values = (var_1771, var_1774))[name = tensor("qo_19")]; tensor var_1777 = mul(x = kr_19, y = rotr_19)[name = tensor("op_1777")]; tensor var_1778 = mul(x = ki_19, y = roti_19)[name = tensor("op_1778")]; tensor var_1779 = sub(x = var_1777, y = var_1778)[name = tensor("op_1779")]; tensor var_1780 = mul(x = kr_19, y = roti_19)[name = tensor("op_1780")]; tensor var_1781 = mul(x = ki_19, y = rotr_19)[name = tensor("op_1781")]; tensor var_1782 = add(x = var_1780, y = var_1781)[name = tensor("op_1782")]; tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; tensor ko_19 = stack(axis = ko_19_axis_0, values = (var_1779, var_1782))[name = tensor("ko_19")]; tensor concat_111x = const()[name = tensor("concat_111x"), val = tensor([1, 4, -1, 64])]; tensor q_39 = reshape(shape = concat_111x, x = qo_19)[name = tensor("q_39")]; tensor concat_112x = const()[name = tensor("concat_112x"), val = tensor([1, 4, -1, 64])]; tensor k_39 = reshape(shape = concat_112x, x = ko_19)[name = tensor("k_39")]; tensor const_86 = const()[name = tensor("const_86"), val = tensor(0)]; tensor const_87 = const()[name = tensor("const_87"), val = tensor(1)]; tensor positions_19 = range_1d(end = gather_100, start = const_86, step = const_87)[name = tensor("positions_19")]; tensor concat_113x = const()[name = tensor("concat_113x"), val = tensor([1, 1, -1])]; tensor var_1794 = reshape(shape = concat_113x, x = positions_19)[name = tensor("op_1794")]; tensor var_1794_promoted_dtype_0 = const()[name = tensor("op_1794_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1794_promoted = cast(dtype = var_1794_promoted_dtype_0, x = var_1794)[name = tensor("cast_430")]; tensor valid_k_19 = less(x = var_1794_promoted, y = var_1659)[name = tensor("valid_k_19")]; tensor concat_114x = const()[name = tensor("concat_114x"), val = tensor([1, -1, 1])]; tensor var_1799 = reshape(shape = concat_114x, x = positions_19)[name = tensor("op_1799")]; tensor delta_19 = sub(x = var_1799, y = var_1794)[name = tensor("delta_19")]; tensor concat_116_axis_0 = const()[name = tensor("concat_116_axis_0"), val = tensor(0)]; tensor concat_116_interleave_0 = const()[name = tensor("concat_116_interleave_0"), val = tensor(false)]; tensor concat_116 = concat(axis = concat_116_axis_0, interleave = concat_116_interleave_0, values = (var_1560, gather_100, gather_100))[name = tensor("concat_116")]; tensor fill_9_value_0 = const()[name = tensor("fill_9_value_0"), val = tensor(0x1p+0)]; tensor fill_9 = fill(shape = concat_116, value = fill_9_value_0)[name = tensor("fill_9")]; tensor attn_bias_73_dtype_0 = const()[name = tensor("attn_bias_73_dtype_0"), val = tensor("bool")]; tensor var_1805 = greater_equal(x = delta_19, y = var_1542)[name = tensor("op_1805")]; tensor attn_bias_73 = cast(dtype = attn_bias_73_dtype_0, x = fill_9)[name = tensor("cast_429")]; tensor attn_bias_75 = logical_and(x = attn_bias_73, y = var_1805)[name = tensor("attn_bias_75")]; tensor var_1807 = less(x = delta_19, y = var_1549)[name = tensor("op_1807")]; tensor attn_bias_77 = logical_and(x = attn_bias_75, y = var_1807)[name = tensor("attn_bias_77")]; tensor var_1809 = logical_and(x = attn_bias_77, y = valid_k_19)[name = tensor("op_1809")]; tensor var_1811_axes_0 = const()[name = tensor("op_1811_axes_0"), val = tensor([1])]; tensor var_1811 = expand_dims(axes = var_1811_axes_0, x = var_1809)[name = tensor("op_1811")]; tensor cast_224_dtype_0 = const()[name = tensor("cast_224_dtype_0"), val = tensor("fp32")]; tensor sub_9_x_0 = const()[name = tensor("sub_9_x_0"), val = tensor(0x1p+0)]; tensor cast_224 = cast(dtype = cast_224_dtype_0, x = var_1811)[name = tensor("cast_428")]; tensor sub_9 = sub(x = sub_9_x_0, y = cast_224)[name = tensor("sub_9")]; tensor mul_18_x_0 = const()[name = tensor("mul_18_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_18 = mul(x = mul_18_x_0, y = sub_9)[name = tensor("mul_18")]; tensor mul_19_y_0 = const()[name = tensor("mul_19_y_0"), val = tensor(0x1p-3)]; tensor mul_19 = mul(x = q_39, y = mul_19_y_0)[name = tensor("mul_19")]; tensor matmul_9_transpose_y_0 = const()[name = tensor("matmul_9_transpose_y_0"), val = tensor(true)]; tensor matmul_9_transpose_x_0 = const()[name = tensor("matmul_9_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_9")]; tensor add_9 = add(x = matmul_9, y = mul_18)[name = tensor("add_9")]; tensor softmax_9_axis_0 = const()[name = tensor("softmax_9_axis_0"), val = tensor(-1)]; tensor softmax_9 = softmax(axis = softmax_9_axis_0, x = add_9)[name = tensor("softmax_9")]; tensor out_37_transpose_x_0 = const()[name = tensor("out_37_transpose_x_0"), val = tensor(false)]; tensor out_37_transpose_y_0 = const()[name = tensor("out_37_transpose_y_0"), val = tensor(false)]; tensor 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("out_37")]; tensor concat_117x = const()[name = tensor("concat_117x"), val = tensor([1, -1])]; tensor var_1817 = reshape(shape = concat_117x, x = positions_19)[name = tensor("op_1817")]; tensor var_1817_promoted_dtype_0 = const()[name = tensor("op_1817_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1817_promoted = cast(dtype = var_1817_promoted_dtype_0, x = var_1817)[name = tensor("cast_427")]; tensor var_1820 = less(x = var_1817_promoted, y = var_1682)[name = tensor("op_1820")]; tensor concat_118x = const()[name = tensor("concat_118x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_19 = reshape(shape = concat_118x, x = var_1820)[name = tensor("valid_q_19")]; tensor var_1824 = const()[name = tensor("op_1824"), val = tensor(0x0p+0)]; tensor out_39 = select(a = out_37, b = var_1824, cond = valid_q_19)[name = tensor("out_39")]; tensor var_1826_perm_0 = const()[name = tensor("op_1826_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_119x = const()[name = tensor("concat_119x"), val = tensor([1, -1, 256])]; tensor var_1826 = transpose(perm = var_1826_perm_0, x = out_39)[name = tensor("transpose_53")]; tensor input_117 = reshape(shape = concat_119x, x = var_1826)[name = tensor("input_117")]; tensor x_81 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_1_self_attn_out_proj_weight, x = input_117)[name = tensor("linear_44")]; tensor var_1832 = mul(x = encoder_7_transformer_layers_1_layer_scale_1_scale, y = x_81)[name = tensor("op_1832")]; tensor input_119 = add(x = input_115, y = var_1832)[name = tensor("input_119")]; tensor input_121_axes_0 = const()[name = tensor("input_121_axes_0"), val = tensor([-1])]; tensor 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("input_121")]; tensor input_123 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_1_ffn_0_weight, x = input_121)[name = tensor("linear_45")]; tensor input_125_mode_0 = const()[name = tensor("input_125_mode_0"), val = tensor("EXACT")]; tensor input_125 = gelu(mode = input_125_mode_0, x = input_123)[name = tensor("input_125")]; tensor x_83 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_1_ffn_2_weight, x = input_125)[name = tensor("linear_46")]; tensor var_1847 = mul(x = encoder_7_transformer_layers_1_layer_scale_2_scale, y = x_83)[name = tensor("op_1847")]; tensor input_127 = add(x = input_119, y = var_1847)[name = tensor("input_127")]; tensor x_85_axes_0 = const()[name = tensor("x_85_axes_0"), val = tensor([-1])]; tensor 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("x_85")]; tensor var_1862_shape = shape(x = x_85)[name = tensor("op_1862_shape")]; tensor gather_110_batch_dims_0 = const()[name = tensor("gather_110_batch_dims_0"), val = tensor(0)]; tensor gather_110_validate_indices_0 = const()[name = tensor("gather_110_validate_indices_0"), val = tensor(false)]; tensor select_36 = const()[name = tensor("select_36"), val = tensor(1)]; tensor gather_110_axis_1 = const()[name = tensor("gather_110_axis_1"), val = tensor(0)]; tensor 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("gather_110")]; tensor projected_41 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_2_self_attn_in_proj_weight, x = x_85)[name = tensor("linear_47")]; tensor concat_120x = const()[name = tensor("concat_120x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_1869 = reshape(shape = concat_120x, x = projected_41)[name = tensor("op_1869")]; tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([2, 0, 3, 1, 4])]; tensor q_41_begin_0 = const()[name = tensor("q_41_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_41_end_0 = const()[name = tensor("q_41_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_41_end_mask_0 = const()[name = tensor("q_41_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_41_squeeze_mask_0 = const()[name = tensor("q_41_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected_43 = transpose(perm = var_1870, x = var_1869)[name = tensor("transpose_52")]; tensor 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("q_41")]; tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_41")]; tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v_21")]; tensor freqs_21 = const()[name = tensor("freqs_21"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44433024)))]; tensor var_1882_shape = shape(x = q_41)[name = tensor("op_1882_shape")]; tensor gather_114_batch_dims_0 = const()[name = tensor("gather_114_batch_dims_0"), val = tensor(0)]; tensor gather_114_validate_indices_0 = const()[name = tensor("gather_114_validate_indices_0"), val = tensor(false)]; tensor select_37 = const()[name = tensor("select_37"), val = tensor(2)]; tensor gather_114_axis_1 = const()[name = tensor("gather_114_axis_1"), val = tensor(0)]; tensor 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("gather_114")]; tensor var_1885 = const()[name = tensor("op_1885"), val = tensor([[0x0p+0]])]; tensor const_93 = const()[name = tensor("const_93"), val = tensor(0)]; tensor const_94 = const()[name = tensor("const_94"), val = tensor(1)]; tensor var_1886 = range_1d(end = gather_114, start = const_93, step = const_94)[name = tensor("op_1886")]; tensor var_1886_promoted_dtype_0 = const()[name = tensor("op_1886_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1886_promoted = cast(dtype = var_1886_promoted_dtype_0, x = var_1886)[name = tensor("cast_426")]; tensor ts_41 = add(x = var_1885, y = var_1886_promoted)[name = tensor("ts_41")]; tensor var_1889 = const()[name = tensor("op_1889"), val = tensor([1, 1, -1, 1])]; tensor ts_43 = reshape(shape = var_1889, x = ts_41)[name = tensor("ts_43")]; tensor concat_121x = const()[name = tensor("concat_121x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2_21 = reshape(shape = concat_121x, x = q_41)[name = tensor("q2_21")]; tensor concat_122x = const()[name = tensor("concat_122x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2_21 = reshape(shape = concat_122x, x = k_41)[name = tensor("k2_21")]; tensor qr_21_begin_0 = const()[name = tensor("qr_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_21_end_0 = const()[name = tensor("qr_21_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_21_end_mask_0 = const()[name = tensor("qr_21_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_21_squeeze_mask_0 = const()[name = tensor("qr_21_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr_21")]; tensor qi_21_begin_0 = const()[name = tensor("qi_21_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_21_end_0 = const()[name = tensor("qi_21_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_21_end_mask_0 = const()[name = tensor("qi_21_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_21_squeeze_mask_0 = const()[name = tensor("qi_21_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi_21")]; tensor kr_21_begin_0 = const()[name = tensor("kr_21_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_21_end_0 = const()[name = tensor("kr_21_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_21_end_mask_0 = const()[name = tensor("kr_21_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_21_squeeze_mask_0 = const()[name = tensor("kr_21_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr_21")]; tensor ki_21_begin_0 = const()[name = tensor("ki_21_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_21_end_0 = const()[name = tensor("ki_21_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_21_end_mask_0 = const()[name = tensor("ki_21_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_21_squeeze_mask_0 = const()[name = tensor("ki_21_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki_21")]; tensor var_1902 = mul(x = freqs_21, y = ts_43)[name = tensor("op_1902")]; tensor rotr_21 = cos(x = var_1902)[name = tensor("rotr_21")]; tensor roti_21 = sin(x = var_1902)[name = tensor("roti_21")]; tensor var_1906 = mul(x = qr_21, y = rotr_21)[name = tensor("op_1906")]; tensor var_1907 = mul(x = qi_21, y = roti_21)[name = tensor("op_1907")]; tensor var_1908 = sub(x = var_1906, y = var_1907)[name = tensor("op_1908")]; tensor var_1909 = mul(x = qr_21, y = roti_21)[name = tensor("op_1909")]; tensor var_1910 = mul(x = qi_21, y = rotr_21)[name = tensor("op_1910")]; tensor var_1911 = add(x = var_1909, y = var_1910)[name = tensor("op_1911")]; tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; tensor qo_21 = stack(axis = qo_21_axis_0, values = (var_1908, var_1911))[name = tensor("qo_21")]; tensor var_1914 = mul(x = kr_21, y = rotr_21)[name = tensor("op_1914")]; tensor var_1915 = mul(x = ki_21, y = roti_21)[name = tensor("op_1915")]; tensor var_1916 = sub(x = var_1914, y = var_1915)[name = tensor("op_1916")]; tensor var_1917 = mul(x = kr_21, y = roti_21)[name = tensor("op_1917")]; tensor var_1918 = mul(x = ki_21, y = rotr_21)[name = tensor("op_1918")]; tensor var_1919 = add(x = var_1917, y = var_1918)[name = tensor("op_1919")]; tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; tensor ko_21 = stack(axis = ko_21_axis_0, values = (var_1916, var_1919))[name = tensor("ko_21")]; tensor concat_123x = const()[name = tensor("concat_123x"), val = tensor([1, 4, -1, 64])]; tensor q_43 = reshape(shape = concat_123x, x = qo_21)[name = tensor("q_43")]; tensor concat_124x = const()[name = tensor("concat_124x"), val = tensor([1, 4, -1, 64])]; tensor k_43 = reshape(shape = concat_124x, x = ko_21)[name = tensor("k_43")]; tensor const_95 = const()[name = tensor("const_95"), val = tensor(0)]; tensor const_96 = const()[name = tensor("const_96"), val = tensor(1)]; tensor positions_21 = range_1d(end = gather_110, start = const_95, step = const_96)[name = tensor("positions_21")]; tensor concat_125x = const()[name = tensor("concat_125x"), val = tensor([1, 1, -1])]; tensor var_1931 = reshape(shape = concat_125x, x = positions_21)[name = tensor("op_1931")]; tensor var_1931_promoted_dtype_0 = const()[name = tensor("op_1931_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1931_promoted = cast(dtype = var_1931_promoted_dtype_0, x = var_1931)[name = tensor("cast_425")]; tensor valid_k_21 = less(x = var_1931_promoted, y = var_1659)[name = tensor("valid_k_21")]; tensor concat_126x = const()[name = tensor("concat_126x"), val = tensor([1, -1, 1])]; tensor var_1936 = reshape(shape = concat_126x, x = positions_21)[name = tensor("op_1936")]; tensor delta_21 = sub(x = var_1936, y = var_1931)[name = tensor("delta_21")]; tensor concat_128_axis_0 = const()[name = tensor("concat_128_axis_0"), val = tensor(0)]; tensor concat_128_interleave_0 = const()[name = tensor("concat_128_interleave_0"), val = tensor(false)]; tensor concat_128 = concat(axis = concat_128_axis_0, interleave = concat_128_interleave_0, values = (var_1560, gather_110, gather_110))[name = tensor("concat_128")]; tensor fill_10_value_0 = const()[name = tensor("fill_10_value_0"), val = tensor(0x1p+0)]; tensor fill_10 = fill(shape = concat_128, value = fill_10_value_0)[name = tensor("fill_10")]; tensor attn_bias_81_dtype_0 = const()[name = tensor("attn_bias_81_dtype_0"), val = tensor("bool")]; tensor var_1942 = greater_equal(x = delta_21, y = var_1542)[name = tensor("op_1942")]; tensor attn_bias_81 = cast(dtype = attn_bias_81_dtype_0, x = fill_10)[name = tensor("cast_424")]; tensor attn_bias_83 = logical_and(x = attn_bias_81, y = var_1942)[name = tensor("attn_bias_83")]; tensor var_1944 = less(x = delta_21, y = var_1549)[name = tensor("op_1944")]; tensor attn_bias_85 = logical_and(x = attn_bias_83, y = var_1944)[name = tensor("attn_bias_85")]; tensor var_1946 = logical_and(x = attn_bias_85, y = valid_k_21)[name = tensor("op_1946")]; tensor var_1948_axes_0 = const()[name = tensor("op_1948_axes_0"), val = tensor([1])]; tensor var_1948 = expand_dims(axes = var_1948_axes_0, x = var_1946)[name = tensor("op_1948")]; tensor cast_246_dtype_0 = const()[name = tensor("cast_246_dtype_0"), val = tensor("fp32")]; tensor sub_10_x_0 = const()[name = tensor("sub_10_x_0"), val = tensor(0x1p+0)]; tensor cast_246 = cast(dtype = cast_246_dtype_0, x = var_1948)[name = tensor("cast_423")]; tensor sub_10 = sub(x = sub_10_x_0, y = cast_246)[name = tensor("sub_10")]; tensor mul_20_x_0 = const()[name = tensor("mul_20_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_20 = mul(x = mul_20_x_0, y = sub_10)[name = tensor("mul_20")]; tensor mul_21_y_0 = const()[name = tensor("mul_21_y_0"), val = tensor(0x1p-3)]; tensor mul_21 = mul(x = q_43, y = mul_21_y_0)[name = tensor("mul_21")]; tensor matmul_10_transpose_y_0 = const()[name = tensor("matmul_10_transpose_y_0"), val = tensor(true)]; tensor matmul_10_transpose_x_0 = const()[name = tensor("matmul_10_transpose_x_0"), val = tensor(false)]; tensor 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("matmul_10")]; tensor add_10 = add(x = matmul_10, y = mul_20)[name = tensor("add_10")]; tensor softmax_10_axis_0 = const()[name = tensor("softmax_10_axis_0"), val = tensor(-1)]; tensor softmax_10 = softmax(axis = softmax_10_axis_0, x = add_10)[name = tensor("softmax_10")]; tensor out_41_transpose_x_0 = const()[name = tensor("out_41_transpose_x_0"), val = tensor(false)]; tensor out_41_transpose_y_0 = const()[name = tensor("out_41_transpose_y_0"), val = tensor(false)]; tensor 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("out_41")]; tensor concat_129x = const()[name = tensor("concat_129x"), val = tensor([1, -1])]; tensor var_1954 = reshape(shape = concat_129x, x = positions_21)[name = tensor("op_1954")]; tensor var_1954_promoted_dtype_0 = const()[name = tensor("op_1954_promoted_dtype_0"), val = tensor("fp32")]; tensor var_1954_promoted = cast(dtype = var_1954_promoted_dtype_0, x = var_1954)[name = tensor("cast_422")]; tensor var_1957 = less(x = var_1954_promoted, y = var_1682)[name = tensor("op_1957")]; tensor concat_130x = const()[name = tensor("concat_130x"), val = tensor([1, 1, -1, 1])]; tensor valid_q_21 = reshape(shape = concat_130x, x = var_1957)[name = tensor("valid_q_21")]; tensor var_1961 = const()[name = tensor("op_1961"), val = tensor(0x0p+0)]; tensor out_43 = select(a = out_41, b = var_1961, cond = valid_q_21)[name = tensor("out_43")]; tensor var_1963_perm_0 = const()[name = tensor("op_1963_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_131x = const()[name = tensor("concat_131x"), val = tensor([1, -1, 256])]; tensor var_1963 = transpose(perm = var_1963_perm_0, x = out_43)[name = tensor("transpose_51")]; tensor input_129 = reshape(shape = concat_131x, x = var_1963)[name = tensor("input_129")]; tensor x_87 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_2_self_attn_out_proj_weight, x = input_129)[name = tensor("linear_48")]; tensor var_1969 = mul(x = encoder_7_transformer_layers_2_layer_scale_1_scale, y = x_87)[name = tensor("op_1969")]; tensor input_131 = add(x = input_127, y = var_1969)[name = tensor("input_131")]; tensor input_133_axes_0 = const()[name = tensor("input_133_axes_0"), val = tensor([-1])]; tensor 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("input_133")]; tensor input_135 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_2_ffn_0_weight, x = input_133)[name = tensor("linear_49")]; tensor input_137_mode_0 = const()[name = tensor("input_137_mode_0"), val = tensor("EXACT")]; tensor input_137 = gelu(mode = input_137_mode_0, x = input_135)[name = tensor("input_137")]; tensor x_89 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_2_ffn_2_weight, x = input_137)[name = tensor("linear_50")]; tensor var_1984 = mul(x = encoder_7_transformer_layers_2_layer_scale_2_scale, y = x_89)[name = tensor("op_1984")]; tensor input_139 = add(x = input_131, y = var_1984)[name = tensor("input_139")]; tensor x_91_axes_0 = const()[name = tensor("x_91_axes_0"), val = tensor([-1])]; tensor 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("x_91")]; tensor var_1999_shape = shape(x = x_91)[name = tensor("op_1999_shape")]; tensor gather_120_batch_dims_0 = const()[name = tensor("gather_120_batch_dims_0"), val = tensor(0)]; tensor gather_120_validate_indices_0 = const()[name = tensor("gather_120_validate_indices_0"), val = tensor(false)]; tensor select_38 = const()[name = tensor("select_38"), val = tensor(1)]; tensor gather_120_axis_1 = const()[name = tensor("gather_120_axis_1"), val = tensor(0)]; tensor 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("gather_120")]; tensor projected_45 = linear(bias = linear_1_bias_0, weight = encoder_7_transformer_layers_3_self_attn_in_proj_weight, x = x_91)[name = tensor("linear_51")]; tensor concat_132x = const()[name = tensor("concat_132x"), val = tensor([1, -1, 3, 4, 64])]; tensor var_2006 = reshape(shape = concat_132x, x = projected_45)[name = tensor("op_2006")]; tensor var_2007 = const()[name = tensor("op_2007"), val = tensor([2, 0, 3, 1, 4])]; tensor q_45_begin_0 = const()[name = tensor("q_45_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor q_45_end_0 = const()[name = tensor("q_45_end_0"), val = tensor([1, 1, 4, 0, 64])]; tensor q_45_end_mask_0 = const()[name = tensor("q_45_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor q_45_squeeze_mask_0 = const()[name = tensor("q_45_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor projected = transpose(perm = var_2007, x = var_2006)[name = tensor("transpose_50")]; tensor 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("q_45")]; tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([2, 1, 4, 0, 64])]; tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("k_45")]; tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([3, 1, 4, 0, 64])]; tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([false, true, true, true, true])]; tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; tensor 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("v")]; tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44433216)))]; tensor var_2019_shape = shape(x = q_45)[name = tensor("op_2019_shape")]; tensor gather_124_batch_dims_0 = const()[name = tensor("gather_124_batch_dims_0"), val = tensor(0)]; tensor gather_124_validate_indices_0 = const()[name = tensor("gather_124_validate_indices_0"), val = tensor(false)]; tensor select_39 = const()[name = tensor("select_39"), val = tensor(2)]; tensor gather_124_axis_1 = const()[name = tensor("gather_124_axis_1"), val = tensor(0)]; tensor 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("gather_124")]; tensor var_2022 = const()[name = tensor("op_2022"), val = tensor([[0x0p+0]])]; tensor const_102 = const()[name = tensor("const_102"), val = tensor(0)]; tensor const_103 = const()[name = tensor("const_103"), val = tensor(1)]; tensor var_2023 = range_1d(end = gather_124, start = const_102, step = const_103)[name = tensor("op_2023")]; tensor var_2023_promoted_dtype_0 = const()[name = tensor("op_2023_promoted_dtype_0"), val = tensor("fp32")]; tensor var_2023_promoted = cast(dtype = var_2023_promoted_dtype_0, x = var_2023)[name = tensor("cast_421")]; tensor ts_45 = add(x = var_2022, y = var_2023_promoted)[name = tensor("ts_45")]; tensor var_2026 = const()[name = tensor("op_2026"), val = tensor([1, 1, -1, 1])]; tensor ts = reshape(shape = var_2026, x = ts_45)[name = tensor("ts")]; tensor concat_133x = const()[name = tensor("concat_133x"), val = tensor([1, 4, -1, 32, 2])]; tensor q2 = reshape(shape = concat_133x, x = q_45)[name = tensor("q2")]; tensor concat_134x = const()[name = tensor("concat_134x"), val = tensor([1, 4, -1, 32, 2])]; tensor k2 = reshape(shape = concat_134x, x = k_45)[name = tensor("k2")]; tensor qr_begin_0 = const()[name = tensor("qr_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor qr_end_0 = const()[name = tensor("qr_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor qr_end_mask_0 = const()[name = tensor("qr_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qr_squeeze_mask_0 = const()[name = tensor("qr_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qr")]; tensor qi_begin_0 = const()[name = tensor("qi_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor qi_end_0 = const()[name = tensor("qi_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor qi_end_mask_0 = const()[name = tensor("qi_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor qi_squeeze_mask_0 = const()[name = tensor("qi_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("qi")]; tensor kr_begin_0 = const()[name = tensor("kr_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor kr_end_0 = const()[name = tensor("kr_end_0"), val = tensor([1, 4, 0, 32, 1])]; tensor kr_end_mask_0 = const()[name = tensor("kr_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor kr_squeeze_mask_0 = const()[name = tensor("kr_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("kr")]; tensor ki_begin_0 = const()[name = tensor("ki_begin_0"), val = tensor([0, 0, 0, 0, 1])]; tensor ki_end_0 = const()[name = tensor("ki_end_0"), val = tensor([1, 4, 0, 32, 2])]; tensor ki_end_mask_0 = const()[name = tensor("ki_end_mask_0"), val = tensor([true, true, true, true, false])]; tensor ki_squeeze_mask_0 = const()[name = tensor("ki_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; tensor 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("ki")]; tensor var_2039 = mul(x = freqs, y = ts)[name = tensor("op_2039")]; tensor rotr = cos(x = var_2039)[name = tensor("rotr")]; tensor roti = sin(x = var_2039)[name = tensor("roti")]; tensor var_2043 = mul(x = qr, y = rotr)[name = tensor("op_2043")]; tensor var_2044 = mul(x = qi, y = roti)[name = tensor("op_2044")]; tensor var_2045 = sub(x = var_2043, y = var_2044)[name = tensor("op_2045")]; tensor var_2046 = mul(x = qr, y = roti)[name = tensor("op_2046")]; tensor var_2047 = mul(x = qi, y = rotr)[name = tensor("op_2047")]; tensor var_2048 = add(x = var_2046, y = var_2047)[name = tensor("op_2048")]; tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; tensor qo = stack(axis = qo_axis_0, values = (var_2045, var_2048))[name = tensor("qo")]; tensor var_2051 = mul(x = kr, y = rotr)[name = tensor("op_2051")]; tensor var_2052 = mul(x = ki, y = roti)[name = tensor("op_2052")]; tensor var_2053 = sub(x = var_2051, y = var_2052)[name = tensor("op_2053")]; tensor var_2054 = mul(x = kr, y = roti)[name = tensor("op_2054")]; tensor var_2055 = mul(x = ki, y = rotr)[name = tensor("op_2055")]; tensor var_2056 = add(x = var_2054, y = var_2055)[name = tensor("op_2056")]; tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; tensor ko = stack(axis = ko_axis_0, values = (var_2053, var_2056))[name = tensor("ko")]; tensor concat_135x = const()[name = tensor("concat_135x"), val = tensor([1, 4, -1, 64])]; tensor q = reshape(shape = concat_135x, x = qo)[name = tensor("q")]; tensor concat_136x = const()[name = tensor("concat_136x"), val = tensor([1, 4, -1, 64])]; tensor k = reshape(shape = concat_136x, x = ko)[name = tensor("k")]; tensor const_104 = const()[name = tensor("const_104"), val = tensor(0)]; tensor const_105 = const()[name = tensor("const_105"), val = tensor(1)]; tensor positions = range_1d(end = gather_120, start = const_104, step = const_105)[name = tensor("positions")]; tensor concat_137x = const()[name = tensor("concat_137x"), val = tensor([1, 1, -1])]; tensor var_2068 = reshape(shape = concat_137x, x = positions)[name = tensor("op_2068")]; tensor var_2068_promoted_dtype_0 = const()[name = tensor("op_2068_promoted_dtype_0"), val = tensor("fp32")]; tensor var_2068_promoted = cast(dtype = var_2068_promoted_dtype_0, x = var_2068)[name = tensor("cast_420")]; tensor valid_k = less(x = var_2068_promoted, y = var_1659)[name = tensor("valid_k")]; tensor concat_138x = const()[name = tensor("concat_138x"), val = tensor([1, -1, 1])]; tensor var_2073 = reshape(shape = concat_138x, x = positions)[name = tensor("op_2073")]; tensor delta = sub(x = var_2073, y = var_2068)[name = tensor("delta")]; tensor concat_140_axis_0 = const()[name = tensor("concat_140_axis_0"), val = tensor(0)]; tensor concat_140_interleave_0 = const()[name = tensor("concat_140_interleave_0"), val = tensor(false)]; tensor concat_140 = concat(axis = concat_140_axis_0, interleave = concat_140_interleave_0, values = (var_1560, gather_120, gather_120))[name = tensor("concat_140")]; tensor fill_11_value_0 = const()[name = tensor("fill_11_value_0"), val = tensor(0x1p+0)]; tensor fill_11 = fill(shape = concat_140, value = fill_11_value_0)[name = tensor("fill_11")]; tensor attn_bias_89_dtype_0 = const()[name = tensor("attn_bias_89_dtype_0"), val = tensor("bool")]; tensor var_2079 = greater_equal(x = delta, y = var_1542)[name = tensor("op_2079")]; tensor attn_bias_89 = cast(dtype = attn_bias_89_dtype_0, x = fill_11)[name = tensor("cast_419")]; tensor attn_bias_91 = logical_and(x = attn_bias_89, y = var_2079)[name = tensor("attn_bias_91")]; tensor var_2081 = less(x = delta, y = var_1549)[name = tensor("op_2081")]; tensor attn_bias_93 = logical_and(x = attn_bias_91, y = var_2081)[name = tensor("attn_bias_93")]; tensor var_2083 = logical_and(x = attn_bias_93, y = valid_k)[name = tensor("op_2083")]; tensor var_2085_axes_0 = const()[name = tensor("op_2085_axes_0"), val = tensor([1])]; tensor var_2085 = expand_dims(axes = var_2085_axes_0, x = var_2083)[name = tensor("op_2085")]; tensor cast_268_dtype_0 = const()[name = tensor("cast_268_dtype_0"), val = tensor("fp32")]; tensor sub_11_x_0 = const()[name = tensor("sub_11_x_0"), val = tensor(0x1p+0)]; tensor cast_268 = cast(dtype = cast_268_dtype_0, x = var_2085)[name = tensor("cast_418")]; tensor sub_11 = sub(x = sub_11_x_0, y = cast_268)[name = tensor("sub_11")]; tensor mul_22_x_0 = const()[name = tensor("mul_22_x_0"), val = tensor(-0x1.d4cp+14)]; tensor mul_22 = mul(x = mul_22_x_0, y = sub_11)[name = tensor("mul_22")]; tensor mul_23_y_0 = const()[name = tensor("mul_23_y_0"), val = tensor(0x1p-3)]; tensor mul_23 = mul(x = q, y = mul_23_y_0)[name = tensor("mul_23")]; tensor matmul_11_transpose_y_0 = const()[name = tensor("matmul_11_transpose_y_0"), val = tensor(true)]; tensor matmul_11_transpose_x_0 = const()[name = tensor("matmul_11_transpose_x_0"), val = tensor(false)]; tensor matmul_11 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = mul_23, y = k)[name = tensor("matmul_11")]; tensor add_11 = add(x = matmul_11, y = mul_22)[name = tensor("add_11")]; tensor softmax_11_axis_0 = const()[name = tensor("softmax_11_axis_0"), val = tensor(-1)]; tensor softmax_11 = softmax(axis = softmax_11_axis_0, x = add_11)[name = tensor("softmax_11")]; tensor out_45_transpose_x_0 = const()[name = tensor("out_45_transpose_x_0"), val = tensor(false)]; tensor out_45_transpose_y_0 = const()[name = tensor("out_45_transpose_y_0"), val = tensor(false)]; tensor out_45 = matmul(transpose_x = out_45_transpose_x_0, transpose_y = out_45_transpose_y_0, x = softmax_11, y = v)[name = tensor("out_45")]; tensor concat_141x = const()[name = tensor("concat_141x"), val = tensor([1, -1])]; tensor var_2091 = reshape(shape = concat_141x, x = positions)[name = tensor("op_2091")]; tensor var_2091_promoted_dtype_0 = const()[name = tensor("op_2091_promoted_dtype_0"), val = tensor("fp32")]; tensor var_2091_promoted = cast(dtype = var_2091_promoted_dtype_0, x = var_2091)[name = tensor("cast_417")]; tensor var_2094 = less(x = var_2091_promoted, y = var_1682)[name = tensor("op_2094")]; tensor concat_142x = const()[name = tensor("concat_142x"), val = tensor([1, 1, -1, 1])]; tensor valid_q = reshape(shape = concat_142x, x = var_2094)[name = tensor("valid_q")]; tensor var_2098 = const()[name = tensor("op_2098"), val = tensor(0x0p+0)]; tensor out = select(a = out_45, b = var_2098, cond = valid_q)[name = tensor("out")]; tensor var_2100_perm_0 = const()[name = tensor("op_2100_perm_0"), val = tensor([0, 2, 1, 3])]; tensor concat_143x = const()[name = tensor("concat_143x"), val = tensor([1, -1, 256])]; tensor var_2100 = transpose(perm = var_2100_perm_0, x = out)[name = tensor("transpose_49")]; tensor input_141 = reshape(shape = concat_143x, x = var_2100)[name = tensor("input_141")]; tensor x_93 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_3_self_attn_out_proj_weight, x = input_141)[name = tensor("linear_52")]; tensor var_2106 = mul(x = encoder_7_transformer_layers_3_layer_scale_1_scale, y = x_93)[name = tensor("op_2106")]; tensor input_143 = add(x = input_139, y = var_2106)[name = tensor("input_143")]; tensor input_145_axes_0 = const()[name = tensor("input_145_axes_0"), val = tensor([-1])]; tensor 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("input_145")]; tensor input_147 = linear(bias = linear_3_bias_0, weight = encoder_7_transformer_layers_3_ffn_0_weight, x = input_145)[name = tensor("linear_53")]; tensor input_149_mode_0 = const()[name = tensor("input_149_mode_0"), val = tensor("EXACT")]; tensor input_149 = gelu(mode = input_149_mode_0, x = input_147)[name = tensor("input_149")]; tensor x_95 = linear(bias = linear_0_bias_0, weight = encoder_7_transformer_layers_3_ffn_2_weight, x = input_149)[name = tensor("linear_54")]; tensor var_2121 = mul(x = encoder_7_transformer_layers_3_layer_scale_2_scale, y = x_95)[name = tensor("op_2121")]; tensor input_151 = add(x = input_143, y = var_2121)[name = tensor("input_151")]; tensor linear_55_bias_0 = const()[name = tensor("linear_55_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44433408)))]; tensor var_2124 = linear(bias = linear_55_bias_0, weight = encoder_7_output_proj_weight, x = input_151)[name = tensor("linear_55")]; tensor x_perm_0 = const()[name = tensor("x_perm_0"), val = tensor([0, 2, 1])]; tensor var_2135 = const()[name = tensor("op_2135"), val = tensor([1, 192, -1, 4])]; tensor x = transpose(perm = x_perm_0, x = var_2124)[name = tensor("transpose_48")]; tensor var_2136 = reshape(shape = var_2135, x = x)[name = tensor("op_2136")]; tensor var_2137 = const()[name = tensor("op_2137"), val = tensor([0, 1, 3, 2])]; tensor var_2142 = const()[name = tensor("op_2142"), val = tensor([1, 768, -1])]; tensor var_2138 = transpose(perm = var_2137, x = var_2136)[name = tensor("transpose_47")]; tensor input_153 = reshape(shape = var_2142, x = var_2138)[name = tensor("input_153")]; tensor input_155_pad_type_0 = const()[name = tensor("input_155_pad_type_0"), val = tensor("valid")]; tensor input_155_strides_0 = const()[name = tensor("input_155_strides_0"), val = tensor([1])]; tensor input_155_pad_0 = const()[name = tensor("input_155_pad_0"), val = tensor([0, 0])]; tensor input_155_dilations_0 = const()[name = tensor("input_155_dilations_0"), val = tensor([1])]; tensor input_155_groups_0 = const()[name = tensor("input_155_groups_0"), val = tensor(1)]; tensor 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("input_155")]; tensor z_e_1_pad_type_0 = const()[name = tensor("z_e_1_pad_type_0"), val = tensor("valid")]; tensor z_e_1_strides_0 = const()[name = tensor("z_e_1_strides_0"), val = tensor([1])]; tensor z_e_1_pad_0 = const()[name = tensor("z_e_1_pad_0"), val = tensor([0, 0])]; tensor z_e_1_dilations_0 = const()[name = tensor("z_e_1_dilations_0"), val = tensor([1])]; tensor z_e_1_groups_0 = const()[name = tensor("z_e_1_groups_0"), val = tensor(1)]; tensor 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("z_e_1")]; tensor var_2168_perm_0 = const()[name = tensor("op_2168_perm_0"), val = tensor([0, 2, 1])]; tensor var_2174 = const()[name = tensor("op_2174"), val = tensor([-1, 8])]; tensor var_2168 = transpose(perm = var_2168_perm_0, x = z_e_1)[name = tensor("transpose_46")]; tensor input_157 = reshape(shape = var_2174, x = var_2168)[name = tensor("input_157")]; tensor var_2178 = const()[name = tensor("op_2178"), val = tensor([-1])]; tensor var_2179 = const()[name = tensor("op_2179"), val = tensor(true)]; tensor var_2181 = reduce_l2_norm(axes = var_2178, keep_dims = var_2179, x = input_157)[name = tensor("op_2181")]; tensor var_2182 = const()[name = tensor("op_2182"), val = tensor(0x1.197998p-40)]; tensor var_2183 = maximum(x = var_2181, y = var_2182)[name = tensor("op_2183")]; tensor shape_0 = shape(x = input_157)[name = tensor("shape_0")]; tensor shape_1 = shape(x = var_2183)[name = tensor("shape_1")]; tensor equal_0_y_0 = const()[name = tensor("equal_0_y_0"), val = tensor(-1)]; tensor equal_0 = equal(x = shape_0, y = equal_0_y_0)[name = tensor("equal_0")]; tensor select_0 = select(a = shape_1, b = shape_0, cond = equal_0)[name = tensor("select_0")]; tensor real_div_0 = real_div(x = select_0, y = shape_1)[name = tensor("real_div_0")]; tensor denom_1 = tile(reps = real_div_0, x = var_2183)[name = tensor("denom_1")]; tensor enc_1 = real_div(x = input_157, y = denom_1)[name = tensor("enc_1")]; tensor cb_1 = const()[name = tensor("cb_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44434240)))]; tensor sim_1 = linear(bias = linear_3_bias_0, weight = cb_1, x = enc_1)[name = tensor("sim_1")]; tensor input_161_axis_0 = const()[name = tensor("input_161_axis_0"), val = tensor(-1)]; tensor input_161_keep_dims_0 = const()[name = tensor("input_161_keep_dims_0"), val = tensor(false)]; tensor input_161_output_dtype_0 = const()[name = tensor("input_161_output_dtype_0"), val = tensor("int32")]; tensor 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("input_161")]; tensor var_2204_batch_dims_0 = const()[name = tensor("op_2204_batch_dims_0"), val = tensor(0)]; tensor var_2204_validate_indices_0 = const()[name = tensor("op_2204_validate_indices_0"), val = tensor(false)]; tensor greater_equal_24_y_0 = const()[name = tensor("greater_equal_24_y_0"), val = tensor(0)]; tensor greater_equal_24 = greater_equal(x = input_161, y = greater_equal_24_y_0)[name = tensor("greater_equal_24")]; tensor slice_by_index_24 = const()[name = tensor("slice_by_index_24"), val = tensor(1024)]; tensor add_36 = add(x = input_161, y = slice_by_index_24)[name = tensor("add_36")]; tensor select_40 = select(a = input_161, b = add_36, cond = greater_equal_24)[name = tensor("select_40")]; tensor var_2204_axis_1 = const()[name = tensor("op_2204_axis_1"), val = tensor(0)]; tensor 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("op_2204")]; tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, -1, 8])]; tensor var_2215 = reshape(shape = var_2214, x = var_2204)[name = tensor("op_2215")]; tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_1_pad_type_0 = const()[name = tensor("z_q_1_pad_type_0"), val = tensor("valid")]; tensor z_q_1_strides_0 = const()[name = tensor("z_q_1_strides_0"), val = tensor([1])]; tensor z_q_1_pad_0 = const()[name = tensor("z_q_1_pad_0"), val = tensor([0, 0])]; tensor z_q_1_dilations_0 = const()[name = tensor("z_q_1_dilations_0"), val = tensor([1])]; tensor z_q_1_groups_0 = const()[name = tensor("z_q_1_groups_0"), val = tensor(1)]; tensor input_163 = transpose(perm = input_163_perm_0, x = var_2215)[name = tensor("transpose_45")]; tensor 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("z_q_1")]; tensor input_165 = sub(x = input_155, y = z_q_1)[name = tensor("input_165")]; tensor var_2237 = const()[name = tensor("op_2237"), val = tensor([1, -1])]; tensor var_2238 = reshape(shape = var_2237, x = input_161)[name = tensor("op_2238")]; tensor z_e_3_pad_type_0 = const()[name = tensor("z_e_3_pad_type_0"), val = tensor("valid")]; tensor z_e_3_strides_0 = const()[name = tensor("z_e_3_strides_0"), val = tensor([1])]; tensor z_e_3_pad_0 = const()[name = tensor("z_e_3_pad_0"), val = tensor([0, 0])]; tensor z_e_3_dilations_0 = const()[name = tensor("z_e_3_dilations_0"), val = tensor([1])]; tensor z_e_3_groups_0 = const()[name = tensor("z_e_3_groups_0"), val = tensor(1)]; tensor 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("z_e_3")]; tensor var_2252_perm_0 = const()[name = tensor("op_2252_perm_0"), val = tensor([0, 2, 1])]; tensor var_2258 = const()[name = tensor("op_2258"), val = tensor([-1, 8])]; tensor var_2252 = transpose(perm = var_2252_perm_0, x = z_e_3)[name = tensor("transpose_44")]; tensor input_167 = reshape(shape = var_2258, x = var_2252)[name = tensor("input_167")]; tensor var_2262 = const()[name = tensor("op_2262"), val = tensor([-1])]; tensor var_2263 = const()[name = tensor("op_2263"), val = tensor(true)]; tensor var_2265 = reduce_l2_norm(axes = var_2262, keep_dims = var_2263, x = input_167)[name = tensor("op_2265")]; tensor var_2266 = const()[name = tensor("op_2266"), val = tensor(0x1.197998p-40)]; tensor var_2267 = maximum(x = var_2265, y = var_2266)[name = tensor("op_2267")]; tensor shape_3 = shape(x = input_167)[name = tensor("shape_3")]; tensor shape_4 = shape(x = var_2267)[name = tensor("shape_4")]; tensor equal_1_y_0 = const()[name = tensor("equal_1_y_0"), val = tensor(-1)]; tensor equal_1 = equal(x = shape_3, y = equal_1_y_0)[name = tensor("equal_1")]; tensor select_1 = select(a = shape_4, b = shape_3, cond = equal_1)[name = tensor("select_1")]; tensor real_div_1 = real_div(x = select_1, y = shape_4)[name = tensor("real_div_1")]; tensor denom_5 = tile(reps = real_div_1, x = var_2267)[name = tensor("denom_5")]; tensor enc_3 = real_div(x = input_167, y = denom_5)[name = tensor("enc_3")]; tensor cb_3 = const()[name = tensor("cb_3"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44467072)))]; tensor sim_3 = linear(bias = linear_3_bias_0, weight = cb_3, x = enc_3)[name = tensor("sim_3")]; tensor input_171_axis_0 = const()[name = tensor("input_171_axis_0"), val = tensor(-1)]; tensor input_171_keep_dims_0 = const()[name = tensor("input_171_keep_dims_0"), val = tensor(false)]; tensor input_171_output_dtype_0 = const()[name = tensor("input_171_output_dtype_0"), val = tensor("int32")]; tensor 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("input_171")]; tensor var_2288_batch_dims_0 = const()[name = tensor("op_2288_batch_dims_0"), val = tensor(0)]; tensor var_2288_validate_indices_0 = const()[name = tensor("op_2288_validate_indices_0"), val = tensor(false)]; tensor greater_equal_25_y_0 = const()[name = tensor("greater_equal_25_y_0"), val = tensor(0)]; tensor greater_equal_25 = greater_equal(x = input_171, y = greater_equal_25_y_0)[name = tensor("greater_equal_25")]; tensor slice_by_index_25 = const()[name = tensor("slice_by_index_25"), val = tensor(1024)]; tensor add_37 = add(x = input_171, y = slice_by_index_25)[name = tensor("add_37")]; tensor select_41 = select(a = input_171, b = add_37, cond = greater_equal_25)[name = tensor("select_41")]; tensor var_2288_axis_1 = const()[name = tensor("op_2288_axis_1"), val = tensor(0)]; tensor 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("op_2288")]; tensor var_2298 = const()[name = tensor("op_2298"), val = tensor([1, -1, 8])]; tensor var_2299 = reshape(shape = var_2298, x = var_2288)[name = tensor("op_2299")]; tensor input_173_perm_0 = const()[name = tensor("input_173_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_3_pad_type_0 = const()[name = tensor("z_q_3_pad_type_0"), val = tensor("valid")]; tensor z_q_3_strides_0 = const()[name = tensor("z_q_3_strides_0"), val = tensor([1])]; tensor z_q_3_pad_0 = const()[name = tensor("z_q_3_pad_0"), val = tensor([0, 0])]; tensor z_q_3_dilations_0 = const()[name = tensor("z_q_3_dilations_0"), val = tensor([1])]; tensor z_q_3_groups_0 = const()[name = tensor("z_q_3_groups_0"), val = tensor(1)]; tensor input_173 = transpose(perm = input_173_perm_0, x = var_2299)[name = tensor("transpose_43")]; tensor 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("z_q_3")]; tensor input_175 = sub(x = input_165, y = z_q_3)[name = tensor("input_175")]; tensor var_2321 = const()[name = tensor("op_2321"), val = tensor([1, -1])]; tensor var_2322 = reshape(shape = var_2321, x = input_171)[name = tensor("op_2322")]; tensor z_e_5_pad_type_0 = const()[name = tensor("z_e_5_pad_type_0"), val = tensor("valid")]; tensor z_e_5_strides_0 = const()[name = tensor("z_e_5_strides_0"), val = tensor([1])]; tensor z_e_5_pad_0 = const()[name = tensor("z_e_5_pad_0"), val = tensor([0, 0])]; tensor z_e_5_dilations_0 = const()[name = tensor("z_e_5_dilations_0"), val = tensor([1])]; tensor z_e_5_groups_0 = const()[name = tensor("z_e_5_groups_0"), val = tensor(1)]; tensor 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("z_e_5")]; tensor var_2336_perm_0 = const()[name = tensor("op_2336_perm_0"), val = tensor([0, 2, 1])]; tensor var_2342 = const()[name = tensor("op_2342"), val = tensor([-1, 8])]; tensor var_2336 = transpose(perm = var_2336_perm_0, x = z_e_5)[name = tensor("transpose_42")]; tensor input_177 = reshape(shape = var_2342, x = var_2336)[name = tensor("input_177")]; tensor var_2346 = const()[name = tensor("op_2346"), val = tensor([-1])]; tensor var_2347 = const()[name = tensor("op_2347"), val = tensor(true)]; tensor var_2349 = reduce_l2_norm(axes = var_2346, keep_dims = var_2347, x = input_177)[name = tensor("op_2349")]; tensor var_2350 = const()[name = tensor("op_2350"), val = tensor(0x1.197998p-40)]; tensor var_2351 = maximum(x = var_2349, y = var_2350)[name = tensor("op_2351")]; tensor shape_6 = shape(x = input_177)[name = tensor("shape_6")]; tensor shape_7 = shape(x = var_2351)[name = tensor("shape_7")]; tensor equal_2_y_0 = const()[name = tensor("equal_2_y_0"), val = tensor(-1)]; tensor equal_2 = equal(x = shape_6, y = equal_2_y_0)[name = tensor("equal_2")]; tensor select_2 = select(a = shape_7, b = shape_6, cond = equal_2)[name = tensor("select_2")]; tensor real_div_2 = real_div(x = select_2, y = shape_7)[name = tensor("real_div_2")]; tensor denom_9 = tile(reps = real_div_2, x = var_2351)[name = tensor("denom_9")]; tensor enc_5 = real_div(x = input_177, y = denom_9)[name = tensor("enc_5")]; tensor cb_5 = const()[name = tensor("cb_5"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44499904)))]; tensor sim_5 = linear(bias = linear_3_bias_0, weight = cb_5, x = enc_5)[name = tensor("sim_5")]; tensor input_181_axis_0 = const()[name = tensor("input_181_axis_0"), val = tensor(-1)]; tensor input_181_keep_dims_0 = const()[name = tensor("input_181_keep_dims_0"), val = tensor(false)]; tensor input_181_output_dtype_0 = const()[name = tensor("input_181_output_dtype_0"), val = tensor("int32")]; tensor 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("input_181")]; tensor var_2372_batch_dims_0 = const()[name = tensor("op_2372_batch_dims_0"), val = tensor(0)]; tensor var_2372_validate_indices_0 = const()[name = tensor("op_2372_validate_indices_0"), val = tensor(false)]; tensor greater_equal_26_y_0 = const()[name = tensor("greater_equal_26_y_0"), val = tensor(0)]; tensor greater_equal_26 = greater_equal(x = input_181, y = greater_equal_26_y_0)[name = tensor("greater_equal_26")]; tensor slice_by_index_26 = const()[name = tensor("slice_by_index_26"), val = tensor(1024)]; tensor add_38 = add(x = input_181, y = slice_by_index_26)[name = tensor("add_38")]; tensor select_42 = select(a = input_181, b = add_38, cond = greater_equal_26)[name = tensor("select_42")]; tensor var_2372_axis_1 = const()[name = tensor("op_2372_axis_1"), val = tensor(0)]; tensor 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("op_2372")]; tensor var_2382 = const()[name = tensor("op_2382"), val = tensor([1, -1, 8])]; tensor var_2383 = reshape(shape = var_2382, x = var_2372)[name = tensor("op_2383")]; tensor input_183_perm_0 = const()[name = tensor("input_183_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_5_pad_type_0 = const()[name = tensor("z_q_5_pad_type_0"), val = tensor("valid")]; tensor z_q_5_strides_0 = const()[name = tensor("z_q_5_strides_0"), val = tensor([1])]; tensor z_q_5_pad_0 = const()[name = tensor("z_q_5_pad_0"), val = tensor([0, 0])]; tensor z_q_5_dilations_0 = const()[name = tensor("z_q_5_dilations_0"), val = tensor([1])]; tensor z_q_5_groups_0 = const()[name = tensor("z_q_5_groups_0"), val = tensor(1)]; tensor input_183 = transpose(perm = input_183_perm_0, x = var_2383)[name = tensor("transpose_41")]; tensor 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("z_q_5")]; tensor input_185 = sub(x = input_175, y = z_q_5)[name = tensor("input_185")]; tensor var_2405 = const()[name = tensor("op_2405"), val = tensor([1, -1])]; tensor var_2406 = reshape(shape = var_2405, x = input_181)[name = tensor("op_2406")]; tensor z_e_7_pad_type_0 = const()[name = tensor("z_e_7_pad_type_0"), val = tensor("valid")]; tensor z_e_7_strides_0 = const()[name = tensor("z_e_7_strides_0"), val = tensor([1])]; tensor z_e_7_pad_0 = const()[name = tensor("z_e_7_pad_0"), val = tensor([0, 0])]; tensor z_e_7_dilations_0 = const()[name = tensor("z_e_7_dilations_0"), val = tensor([1])]; tensor z_e_7_groups_0 = const()[name = tensor("z_e_7_groups_0"), val = tensor(1)]; tensor 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("z_e_7")]; tensor var_2420_perm_0 = const()[name = tensor("op_2420_perm_0"), val = tensor([0, 2, 1])]; tensor var_2426 = const()[name = tensor("op_2426"), val = tensor([-1, 8])]; tensor var_2420 = transpose(perm = var_2420_perm_0, x = z_e_7)[name = tensor("transpose_40")]; tensor input_187 = reshape(shape = var_2426, x = var_2420)[name = tensor("input_187")]; tensor var_2430 = const()[name = tensor("op_2430"), val = tensor([-1])]; tensor var_2431 = const()[name = tensor("op_2431"), val = tensor(true)]; tensor var_2433 = reduce_l2_norm(axes = var_2430, keep_dims = var_2431, x = input_187)[name = tensor("op_2433")]; tensor var_2434 = const()[name = tensor("op_2434"), val = tensor(0x1.197998p-40)]; tensor var_2435 = maximum(x = var_2433, y = var_2434)[name = tensor("op_2435")]; tensor shape_9 = shape(x = input_187)[name = tensor("shape_9")]; tensor shape_10 = shape(x = var_2435)[name = tensor("shape_10")]; tensor equal_3_y_0 = const()[name = tensor("equal_3_y_0"), val = tensor(-1)]; tensor equal_3 = equal(x = shape_9, y = equal_3_y_0)[name = tensor("equal_3")]; tensor select_3 = select(a = shape_10, b = shape_9, cond = equal_3)[name = tensor("select_3")]; tensor real_div_3 = real_div(x = select_3, y = shape_10)[name = tensor("real_div_3")]; tensor denom_13 = tile(reps = real_div_3, x = var_2435)[name = tensor("denom_13")]; tensor enc_7 = real_div(x = input_187, y = denom_13)[name = tensor("enc_7")]; tensor cb_7 = const()[name = tensor("cb_7"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44532736)))]; tensor sim_7 = linear(bias = linear_3_bias_0, weight = cb_7, x = enc_7)[name = tensor("sim_7")]; tensor input_191_axis_0 = const()[name = tensor("input_191_axis_0"), val = tensor(-1)]; tensor input_191_keep_dims_0 = const()[name = tensor("input_191_keep_dims_0"), val = tensor(false)]; tensor input_191_output_dtype_0 = const()[name = tensor("input_191_output_dtype_0"), val = tensor("int32")]; tensor 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("input_191")]; tensor var_2456_batch_dims_0 = const()[name = tensor("op_2456_batch_dims_0"), val = tensor(0)]; tensor var_2456_validate_indices_0 = const()[name = tensor("op_2456_validate_indices_0"), val = tensor(false)]; tensor greater_equal_27_y_0 = const()[name = tensor("greater_equal_27_y_0"), val = tensor(0)]; tensor greater_equal_27 = greater_equal(x = input_191, y = greater_equal_27_y_0)[name = tensor("greater_equal_27")]; tensor slice_by_index_27 = const()[name = tensor("slice_by_index_27"), val = tensor(1024)]; tensor add_39 = add(x = input_191, y = slice_by_index_27)[name = tensor("add_39")]; tensor select_43 = select(a = input_191, b = add_39, cond = greater_equal_27)[name = tensor("select_43")]; tensor var_2456_axis_1 = const()[name = tensor("op_2456_axis_1"), val = tensor(0)]; tensor 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("op_2456")]; tensor var_2466 = const()[name = tensor("op_2466"), val = tensor([1, -1, 8])]; tensor var_2467 = reshape(shape = var_2466, x = var_2456)[name = tensor("op_2467")]; tensor input_193_perm_0 = const()[name = tensor("input_193_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_7_pad_type_0 = const()[name = tensor("z_q_7_pad_type_0"), val = tensor("valid")]; tensor z_q_7_strides_0 = const()[name = tensor("z_q_7_strides_0"), val = tensor([1])]; tensor z_q_7_pad_0 = const()[name = tensor("z_q_7_pad_0"), val = tensor([0, 0])]; tensor z_q_7_dilations_0 = const()[name = tensor("z_q_7_dilations_0"), val = tensor([1])]; tensor z_q_7_groups_0 = const()[name = tensor("z_q_7_groups_0"), val = tensor(1)]; tensor input_193 = transpose(perm = input_193_perm_0, x = var_2467)[name = tensor("transpose_39")]; tensor 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("z_q_7")]; tensor input_195 = sub(x = input_185, y = z_q_7)[name = tensor("input_195")]; tensor var_2489 = const()[name = tensor("op_2489"), val = tensor([1, -1])]; tensor var_2490 = reshape(shape = var_2489, x = input_191)[name = tensor("op_2490")]; tensor z_e_9_pad_type_0 = const()[name = tensor("z_e_9_pad_type_0"), val = tensor("valid")]; tensor z_e_9_strides_0 = const()[name = tensor("z_e_9_strides_0"), val = tensor([1])]; tensor z_e_9_pad_0 = const()[name = tensor("z_e_9_pad_0"), val = tensor([0, 0])]; tensor z_e_9_dilations_0 = const()[name = tensor("z_e_9_dilations_0"), val = tensor([1])]; tensor z_e_9_groups_0 = const()[name = tensor("z_e_9_groups_0"), val = tensor(1)]; tensor 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("z_e_9")]; tensor var_2504_perm_0 = const()[name = tensor("op_2504_perm_0"), val = tensor([0, 2, 1])]; tensor var_2510 = const()[name = tensor("op_2510"), val = tensor([-1, 8])]; tensor var_2504 = transpose(perm = var_2504_perm_0, x = z_e_9)[name = tensor("transpose_38")]; tensor input_197 = reshape(shape = var_2510, x = var_2504)[name = tensor("input_197")]; tensor var_2514 = const()[name = tensor("op_2514"), val = tensor([-1])]; tensor var_2515 = const()[name = tensor("op_2515"), val = tensor(true)]; tensor var_2517 = reduce_l2_norm(axes = var_2514, keep_dims = var_2515, x = input_197)[name = tensor("op_2517")]; tensor var_2518 = const()[name = tensor("op_2518"), val = tensor(0x1.197998p-40)]; tensor var_2519 = maximum(x = var_2517, y = var_2518)[name = tensor("op_2519")]; tensor shape_12 = shape(x = input_197)[name = tensor("shape_12")]; tensor shape_13 = shape(x = var_2519)[name = tensor("shape_13")]; tensor equal_4_y_0 = const()[name = tensor("equal_4_y_0"), val = tensor(-1)]; tensor equal_4 = equal(x = shape_12, y = equal_4_y_0)[name = tensor("equal_4")]; tensor select_4 = select(a = shape_13, b = shape_12, cond = equal_4)[name = tensor("select_4")]; tensor real_div_4 = real_div(x = select_4, y = shape_13)[name = tensor("real_div_4")]; tensor denom_17 = tile(reps = real_div_4, x = var_2519)[name = tensor("denom_17")]; tensor enc_9 = real_div(x = input_197, y = denom_17)[name = tensor("enc_9")]; tensor cb_9 = const()[name = tensor("cb_9"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44565568)))]; tensor sim_9 = linear(bias = linear_3_bias_0, weight = cb_9, x = enc_9)[name = tensor("sim_9")]; tensor input_201_axis_0 = const()[name = tensor("input_201_axis_0"), val = tensor(-1)]; tensor input_201_keep_dims_0 = const()[name = tensor("input_201_keep_dims_0"), val = tensor(false)]; tensor input_201_output_dtype_0 = const()[name = tensor("input_201_output_dtype_0"), val = tensor("int32")]; tensor 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("input_201")]; tensor var_2540_batch_dims_0 = const()[name = tensor("op_2540_batch_dims_0"), val = tensor(0)]; tensor var_2540_validate_indices_0 = const()[name = tensor("op_2540_validate_indices_0"), val = tensor(false)]; tensor greater_equal_28_y_0 = const()[name = tensor("greater_equal_28_y_0"), val = tensor(0)]; tensor greater_equal_28 = greater_equal(x = input_201, y = greater_equal_28_y_0)[name = tensor("greater_equal_28")]; tensor slice_by_index_28 = const()[name = tensor("slice_by_index_28"), val = tensor(1024)]; tensor add_40 = add(x = input_201, y = slice_by_index_28)[name = tensor("add_40")]; tensor select_44 = select(a = input_201, b = add_40, cond = greater_equal_28)[name = tensor("select_44")]; tensor var_2540_axis_1 = const()[name = tensor("op_2540_axis_1"), val = tensor(0)]; tensor 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("op_2540")]; tensor var_2550 = const()[name = tensor("op_2550"), val = tensor([1, -1, 8])]; tensor var_2551 = reshape(shape = var_2550, x = var_2540)[name = tensor("op_2551")]; tensor input_203_perm_0 = const()[name = tensor("input_203_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_9_pad_type_0 = const()[name = tensor("z_q_9_pad_type_0"), val = tensor("valid")]; tensor z_q_9_strides_0 = const()[name = tensor("z_q_9_strides_0"), val = tensor([1])]; tensor z_q_9_pad_0 = const()[name = tensor("z_q_9_pad_0"), val = tensor([0, 0])]; tensor z_q_9_dilations_0 = const()[name = tensor("z_q_9_dilations_0"), val = tensor([1])]; tensor z_q_9_groups_0 = const()[name = tensor("z_q_9_groups_0"), val = tensor(1)]; tensor input_203 = transpose(perm = input_203_perm_0, x = var_2551)[name = tensor("transpose_37")]; tensor 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("z_q_9")]; tensor input_205 = sub(x = input_195, y = z_q_9)[name = tensor("input_205")]; tensor var_2573 = const()[name = tensor("op_2573"), val = tensor([1, -1])]; tensor var_2574 = reshape(shape = var_2573, x = input_201)[name = tensor("op_2574")]; tensor z_e_11_pad_type_0 = const()[name = tensor("z_e_11_pad_type_0"), val = tensor("valid")]; tensor z_e_11_strides_0 = const()[name = tensor("z_e_11_strides_0"), val = tensor([1])]; tensor z_e_11_pad_0 = const()[name = tensor("z_e_11_pad_0"), val = tensor([0, 0])]; tensor z_e_11_dilations_0 = const()[name = tensor("z_e_11_dilations_0"), val = tensor([1])]; tensor z_e_11_groups_0 = const()[name = tensor("z_e_11_groups_0"), val = tensor(1)]; tensor 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("z_e_11")]; tensor var_2588_perm_0 = const()[name = tensor("op_2588_perm_0"), val = tensor([0, 2, 1])]; tensor var_2594 = const()[name = tensor("op_2594"), val = tensor([-1, 8])]; tensor var_2588 = transpose(perm = var_2588_perm_0, x = z_e_11)[name = tensor("transpose_36")]; tensor input_207 = reshape(shape = var_2594, x = var_2588)[name = tensor("input_207")]; tensor var_2598 = const()[name = tensor("op_2598"), val = tensor([-1])]; tensor var_2599 = const()[name = tensor("op_2599"), val = tensor(true)]; tensor var_2601 = reduce_l2_norm(axes = var_2598, keep_dims = var_2599, x = input_207)[name = tensor("op_2601")]; tensor var_2602 = const()[name = tensor("op_2602"), val = tensor(0x1.197998p-40)]; tensor var_2603 = maximum(x = var_2601, y = var_2602)[name = tensor("op_2603")]; tensor shape_15 = shape(x = input_207)[name = tensor("shape_15")]; tensor shape_16 = shape(x = var_2603)[name = tensor("shape_16")]; tensor equal_5_y_0 = const()[name = tensor("equal_5_y_0"), val = tensor(-1)]; tensor equal_5 = equal(x = shape_15, y = equal_5_y_0)[name = tensor("equal_5")]; tensor select_5 = select(a = shape_16, b = shape_15, cond = equal_5)[name = tensor("select_5")]; tensor real_div_5 = real_div(x = select_5, y = shape_16)[name = tensor("real_div_5")]; tensor denom_21 = tile(reps = real_div_5, x = var_2603)[name = tensor("denom_21")]; tensor enc_11 = real_div(x = input_207, y = denom_21)[name = tensor("enc_11")]; tensor cb_11 = const()[name = tensor("cb_11"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44598400)))]; tensor sim_11 = linear(bias = linear_3_bias_0, weight = cb_11, x = enc_11)[name = tensor("sim_11")]; tensor input_211_axis_0 = const()[name = tensor("input_211_axis_0"), val = tensor(-1)]; tensor input_211_keep_dims_0 = const()[name = tensor("input_211_keep_dims_0"), val = tensor(false)]; tensor input_211_output_dtype_0 = const()[name = tensor("input_211_output_dtype_0"), val = tensor("int32")]; tensor 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("input_211")]; tensor var_2624_batch_dims_0 = const()[name = tensor("op_2624_batch_dims_0"), val = tensor(0)]; tensor var_2624_validate_indices_0 = const()[name = tensor("op_2624_validate_indices_0"), val = tensor(false)]; tensor greater_equal_29_y_0 = const()[name = tensor("greater_equal_29_y_0"), val = tensor(0)]; tensor greater_equal_29 = greater_equal(x = input_211, y = greater_equal_29_y_0)[name = tensor("greater_equal_29")]; tensor slice_by_index_29 = const()[name = tensor("slice_by_index_29"), val = tensor(1024)]; tensor add_41 = add(x = input_211, y = slice_by_index_29)[name = tensor("add_41")]; tensor select_45 = select(a = input_211, b = add_41, cond = greater_equal_29)[name = tensor("select_45")]; tensor var_2624_axis_1 = const()[name = tensor("op_2624_axis_1"), val = tensor(0)]; tensor 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("op_2624")]; tensor var_2634 = const()[name = tensor("op_2634"), val = tensor([1, -1, 8])]; tensor var_2635 = reshape(shape = var_2634, x = var_2624)[name = tensor("op_2635")]; tensor input_213_perm_0 = const()[name = tensor("input_213_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_11_pad_type_0 = const()[name = tensor("z_q_11_pad_type_0"), val = tensor("valid")]; tensor z_q_11_strides_0 = const()[name = tensor("z_q_11_strides_0"), val = tensor([1])]; tensor z_q_11_pad_0 = const()[name = tensor("z_q_11_pad_0"), val = tensor([0, 0])]; tensor z_q_11_dilations_0 = const()[name = tensor("z_q_11_dilations_0"), val = tensor([1])]; tensor z_q_11_groups_0 = const()[name = tensor("z_q_11_groups_0"), val = tensor(1)]; tensor input_213 = transpose(perm = input_213_perm_0, x = var_2635)[name = tensor("transpose_35")]; tensor 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("z_q_11")]; tensor input_215 = sub(x = input_205, y = z_q_11)[name = tensor("input_215")]; tensor var_2657 = const()[name = tensor("op_2657"), val = tensor([1, -1])]; tensor var_2658 = reshape(shape = var_2657, x = input_211)[name = tensor("op_2658")]; tensor z_e_13_pad_type_0 = const()[name = tensor("z_e_13_pad_type_0"), val = tensor("valid")]; tensor z_e_13_strides_0 = const()[name = tensor("z_e_13_strides_0"), val = tensor([1])]; tensor z_e_13_pad_0 = const()[name = tensor("z_e_13_pad_0"), val = tensor([0, 0])]; tensor z_e_13_dilations_0 = const()[name = tensor("z_e_13_dilations_0"), val = tensor([1])]; tensor z_e_13_groups_0 = const()[name = tensor("z_e_13_groups_0"), val = tensor(1)]; tensor 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("z_e_13")]; tensor var_2672_perm_0 = const()[name = tensor("op_2672_perm_0"), val = tensor([0, 2, 1])]; tensor var_2678 = const()[name = tensor("op_2678"), val = tensor([-1, 8])]; tensor var_2672 = transpose(perm = var_2672_perm_0, x = z_e_13)[name = tensor("transpose_34")]; tensor input_217 = reshape(shape = var_2678, x = var_2672)[name = tensor("input_217")]; tensor var_2682 = const()[name = tensor("op_2682"), val = tensor([-1])]; tensor var_2683 = const()[name = tensor("op_2683"), val = tensor(true)]; tensor var_2685 = reduce_l2_norm(axes = var_2682, keep_dims = var_2683, x = input_217)[name = tensor("op_2685")]; tensor var_2686 = const()[name = tensor("op_2686"), val = tensor(0x1.197998p-40)]; tensor var_2687 = maximum(x = var_2685, y = var_2686)[name = tensor("op_2687")]; tensor shape_18 = shape(x = input_217)[name = tensor("shape_18")]; tensor shape_19 = shape(x = var_2687)[name = tensor("shape_19")]; tensor equal_6_y_0 = const()[name = tensor("equal_6_y_0"), val = tensor(-1)]; tensor equal_6 = equal(x = shape_18, y = equal_6_y_0)[name = tensor("equal_6")]; tensor select_6 = select(a = shape_19, b = shape_18, cond = equal_6)[name = tensor("select_6")]; tensor real_div_6 = real_div(x = select_6, y = shape_19)[name = tensor("real_div_6")]; tensor denom_25 = tile(reps = real_div_6, x = var_2687)[name = tensor("denom_25")]; tensor enc_13 = real_div(x = input_217, y = denom_25)[name = tensor("enc_13")]; tensor cb_13 = const()[name = tensor("cb_13"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44631232)))]; tensor sim_13 = linear(bias = linear_3_bias_0, weight = cb_13, x = enc_13)[name = tensor("sim_13")]; tensor input_221_axis_0 = const()[name = tensor("input_221_axis_0"), val = tensor(-1)]; tensor input_221_keep_dims_0 = const()[name = tensor("input_221_keep_dims_0"), val = tensor(false)]; tensor input_221_output_dtype_0 = const()[name = tensor("input_221_output_dtype_0"), val = tensor("int32")]; tensor 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("input_221")]; tensor var_2708_batch_dims_0 = const()[name = tensor("op_2708_batch_dims_0"), val = tensor(0)]; tensor var_2708_validate_indices_0 = const()[name = tensor("op_2708_validate_indices_0"), val = tensor(false)]; tensor greater_equal_30_y_0 = const()[name = tensor("greater_equal_30_y_0"), val = tensor(0)]; tensor greater_equal_30 = greater_equal(x = input_221, y = greater_equal_30_y_0)[name = tensor("greater_equal_30")]; tensor slice_by_index_30 = const()[name = tensor("slice_by_index_30"), val = tensor(1024)]; tensor add_42 = add(x = input_221, y = slice_by_index_30)[name = tensor("add_42")]; tensor select_46 = select(a = input_221, b = add_42, cond = greater_equal_30)[name = tensor("select_46")]; tensor var_2708_axis_1 = const()[name = tensor("op_2708_axis_1"), val = tensor(0)]; tensor 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("op_2708")]; tensor var_2718 = const()[name = tensor("op_2718"), val = tensor([1, -1, 8])]; tensor var_2719 = reshape(shape = var_2718, x = var_2708)[name = tensor("op_2719")]; tensor input_223_perm_0 = const()[name = tensor("input_223_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_13_pad_type_0 = const()[name = tensor("z_q_13_pad_type_0"), val = tensor("valid")]; tensor z_q_13_strides_0 = const()[name = tensor("z_q_13_strides_0"), val = tensor([1])]; tensor z_q_13_pad_0 = const()[name = tensor("z_q_13_pad_0"), val = tensor([0, 0])]; tensor z_q_13_dilations_0 = const()[name = tensor("z_q_13_dilations_0"), val = tensor([1])]; tensor z_q_13_groups_0 = const()[name = tensor("z_q_13_groups_0"), val = tensor(1)]; tensor input_223 = transpose(perm = input_223_perm_0, x = var_2719)[name = tensor("transpose_33")]; tensor 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("z_q_13")]; tensor input_225 = sub(x = input_215, y = z_q_13)[name = tensor("input_225")]; tensor var_2741 = const()[name = tensor("op_2741"), val = tensor([1, -1])]; tensor var_2742 = reshape(shape = var_2741, x = input_221)[name = tensor("op_2742")]; tensor z_e_15_pad_type_0 = const()[name = tensor("z_e_15_pad_type_0"), val = tensor("valid")]; tensor z_e_15_strides_0 = const()[name = tensor("z_e_15_strides_0"), val = tensor([1])]; tensor z_e_15_pad_0 = const()[name = tensor("z_e_15_pad_0"), val = tensor([0, 0])]; tensor z_e_15_dilations_0 = const()[name = tensor("z_e_15_dilations_0"), val = tensor([1])]; tensor z_e_15_groups_0 = const()[name = tensor("z_e_15_groups_0"), val = tensor(1)]; tensor 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("z_e_15")]; tensor var_2756_perm_0 = const()[name = tensor("op_2756_perm_0"), val = tensor([0, 2, 1])]; tensor var_2762 = const()[name = tensor("op_2762"), val = tensor([-1, 8])]; tensor var_2756 = transpose(perm = var_2756_perm_0, x = z_e_15)[name = tensor("transpose_32")]; tensor input_227 = reshape(shape = var_2762, x = var_2756)[name = tensor("input_227")]; tensor var_2766 = const()[name = tensor("op_2766"), val = tensor([-1])]; tensor var_2767 = const()[name = tensor("op_2767"), val = tensor(true)]; tensor var_2769 = reduce_l2_norm(axes = var_2766, keep_dims = var_2767, x = input_227)[name = tensor("op_2769")]; tensor var_2770 = const()[name = tensor("op_2770"), val = tensor(0x1.197998p-40)]; tensor var_2771 = maximum(x = var_2769, y = var_2770)[name = tensor("op_2771")]; tensor shape_21 = shape(x = input_227)[name = tensor("shape_21")]; tensor shape_22 = shape(x = var_2771)[name = tensor("shape_22")]; tensor equal_7_y_0 = const()[name = tensor("equal_7_y_0"), val = tensor(-1)]; tensor equal_7 = equal(x = shape_21, y = equal_7_y_0)[name = tensor("equal_7")]; tensor select_7 = select(a = shape_22, b = shape_21, cond = equal_7)[name = tensor("select_7")]; tensor real_div_7 = real_div(x = select_7, y = shape_22)[name = tensor("real_div_7")]; tensor denom_29 = tile(reps = real_div_7, x = var_2771)[name = tensor("denom_29")]; tensor enc_15 = real_div(x = input_227, y = denom_29)[name = tensor("enc_15")]; tensor cb_15 = const()[name = tensor("cb_15"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44664064)))]; tensor sim_15 = linear(bias = linear_3_bias_0, weight = cb_15, x = enc_15)[name = tensor("sim_15")]; tensor input_231_axis_0 = const()[name = tensor("input_231_axis_0"), val = tensor(-1)]; tensor input_231_keep_dims_0 = const()[name = tensor("input_231_keep_dims_0"), val = tensor(false)]; tensor input_231_output_dtype_0 = const()[name = tensor("input_231_output_dtype_0"), val = tensor("int32")]; tensor 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("input_231")]; tensor var_2792_batch_dims_0 = const()[name = tensor("op_2792_batch_dims_0"), val = tensor(0)]; tensor var_2792_validate_indices_0 = const()[name = tensor("op_2792_validate_indices_0"), val = tensor(false)]; tensor greater_equal_31_y_0 = const()[name = tensor("greater_equal_31_y_0"), val = tensor(0)]; tensor greater_equal_31 = greater_equal(x = input_231, y = greater_equal_31_y_0)[name = tensor("greater_equal_31")]; tensor slice_by_index_31 = const()[name = tensor("slice_by_index_31"), val = tensor(1024)]; tensor add_43 = add(x = input_231, y = slice_by_index_31)[name = tensor("add_43")]; tensor select_47 = select(a = input_231, b = add_43, cond = greater_equal_31)[name = tensor("select_47")]; tensor var_2792_axis_1 = const()[name = tensor("op_2792_axis_1"), val = tensor(0)]; tensor 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("op_2792")]; tensor var_2802 = const()[name = tensor("op_2802"), val = tensor([1, -1, 8])]; tensor var_2803 = reshape(shape = var_2802, x = var_2792)[name = tensor("op_2803")]; tensor input_233_perm_0 = const()[name = tensor("input_233_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_15_pad_type_0 = const()[name = tensor("z_q_15_pad_type_0"), val = tensor("valid")]; tensor z_q_15_strides_0 = const()[name = tensor("z_q_15_strides_0"), val = tensor([1])]; tensor z_q_15_pad_0 = const()[name = tensor("z_q_15_pad_0"), val = tensor([0, 0])]; tensor z_q_15_dilations_0 = const()[name = tensor("z_q_15_dilations_0"), val = tensor([1])]; tensor z_q_15_groups_0 = const()[name = tensor("z_q_15_groups_0"), val = tensor(1)]; tensor input_233 = transpose(perm = input_233_perm_0, x = var_2803)[name = tensor("transpose_31")]; tensor 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("z_q_15")]; tensor input_235 = sub(x = input_225, y = z_q_15)[name = tensor("input_235")]; tensor var_2825 = const()[name = tensor("op_2825"), val = tensor([1, -1])]; tensor var_2826 = reshape(shape = var_2825, x = input_231)[name = tensor("op_2826")]; tensor z_e_17_pad_type_0 = const()[name = tensor("z_e_17_pad_type_0"), val = tensor("valid")]; tensor z_e_17_strides_0 = const()[name = tensor("z_e_17_strides_0"), val = tensor([1])]; tensor z_e_17_pad_0 = const()[name = tensor("z_e_17_pad_0"), val = tensor([0, 0])]; tensor z_e_17_dilations_0 = const()[name = tensor("z_e_17_dilations_0"), val = tensor([1])]; tensor z_e_17_groups_0 = const()[name = tensor("z_e_17_groups_0"), val = tensor(1)]; tensor 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("z_e_17")]; tensor var_2840_perm_0 = const()[name = tensor("op_2840_perm_0"), val = tensor([0, 2, 1])]; tensor var_2846 = const()[name = tensor("op_2846"), val = tensor([-1, 8])]; tensor var_2840 = transpose(perm = var_2840_perm_0, x = z_e_17)[name = tensor("transpose_30")]; tensor input_237 = reshape(shape = var_2846, x = var_2840)[name = tensor("input_237")]; tensor var_2850 = const()[name = tensor("op_2850"), val = tensor([-1])]; tensor var_2851 = const()[name = tensor("op_2851"), val = tensor(true)]; tensor var_2853 = reduce_l2_norm(axes = var_2850, keep_dims = var_2851, x = input_237)[name = tensor("op_2853")]; tensor var_2854 = const()[name = tensor("op_2854"), val = tensor(0x1.197998p-40)]; tensor var_2855 = maximum(x = var_2853, y = var_2854)[name = tensor("op_2855")]; tensor shape_24 = shape(x = input_237)[name = tensor("shape_24")]; tensor shape_25 = shape(x = var_2855)[name = tensor("shape_25")]; tensor equal_8_y_0 = const()[name = tensor("equal_8_y_0"), val = tensor(-1)]; tensor equal_8 = equal(x = shape_24, y = equal_8_y_0)[name = tensor("equal_8")]; tensor select_8 = select(a = shape_25, b = shape_24, cond = equal_8)[name = tensor("select_8")]; tensor real_div_8 = real_div(x = select_8, y = shape_25)[name = tensor("real_div_8")]; tensor denom_33 = tile(reps = real_div_8, x = var_2855)[name = tensor("denom_33")]; tensor enc_17 = real_div(x = input_237, y = denom_33)[name = tensor("enc_17")]; tensor cb_17 = const()[name = tensor("cb_17"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44696896)))]; tensor sim_17 = linear(bias = linear_3_bias_0, weight = cb_17, x = enc_17)[name = tensor("sim_17")]; tensor input_241_axis_0 = const()[name = tensor("input_241_axis_0"), val = tensor(-1)]; tensor input_241_keep_dims_0 = const()[name = tensor("input_241_keep_dims_0"), val = tensor(false)]; tensor input_241_output_dtype_0 = const()[name = tensor("input_241_output_dtype_0"), val = tensor("int32")]; tensor 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("input_241")]; tensor var_2876_batch_dims_0 = const()[name = tensor("op_2876_batch_dims_0"), val = tensor(0)]; tensor var_2876_validate_indices_0 = const()[name = tensor("op_2876_validate_indices_0"), val = tensor(false)]; tensor greater_equal_32_y_0 = const()[name = tensor("greater_equal_32_y_0"), val = tensor(0)]; tensor greater_equal_32 = greater_equal(x = input_241, y = greater_equal_32_y_0)[name = tensor("greater_equal_32")]; tensor slice_by_index_32 = const()[name = tensor("slice_by_index_32"), val = tensor(1024)]; tensor add_44 = add(x = input_241, y = slice_by_index_32)[name = tensor("add_44")]; tensor select_48 = select(a = input_241, b = add_44, cond = greater_equal_32)[name = tensor("select_48")]; tensor var_2876_axis_1 = const()[name = tensor("op_2876_axis_1"), val = tensor(0)]; tensor 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("op_2876")]; tensor var_2886 = const()[name = tensor("op_2886"), val = tensor([1, -1, 8])]; tensor var_2887 = reshape(shape = var_2886, x = var_2876)[name = tensor("op_2887")]; tensor input_243_perm_0 = const()[name = tensor("input_243_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_17_pad_type_0 = const()[name = tensor("z_q_17_pad_type_0"), val = tensor("valid")]; tensor z_q_17_strides_0 = const()[name = tensor("z_q_17_strides_0"), val = tensor([1])]; tensor z_q_17_pad_0 = const()[name = tensor("z_q_17_pad_0"), val = tensor([0, 0])]; tensor z_q_17_dilations_0 = const()[name = tensor("z_q_17_dilations_0"), val = tensor([1])]; tensor z_q_17_groups_0 = const()[name = tensor("z_q_17_groups_0"), val = tensor(1)]; tensor input_243 = transpose(perm = input_243_perm_0, x = var_2887)[name = tensor("transpose_29")]; tensor 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("z_q_17")]; tensor input_245 = sub(x = input_235, y = z_q_17)[name = tensor("input_245")]; tensor var_2909 = const()[name = tensor("op_2909"), val = tensor([1, -1])]; tensor var_2910 = reshape(shape = var_2909, x = input_241)[name = tensor("op_2910")]; tensor z_e_19_pad_type_0 = const()[name = tensor("z_e_19_pad_type_0"), val = tensor("valid")]; tensor z_e_19_strides_0 = const()[name = tensor("z_e_19_strides_0"), val = tensor([1])]; tensor z_e_19_pad_0 = const()[name = tensor("z_e_19_pad_0"), val = tensor([0, 0])]; tensor z_e_19_dilations_0 = const()[name = tensor("z_e_19_dilations_0"), val = tensor([1])]; tensor z_e_19_groups_0 = const()[name = tensor("z_e_19_groups_0"), val = tensor(1)]; tensor 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("z_e_19")]; tensor var_2924_perm_0 = const()[name = tensor("op_2924_perm_0"), val = tensor([0, 2, 1])]; tensor var_2930 = const()[name = tensor("op_2930"), val = tensor([-1, 8])]; tensor var_2924 = transpose(perm = var_2924_perm_0, x = z_e_19)[name = tensor("transpose_28")]; tensor input_247 = reshape(shape = var_2930, x = var_2924)[name = tensor("input_247")]; tensor var_2934 = const()[name = tensor("op_2934"), val = tensor([-1])]; tensor var_2935 = const()[name = tensor("op_2935"), val = tensor(true)]; tensor var_2937 = reduce_l2_norm(axes = var_2934, keep_dims = var_2935, x = input_247)[name = tensor("op_2937")]; tensor var_2938 = const()[name = tensor("op_2938"), val = tensor(0x1.197998p-40)]; tensor var_2939 = maximum(x = var_2937, y = var_2938)[name = tensor("op_2939")]; tensor shape_27 = shape(x = input_247)[name = tensor("shape_27")]; tensor shape_28 = shape(x = var_2939)[name = tensor("shape_28")]; tensor equal_9_y_0 = const()[name = tensor("equal_9_y_0"), val = tensor(-1)]; tensor equal_9 = equal(x = shape_27, y = equal_9_y_0)[name = tensor("equal_9")]; tensor select_9 = select(a = shape_28, b = shape_27, cond = equal_9)[name = tensor("select_9")]; tensor real_div_9 = real_div(x = select_9, y = shape_28)[name = tensor("real_div_9")]; tensor denom_37 = tile(reps = real_div_9, x = var_2939)[name = tensor("denom_37")]; tensor enc_19 = real_div(x = input_247, y = denom_37)[name = tensor("enc_19")]; tensor cb_19 = const()[name = tensor("cb_19"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44729728)))]; tensor sim_19 = linear(bias = linear_3_bias_0, weight = cb_19, x = enc_19)[name = tensor("sim_19")]; tensor input_251_axis_0 = const()[name = tensor("input_251_axis_0"), val = tensor(-1)]; tensor input_251_keep_dims_0 = const()[name = tensor("input_251_keep_dims_0"), val = tensor(false)]; tensor input_251_output_dtype_0 = const()[name = tensor("input_251_output_dtype_0"), val = tensor("int32")]; tensor 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("input_251")]; tensor var_2960_batch_dims_0 = const()[name = tensor("op_2960_batch_dims_0"), val = tensor(0)]; tensor var_2960_validate_indices_0 = const()[name = tensor("op_2960_validate_indices_0"), val = tensor(false)]; tensor greater_equal_33_y_0 = const()[name = tensor("greater_equal_33_y_0"), val = tensor(0)]; tensor greater_equal_33 = greater_equal(x = input_251, y = greater_equal_33_y_0)[name = tensor("greater_equal_33")]; tensor slice_by_index_33 = const()[name = tensor("slice_by_index_33"), val = tensor(1024)]; tensor add_45 = add(x = input_251, y = slice_by_index_33)[name = tensor("add_45")]; tensor select_49 = select(a = input_251, b = add_45, cond = greater_equal_33)[name = tensor("select_49")]; tensor var_2960_axis_1 = const()[name = tensor("op_2960_axis_1"), val = tensor(0)]; tensor 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("op_2960")]; tensor var_2970 = const()[name = tensor("op_2970"), val = tensor([1, -1, 8])]; tensor var_2971 = reshape(shape = var_2970, x = var_2960)[name = tensor("op_2971")]; tensor input_253_perm_0 = const()[name = tensor("input_253_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_19_pad_type_0 = const()[name = tensor("z_q_19_pad_type_0"), val = tensor("valid")]; tensor z_q_19_strides_0 = const()[name = tensor("z_q_19_strides_0"), val = tensor([1])]; tensor z_q_19_pad_0 = const()[name = tensor("z_q_19_pad_0"), val = tensor([0, 0])]; tensor z_q_19_dilations_0 = const()[name = tensor("z_q_19_dilations_0"), val = tensor([1])]; tensor z_q_19_groups_0 = const()[name = tensor("z_q_19_groups_0"), val = tensor(1)]; tensor input_253 = transpose(perm = input_253_perm_0, x = var_2971)[name = tensor("transpose_27")]; tensor 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("z_q_19")]; tensor input_255 = sub(x = input_245, y = z_q_19)[name = tensor("input_255")]; tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, -1])]; tensor var_2994 = reshape(shape = var_2993, x = input_251)[name = tensor("op_2994")]; tensor z_e_21_pad_type_0 = const()[name = tensor("z_e_21_pad_type_0"), val = tensor("valid")]; tensor z_e_21_strides_0 = const()[name = tensor("z_e_21_strides_0"), val = tensor([1])]; tensor z_e_21_pad_0 = const()[name = tensor("z_e_21_pad_0"), val = tensor([0, 0])]; tensor z_e_21_dilations_0 = const()[name = tensor("z_e_21_dilations_0"), val = tensor([1])]; tensor z_e_21_groups_0 = const()[name = tensor("z_e_21_groups_0"), val = tensor(1)]; tensor 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("z_e_21")]; tensor var_3008_perm_0 = const()[name = tensor("op_3008_perm_0"), val = tensor([0, 2, 1])]; tensor var_3014 = const()[name = tensor("op_3014"), val = tensor([-1, 8])]; tensor var_3008 = transpose(perm = var_3008_perm_0, x = z_e_21)[name = tensor("transpose_26")]; tensor input_257 = reshape(shape = var_3014, x = var_3008)[name = tensor("input_257")]; tensor var_3018 = const()[name = tensor("op_3018"), val = tensor([-1])]; tensor var_3019 = const()[name = tensor("op_3019"), val = tensor(true)]; tensor var_3021 = reduce_l2_norm(axes = var_3018, keep_dims = var_3019, x = input_257)[name = tensor("op_3021")]; tensor var_3022 = const()[name = tensor("op_3022"), val = tensor(0x1.197998p-40)]; tensor var_3023 = maximum(x = var_3021, y = var_3022)[name = tensor("op_3023")]; tensor shape_30 = shape(x = input_257)[name = tensor("shape_30")]; tensor shape_31 = shape(x = var_3023)[name = tensor("shape_31")]; tensor equal_10_y_0 = const()[name = tensor("equal_10_y_0"), val = tensor(-1)]; tensor equal_10 = equal(x = shape_30, y = equal_10_y_0)[name = tensor("equal_10")]; tensor select_10 = select(a = shape_31, b = shape_30, cond = equal_10)[name = tensor("select_10")]; tensor real_div_10 = real_div(x = select_10, y = shape_31)[name = tensor("real_div_10")]; tensor denom_41 = tile(reps = real_div_10, x = var_3023)[name = tensor("denom_41")]; tensor enc_21 = real_div(x = input_257, y = denom_41)[name = tensor("enc_21")]; tensor cb_21 = const()[name = tensor("cb_21"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44762560)))]; tensor sim_21 = linear(bias = linear_3_bias_0, weight = cb_21, x = enc_21)[name = tensor("sim_21")]; tensor input_261_axis_0 = const()[name = tensor("input_261_axis_0"), val = tensor(-1)]; tensor input_261_keep_dims_0 = const()[name = tensor("input_261_keep_dims_0"), val = tensor(false)]; tensor input_261_output_dtype_0 = const()[name = tensor("input_261_output_dtype_0"), val = tensor("int32")]; tensor 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("input_261")]; tensor var_3044_batch_dims_0 = const()[name = tensor("op_3044_batch_dims_0"), val = tensor(0)]; tensor var_3044_validate_indices_0 = const()[name = tensor("op_3044_validate_indices_0"), val = tensor(false)]; tensor greater_equal_34_y_0 = const()[name = tensor("greater_equal_34_y_0"), val = tensor(0)]; tensor greater_equal_34 = greater_equal(x = input_261, y = greater_equal_34_y_0)[name = tensor("greater_equal_34")]; tensor slice_by_index_34 = const()[name = tensor("slice_by_index_34"), val = tensor(1024)]; tensor add_46 = add(x = input_261, y = slice_by_index_34)[name = tensor("add_46")]; tensor select_50 = select(a = input_261, b = add_46, cond = greater_equal_34)[name = tensor("select_50")]; tensor var_3044_axis_1 = const()[name = tensor("op_3044_axis_1"), val = tensor(0)]; tensor 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("op_3044")]; tensor var_3054 = const()[name = tensor("op_3054"), val = tensor([1, -1, 8])]; tensor var_3055 = reshape(shape = var_3054, x = var_3044)[name = tensor("op_3055")]; tensor input_263_perm_0 = const()[name = tensor("input_263_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_21_pad_type_0 = const()[name = tensor("z_q_21_pad_type_0"), val = tensor("valid")]; tensor z_q_21_strides_0 = const()[name = tensor("z_q_21_strides_0"), val = tensor([1])]; tensor z_q_21_pad_0 = const()[name = tensor("z_q_21_pad_0"), val = tensor([0, 0])]; tensor z_q_21_dilations_0 = const()[name = tensor("z_q_21_dilations_0"), val = tensor([1])]; tensor z_q_21_groups_0 = const()[name = tensor("z_q_21_groups_0"), val = tensor(1)]; tensor input_263 = transpose(perm = input_263_perm_0, x = var_3055)[name = tensor("transpose_25")]; tensor 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("z_q_21")]; tensor input_265 = sub(x = input_255, y = z_q_21)[name = tensor("input_265")]; tensor var_3077 = const()[name = tensor("op_3077"), val = tensor([1, -1])]; tensor var_3078 = reshape(shape = var_3077, x = input_261)[name = tensor("op_3078")]; tensor z_e_23_pad_type_0 = const()[name = tensor("z_e_23_pad_type_0"), val = tensor("valid")]; tensor z_e_23_strides_0 = const()[name = tensor("z_e_23_strides_0"), val = tensor([1])]; tensor z_e_23_pad_0 = const()[name = tensor("z_e_23_pad_0"), val = tensor([0, 0])]; tensor z_e_23_dilations_0 = const()[name = tensor("z_e_23_dilations_0"), val = tensor([1])]; tensor z_e_23_groups_0 = const()[name = tensor("z_e_23_groups_0"), val = tensor(1)]; tensor 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("z_e_23")]; tensor var_3092_perm_0 = const()[name = tensor("op_3092_perm_0"), val = tensor([0, 2, 1])]; tensor var_3098 = const()[name = tensor("op_3098"), val = tensor([-1, 8])]; tensor var_3092 = transpose(perm = var_3092_perm_0, x = z_e_23)[name = tensor("transpose_24")]; tensor input_267 = reshape(shape = var_3098, x = var_3092)[name = tensor("input_267")]; tensor var_3102 = const()[name = tensor("op_3102"), val = tensor([-1])]; tensor var_3103 = const()[name = tensor("op_3103"), val = tensor(true)]; tensor var_3105 = reduce_l2_norm(axes = var_3102, keep_dims = var_3103, x = input_267)[name = tensor("op_3105")]; tensor var_3106 = const()[name = tensor("op_3106"), val = tensor(0x1.197998p-40)]; tensor var_3107 = maximum(x = var_3105, y = var_3106)[name = tensor("op_3107")]; tensor shape_33 = shape(x = input_267)[name = tensor("shape_33")]; tensor shape_34 = shape(x = var_3107)[name = tensor("shape_34")]; tensor equal_11_y_0 = const()[name = tensor("equal_11_y_0"), val = tensor(-1)]; tensor equal_11 = equal(x = shape_33, y = equal_11_y_0)[name = tensor("equal_11")]; tensor select_11 = select(a = shape_34, b = shape_33, cond = equal_11)[name = tensor("select_11")]; tensor real_div_11 = real_div(x = select_11, y = shape_34)[name = tensor("real_div_11")]; tensor denom_45 = tile(reps = real_div_11, x = var_3107)[name = tensor("denom_45")]; tensor enc_23 = real_div(x = input_267, y = denom_45)[name = tensor("enc_23")]; tensor cb_23 = const()[name = tensor("cb_23"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44795392)))]; tensor sim_23 = linear(bias = linear_3_bias_0, weight = cb_23, x = enc_23)[name = tensor("sim_23")]; tensor input_271_axis_0 = const()[name = tensor("input_271_axis_0"), val = tensor(-1)]; tensor input_271_keep_dims_0 = const()[name = tensor("input_271_keep_dims_0"), val = tensor(false)]; tensor input_271_output_dtype_0 = const()[name = tensor("input_271_output_dtype_0"), val = tensor("int32")]; tensor 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("input_271")]; tensor var_3128_batch_dims_0 = const()[name = tensor("op_3128_batch_dims_0"), val = tensor(0)]; tensor var_3128_validate_indices_0 = const()[name = tensor("op_3128_validate_indices_0"), val = tensor(false)]; tensor greater_equal_35_y_0 = const()[name = tensor("greater_equal_35_y_0"), val = tensor(0)]; tensor greater_equal_35 = greater_equal(x = input_271, y = greater_equal_35_y_0)[name = tensor("greater_equal_35")]; tensor slice_by_index_35 = const()[name = tensor("slice_by_index_35"), val = tensor(1024)]; tensor add_47 = add(x = input_271, y = slice_by_index_35)[name = tensor("add_47")]; tensor select_51 = select(a = input_271, b = add_47, cond = greater_equal_35)[name = tensor("select_51")]; tensor var_3128_axis_1 = const()[name = tensor("op_3128_axis_1"), val = tensor(0)]; tensor 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("op_3128")]; tensor var_3138 = const()[name = tensor("op_3138"), val = tensor([1, -1, 8])]; tensor var_3139 = reshape(shape = var_3138, x = var_3128)[name = tensor("op_3139")]; tensor input_273_perm_0 = const()[name = tensor("input_273_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_23_pad_type_0 = const()[name = tensor("z_q_23_pad_type_0"), val = tensor("valid")]; tensor z_q_23_strides_0 = const()[name = tensor("z_q_23_strides_0"), val = tensor([1])]; tensor z_q_23_pad_0 = const()[name = tensor("z_q_23_pad_0"), val = tensor([0, 0])]; tensor z_q_23_dilations_0 = const()[name = tensor("z_q_23_dilations_0"), val = tensor([1])]; tensor z_q_23_groups_0 = const()[name = tensor("z_q_23_groups_0"), val = tensor(1)]; tensor input_273 = transpose(perm = input_273_perm_0, x = var_3139)[name = tensor("transpose_23")]; tensor 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("z_q_23")]; tensor input_275 = sub(x = input_265, y = z_q_23)[name = tensor("input_275")]; tensor var_3161 = const()[name = tensor("op_3161"), val = tensor([1, -1])]; tensor var_3162 = reshape(shape = var_3161, x = input_271)[name = tensor("op_3162")]; tensor z_e_25_pad_type_0 = const()[name = tensor("z_e_25_pad_type_0"), val = tensor("valid")]; tensor z_e_25_strides_0 = const()[name = tensor("z_e_25_strides_0"), val = tensor([1])]; tensor z_e_25_pad_0 = const()[name = tensor("z_e_25_pad_0"), val = tensor([0, 0])]; tensor z_e_25_dilations_0 = const()[name = tensor("z_e_25_dilations_0"), val = tensor([1])]; tensor z_e_25_groups_0 = const()[name = tensor("z_e_25_groups_0"), val = tensor(1)]; tensor 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("z_e_25")]; tensor var_3176_perm_0 = const()[name = tensor("op_3176_perm_0"), val = tensor([0, 2, 1])]; tensor var_3182 = const()[name = tensor("op_3182"), val = tensor([-1, 8])]; tensor var_3176 = transpose(perm = var_3176_perm_0, x = z_e_25)[name = tensor("transpose_22")]; tensor input_277 = reshape(shape = var_3182, x = var_3176)[name = tensor("input_277")]; tensor var_3186 = const()[name = tensor("op_3186"), val = tensor([-1])]; tensor var_3187 = const()[name = tensor("op_3187"), val = tensor(true)]; tensor var_3189 = reduce_l2_norm(axes = var_3186, keep_dims = var_3187, x = input_277)[name = tensor("op_3189")]; tensor var_3190 = const()[name = tensor("op_3190"), val = tensor(0x1.197998p-40)]; tensor var_3191 = maximum(x = var_3189, y = var_3190)[name = tensor("op_3191")]; tensor shape_36 = shape(x = input_277)[name = tensor("shape_36")]; tensor shape_37 = shape(x = var_3191)[name = tensor("shape_37")]; tensor equal_12_y_0 = const()[name = tensor("equal_12_y_0"), val = tensor(-1)]; tensor equal_12 = equal(x = shape_36, y = equal_12_y_0)[name = tensor("equal_12")]; tensor select_12 = select(a = shape_37, b = shape_36, cond = equal_12)[name = tensor("select_12")]; tensor real_div_12 = real_div(x = select_12, y = shape_37)[name = tensor("real_div_12")]; tensor denom_49 = tile(reps = real_div_12, x = var_3191)[name = tensor("denom_49")]; tensor enc_25 = real_div(x = input_277, y = denom_49)[name = tensor("enc_25")]; tensor cb_25 = const()[name = tensor("cb_25"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44828224)))]; tensor sim_25 = linear(bias = linear_3_bias_0, weight = cb_25, x = enc_25)[name = tensor("sim_25")]; tensor input_281_axis_0 = const()[name = tensor("input_281_axis_0"), val = tensor(-1)]; tensor input_281_keep_dims_0 = const()[name = tensor("input_281_keep_dims_0"), val = tensor(false)]; tensor input_281_output_dtype_0 = const()[name = tensor("input_281_output_dtype_0"), val = tensor("int32")]; tensor 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("input_281")]; tensor var_3212_batch_dims_0 = const()[name = tensor("op_3212_batch_dims_0"), val = tensor(0)]; tensor var_3212_validate_indices_0 = const()[name = tensor("op_3212_validate_indices_0"), val = tensor(false)]; tensor greater_equal_36_y_0 = const()[name = tensor("greater_equal_36_y_0"), val = tensor(0)]; tensor greater_equal_36 = greater_equal(x = input_281, y = greater_equal_36_y_0)[name = tensor("greater_equal_36")]; tensor slice_by_index_36 = const()[name = tensor("slice_by_index_36"), val = tensor(1024)]; tensor add_48 = add(x = input_281, y = slice_by_index_36)[name = tensor("add_48")]; tensor select_52 = select(a = input_281, b = add_48, cond = greater_equal_36)[name = tensor("select_52")]; tensor var_3212_axis_1 = const()[name = tensor("op_3212_axis_1"), val = tensor(0)]; tensor 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("op_3212")]; tensor var_3222 = const()[name = tensor("op_3222"), val = tensor([1, -1, 8])]; tensor var_3223 = reshape(shape = var_3222, x = var_3212)[name = tensor("op_3223")]; tensor input_283_perm_0 = const()[name = tensor("input_283_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_25_pad_type_0 = const()[name = tensor("z_q_25_pad_type_0"), val = tensor("valid")]; tensor z_q_25_strides_0 = const()[name = tensor("z_q_25_strides_0"), val = tensor([1])]; tensor z_q_25_pad_0 = const()[name = tensor("z_q_25_pad_0"), val = tensor([0, 0])]; tensor z_q_25_dilations_0 = const()[name = tensor("z_q_25_dilations_0"), val = tensor([1])]; tensor z_q_25_groups_0 = const()[name = tensor("z_q_25_groups_0"), val = tensor(1)]; tensor input_283 = transpose(perm = input_283_perm_0, x = var_3223)[name = tensor("transpose_21")]; tensor 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("z_q_25")]; tensor input_285 = sub(x = input_275, y = z_q_25)[name = tensor("input_285")]; tensor var_3245 = const()[name = tensor("op_3245"), val = tensor([1, -1])]; tensor var_3246 = reshape(shape = var_3245, x = input_281)[name = tensor("op_3246")]; tensor z_e_27_pad_type_0 = const()[name = tensor("z_e_27_pad_type_0"), val = tensor("valid")]; tensor z_e_27_strides_0 = const()[name = tensor("z_e_27_strides_0"), val = tensor([1])]; tensor z_e_27_pad_0 = const()[name = tensor("z_e_27_pad_0"), val = tensor([0, 0])]; tensor z_e_27_dilations_0 = const()[name = tensor("z_e_27_dilations_0"), val = tensor([1])]; tensor z_e_27_groups_0 = const()[name = tensor("z_e_27_groups_0"), val = tensor(1)]; tensor 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("z_e_27")]; tensor var_3260_perm_0 = const()[name = tensor("op_3260_perm_0"), val = tensor([0, 2, 1])]; tensor var_3266 = const()[name = tensor("op_3266"), val = tensor([-1, 8])]; tensor var_3260 = transpose(perm = var_3260_perm_0, x = z_e_27)[name = tensor("transpose_20")]; tensor input_287 = reshape(shape = var_3266, x = var_3260)[name = tensor("input_287")]; tensor var_3270 = const()[name = tensor("op_3270"), val = tensor([-1])]; tensor var_3271 = const()[name = tensor("op_3271"), val = tensor(true)]; tensor var_3273 = reduce_l2_norm(axes = var_3270, keep_dims = var_3271, x = input_287)[name = tensor("op_3273")]; tensor var_3274 = const()[name = tensor("op_3274"), val = tensor(0x1.197998p-40)]; tensor var_3275 = maximum(x = var_3273, y = var_3274)[name = tensor("op_3275")]; tensor shape_39 = shape(x = input_287)[name = tensor("shape_39")]; tensor shape_40 = shape(x = var_3275)[name = tensor("shape_40")]; tensor equal_13_y_0 = const()[name = tensor("equal_13_y_0"), val = tensor(-1)]; tensor equal_13 = equal(x = shape_39, y = equal_13_y_0)[name = tensor("equal_13")]; tensor select_13 = select(a = shape_40, b = shape_39, cond = equal_13)[name = tensor("select_13")]; tensor real_div_13 = real_div(x = select_13, y = shape_40)[name = tensor("real_div_13")]; tensor denom_53 = tile(reps = real_div_13, x = var_3275)[name = tensor("denom_53")]; tensor enc_27 = real_div(x = input_287, y = denom_53)[name = tensor("enc_27")]; tensor cb_27 = const()[name = tensor("cb_27"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44861056)))]; tensor sim_27 = linear(bias = linear_3_bias_0, weight = cb_27, x = enc_27)[name = tensor("sim_27")]; tensor input_291_axis_0 = const()[name = tensor("input_291_axis_0"), val = tensor(-1)]; tensor input_291_keep_dims_0 = const()[name = tensor("input_291_keep_dims_0"), val = tensor(false)]; tensor input_291_output_dtype_0 = const()[name = tensor("input_291_output_dtype_0"), val = tensor("int32")]; tensor 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("input_291")]; tensor var_3296_batch_dims_0 = const()[name = tensor("op_3296_batch_dims_0"), val = tensor(0)]; tensor var_3296_validate_indices_0 = const()[name = tensor("op_3296_validate_indices_0"), val = tensor(false)]; tensor greater_equal_37_y_0 = const()[name = tensor("greater_equal_37_y_0"), val = tensor(0)]; tensor greater_equal_37 = greater_equal(x = input_291, y = greater_equal_37_y_0)[name = tensor("greater_equal_37")]; tensor slice_by_index_37 = const()[name = tensor("slice_by_index_37"), val = tensor(1024)]; tensor add_49 = add(x = input_291, y = slice_by_index_37)[name = tensor("add_49")]; tensor select_53 = select(a = input_291, b = add_49, cond = greater_equal_37)[name = tensor("select_53")]; tensor var_3296_axis_1 = const()[name = tensor("op_3296_axis_1"), val = tensor(0)]; tensor 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("op_3296")]; tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, -1, 8])]; tensor var_3307 = reshape(shape = var_3306, x = var_3296)[name = tensor("op_3307")]; tensor input_293_perm_0 = const()[name = tensor("input_293_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_27_pad_type_0 = const()[name = tensor("z_q_27_pad_type_0"), val = tensor("valid")]; tensor z_q_27_strides_0 = const()[name = tensor("z_q_27_strides_0"), val = tensor([1])]; tensor z_q_27_pad_0 = const()[name = tensor("z_q_27_pad_0"), val = tensor([0, 0])]; tensor z_q_27_dilations_0 = const()[name = tensor("z_q_27_dilations_0"), val = tensor([1])]; tensor z_q_27_groups_0 = const()[name = tensor("z_q_27_groups_0"), val = tensor(1)]; tensor input_293 = transpose(perm = input_293_perm_0, x = var_3307)[name = tensor("transpose_19")]; tensor 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("z_q_27")]; tensor input_295 = sub(x = input_285, y = z_q_27)[name = tensor("input_295")]; tensor var_3329 = const()[name = tensor("op_3329"), val = tensor([1, -1])]; tensor var_3330 = reshape(shape = var_3329, x = input_291)[name = tensor("op_3330")]; tensor z_e_29_pad_type_0 = const()[name = tensor("z_e_29_pad_type_0"), val = tensor("valid")]; tensor z_e_29_strides_0 = const()[name = tensor("z_e_29_strides_0"), val = tensor([1])]; tensor z_e_29_pad_0 = const()[name = tensor("z_e_29_pad_0"), val = tensor([0, 0])]; tensor z_e_29_dilations_0 = const()[name = tensor("z_e_29_dilations_0"), val = tensor([1])]; tensor z_e_29_groups_0 = const()[name = tensor("z_e_29_groups_0"), val = tensor(1)]; tensor 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("z_e_29")]; tensor var_3344_perm_0 = const()[name = tensor("op_3344_perm_0"), val = tensor([0, 2, 1])]; tensor var_3350 = const()[name = tensor("op_3350"), val = tensor([-1, 8])]; tensor var_3344 = transpose(perm = var_3344_perm_0, x = z_e_29)[name = tensor("transpose_18")]; tensor input_297 = reshape(shape = var_3350, x = var_3344)[name = tensor("input_297")]; tensor var_3354 = const()[name = tensor("op_3354"), val = tensor([-1])]; tensor var_3355 = const()[name = tensor("op_3355"), val = tensor(true)]; tensor var_3357 = reduce_l2_norm(axes = var_3354, keep_dims = var_3355, x = input_297)[name = tensor("op_3357")]; tensor var_3358 = const()[name = tensor("op_3358"), val = tensor(0x1.197998p-40)]; tensor var_3359 = maximum(x = var_3357, y = var_3358)[name = tensor("op_3359")]; tensor shape_42 = shape(x = input_297)[name = tensor("shape_42")]; tensor shape_43 = shape(x = var_3359)[name = tensor("shape_43")]; tensor equal_14_y_0 = const()[name = tensor("equal_14_y_0"), val = tensor(-1)]; tensor equal_14 = equal(x = shape_42, y = equal_14_y_0)[name = tensor("equal_14")]; tensor select_14 = select(a = shape_43, b = shape_42, cond = equal_14)[name = tensor("select_14")]; tensor real_div_14 = real_div(x = select_14, y = shape_43)[name = tensor("real_div_14")]; tensor denom_57 = tile(reps = real_div_14, x = var_3359)[name = tensor("denom_57")]; tensor enc_29 = real_div(x = input_297, y = denom_57)[name = tensor("enc_29")]; tensor cb_29 = const()[name = tensor("cb_29"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44893888)))]; tensor sim_29 = linear(bias = linear_3_bias_0, weight = cb_29, x = enc_29)[name = tensor("sim_29")]; tensor input_301_axis_0 = const()[name = tensor("input_301_axis_0"), val = tensor(-1)]; tensor input_301_keep_dims_0 = const()[name = tensor("input_301_keep_dims_0"), val = tensor(false)]; tensor input_301_output_dtype_0 = const()[name = tensor("input_301_output_dtype_0"), val = tensor("int32")]; tensor 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("input_301")]; tensor var_3380_batch_dims_0 = const()[name = tensor("op_3380_batch_dims_0"), val = tensor(0)]; tensor var_3380_validate_indices_0 = const()[name = tensor("op_3380_validate_indices_0"), val = tensor(false)]; tensor greater_equal_38_y_0 = const()[name = tensor("greater_equal_38_y_0"), val = tensor(0)]; tensor greater_equal_38 = greater_equal(x = input_301, y = greater_equal_38_y_0)[name = tensor("greater_equal_38")]; tensor slice_by_index_38 = const()[name = tensor("slice_by_index_38"), val = tensor(1024)]; tensor add_50 = add(x = input_301, y = slice_by_index_38)[name = tensor("add_50")]; tensor select_54 = select(a = input_301, b = add_50, cond = greater_equal_38)[name = tensor("select_54")]; tensor var_3380_axis_1 = const()[name = tensor("op_3380_axis_1"), val = tensor(0)]; tensor 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("op_3380")]; tensor var_3390 = const()[name = tensor("op_3390"), val = tensor([1, -1, 8])]; tensor var_3391 = reshape(shape = var_3390, x = var_3380)[name = tensor("op_3391")]; tensor input_303_perm_0 = const()[name = tensor("input_303_perm_0"), val = tensor([0, 2, 1])]; tensor z_q_29_pad_type_0 = const()[name = tensor("z_q_29_pad_type_0"), val = tensor("valid")]; tensor z_q_29_strides_0 = const()[name = tensor("z_q_29_strides_0"), val = tensor([1])]; tensor z_q_29_pad_0 = const()[name = tensor("z_q_29_pad_0"), val = tensor([0, 0])]; tensor z_q_29_dilations_0 = const()[name = tensor("z_q_29_dilations_0"), val = tensor([1])]; tensor z_q_29_groups_0 = const()[name = tensor("z_q_29_groups_0"), val = tensor(1)]; tensor input_303 = transpose(perm = input_303_perm_0, x = var_3391)[name = tensor("transpose_17")]; tensor 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("z_q_29")]; tensor input_305 = sub(x = input_295, y = z_q_29)[name = tensor("input_305")]; tensor var_3413 = const()[name = tensor("op_3413"), val = tensor([1, -1])]; tensor var_3414 = reshape(shape = var_3413, x = input_301)[name = tensor("op_3414")]; tensor z_e_pad_type_0 = const()[name = tensor("z_e_pad_type_0"), val = tensor("valid")]; tensor z_e_strides_0 = const()[name = tensor("z_e_strides_0"), val = tensor([1])]; tensor z_e_pad_0 = const()[name = tensor("z_e_pad_0"), val = tensor([0, 0])]; tensor z_e_dilations_0 = const()[name = tensor("z_e_dilations_0"), val = tensor([1])]; tensor z_e_groups_0 = const()[name = tensor("z_e_groups_0"), val = tensor(1)]; tensor 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("z_e")]; tensor var_3428_perm_0 = const()[name = tensor("op_3428_perm_0"), val = tensor([0, 2, 1])]; tensor var_3434 = const()[name = tensor("op_3434"), val = tensor([-1, 8])]; tensor var_3428 = transpose(perm = var_3428_perm_0, x = z_e)[name = tensor("transpose_16")]; tensor input_307 = reshape(shape = var_3434, x = var_3428)[name = tensor("input_307")]; tensor var_3438 = const()[name = tensor("op_3438"), val = tensor([-1])]; tensor var_3439 = const()[name = tensor("op_3439"), val = tensor(true)]; tensor var_3441 = reduce_l2_norm(axes = var_3438, keep_dims = var_3439, x = input_307)[name = tensor("op_3441")]; tensor var_3442 = const()[name = tensor("op_3442"), val = tensor(0x1.197998p-40)]; tensor var_3443 = maximum(x = var_3441, y = var_3442)[name = tensor("op_3443")]; tensor shape_45 = shape(x = input_307)[name = tensor("shape_45")]; tensor shape_46 = shape(x = var_3443)[name = tensor("shape_46")]; tensor equal_15_y_0 = const()[name = tensor("equal_15_y_0"), val = tensor(-1)]; tensor equal_15 = equal(x = shape_45, y = equal_15_y_0)[name = tensor("equal_15")]; tensor select_15 = select(a = shape_46, b = shape_45, cond = equal_15)[name = tensor("select_15")]; tensor real_div_15 = real_div(x = select_15, y = shape_46)[name = tensor("real_div_15")]; tensor denom_61 = tile(reps = real_div_15, x = var_3443)[name = tensor("denom_61")]; tensor enc = real_div(x = input_307, y = denom_61)[name = tensor("enc")]; tensor cb = const()[name = tensor("cb"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44926720)))]; tensor sim = linear(bias = linear_3_bias_0, weight = cb, x = enc)[name = tensor("sim")]; tensor input_311_axis_0 = const()[name = tensor("input_311_axis_0"), val = tensor(-1)]; tensor input_311_keep_dims_0 = const()[name = tensor("input_311_keep_dims_0"), val = tensor(false)]; tensor input_311_output_dtype_0 = const()[name = tensor("input_311_output_dtype_0"), val = tensor("int32")]; tensor 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("input_311")]; tensor var_3466 = const()[name = tensor("op_3466"), val = tensor([1, -1])]; tensor var_3467 = reshape(shape = var_3466, x = input_311)[name = tensor("op_3467")]; tensor var_3470_axis_0 = const()[name = tensor("op_3470_axis_0"), val = tensor(0)]; tensor 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("op_3470")]; } -> (codes); }