| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})] |
| { |
| func main<ios18>(tensor<fp32, [1, ?]> audio) [FlexibleShapeInformation = tuple<tuple<string, dict<string, tensor<int32, [?]>>>, tuple<string, dict<string, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 64000]}}), ("RangeDims", {{"audio", [[1, 1], [16000, 320000]]}})))] { |
| tensor<fp32, [512, 1, 10]> net_wavlm_feature_extractor_conv_layers_0_conv_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 1, 10]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp32, [512, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5248))))[name = string("net_wavlm_feature_extractor_conv_layers_0_conv_weight_quantized")]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_0_layer_norm_bias = const()[name = string("net_wavlm_feature_extractor_conv_layers_0_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7360)))]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_0_layer_norm_weight = const()[name = string("net_wavlm_feature_extractor_conv_layers_0_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9472)))]; |
| tensor<fp32, [512, 512, 3]> net_wavlm_feature_extractor_conv_layers_1_conv_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11584))), scale = tensor<fp32, [512, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(798080))))[name = string("net_wavlm_feature_extractor_conv_layers_1_conv_weight_quantized")]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_1_layer_norm_bias = const()[name = string("net_wavlm_feature_extractor_conv_layers_1_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800192)))]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_1_layer_norm_weight = const()[name = string("net_wavlm_feature_extractor_conv_layers_1_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(802304)))]; |
| tensor<fp32, [512, 512, 3]> net_wavlm_feature_extractor_conv_layers_2_conv_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(804416))), scale = tensor<fp32, [512, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1590912))))[name = string("net_wavlm_feature_extractor_conv_layers_2_conv_weight_quantized")]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_2_layer_norm_bias = const()[name = string("net_wavlm_feature_extractor_conv_layers_2_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1593024)))]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_2_layer_norm_weight = const()[name = string("net_wavlm_feature_extractor_conv_layers_2_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1595136)))]; |
| tensor<fp32, [512, 512, 3]> net_wavlm_feature_extractor_conv_layers_3_conv_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1597248))), scale = tensor<fp32, [512, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2383744))))[name = string("net_wavlm_feature_extractor_conv_layers_3_conv_weight_quantized")]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_3_layer_norm_bias = const()[name = string("net_wavlm_feature_extractor_conv_layers_3_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2385856)))]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_3_layer_norm_weight = const()[name = string("net_wavlm_feature_extractor_conv_layers_3_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2387968)))]; |
| tensor<fp32, [512, 512, 3]> net_wavlm_feature_extractor_conv_layers_4_conv_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2390080))), scale = tensor<fp32, [512, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3176576))))[name = string("net_wavlm_feature_extractor_conv_layers_4_conv_weight_quantized")]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_4_layer_norm_bias = const()[name = string("net_wavlm_feature_extractor_conv_layers_4_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3178688)))]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_4_layer_norm_weight = const()[name = string("net_wavlm_feature_extractor_conv_layers_4_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3180800)))]; |
| tensor<fp32, [512, 512, 2]> net_wavlm_feature_extractor_conv_layers_5_conv_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512, 2]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3182912))), scale = tensor<fp32, [512, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3707264))))[name = string("net_wavlm_feature_extractor_conv_layers_5_conv_weight_quantized")]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_5_layer_norm_bias = const()[name = string("net_wavlm_feature_extractor_conv_layers_5_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3709376)))]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_5_layer_norm_weight = const()[name = string("net_wavlm_feature_extractor_conv_layers_5_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3711488)))]; |
| tensor<fp32, [512, 512, 2]> net_wavlm_feature_extractor_conv_layers_6_conv_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 512, 2]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3713600))), scale = tensor<fp32, [512, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4237952))))[name = string("net_wavlm_feature_extractor_conv_layers_6_conv_weight_quantized")]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_6_layer_norm_bias = const()[name = string("net_wavlm_feature_extractor_conv_layers_6_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4240064)))]; |
| tensor<fp32, [512]> net_wavlm_feature_extractor_conv_layers_6_layer_norm_weight = const()[name = string("net_wavlm_feature_extractor_conv_layers_6_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4242176)))]; |
| tensor<fp32, [512]> net_wavlm_feature_projection_layer_norm_bias = const()[name = string("net_wavlm_feature_projection_layer_norm_bias"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4244288)))]; |
| tensor<fp32, [512]> net_wavlm_feature_projection_layer_norm_weight = const()[name = string("net_wavlm_feature_projection_layer_norm_weight"), val = tensor<fp32, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4246400)))]; |
| tensor<fp32, [1024]> net_wavlm_feature_projection_projection_bias = const()[name = string("net_wavlm_feature_projection_projection_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4248512)))]; |
| tensor<fp32, [1024, 512]> net_wavlm_feature_projection_projection_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4252672))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4777024))))[name = string("net_wavlm_feature_projection_projection_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_pos_conv_embed_conv_bias = const()[name = string("net_wavlm_encoder_pos_conv_embed_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4781184)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_0_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4785344)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_0_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4789504)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_0_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4793664)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_0_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4797824))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5846464))))[name = string("net_wavlm_encoder_layers_0_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_0_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5850624))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6899264))))[name = string("net_wavlm_encoder_layers_0_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_0_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6903424))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7952064))))[name = string("net_wavlm_encoder_layers_0_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_0_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7956224))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9004864))))[name = string("net_wavlm_encoder_layers_0_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_0_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9009024)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_0_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9013184)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_0_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9017344)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_0_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_0_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9021504)))]; |
| tensor<fp32, [320, 16]> net_wavlm_encoder_layers_0_attention_rel_attn_embed_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [320, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9021632))), scale = tensor<fp32, [320, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9026816))))[name = string("net_wavlm_encoder_layers_0_attention_rel_attn_embed_weight_quantized")]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_0_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_0_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.e4f9f8p-1, 0x1.e4f9f8p-1, 0x1.e4f9f8p-1, 0x1.e4f9f8p-1, 0x1.a27a0ap+0, 0x1.a27a0ap+0, 0x1.a27a0ap+0, 0x1.a27a0ap+0])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_0_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_0_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9028160)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_0_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9030272)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_0_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9034432)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_0_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_0_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9038592)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_0_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9055040))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13249408))))[name = string("net_wavlm_encoder_layers_0_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_0_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_0_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13265856)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_0_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13270016))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17464384))))[name = string("net_wavlm_encoder_layers_0_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_1_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17468544)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_1_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17472704)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_1_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17476864)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_1_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17481024))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18529664))))[name = string("net_wavlm_encoder_layers_1_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_1_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(18533824))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19582464))))[name = string("net_wavlm_encoder_layers_1_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_1_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19586624))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20635264))))[name = string("net_wavlm_encoder_layers_1_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_1_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20639424))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21688064))))[name = string("net_wavlm_encoder_layers_1_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_1_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21692224)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_1_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21696384)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_1_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21700544)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_1_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_1_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21704704)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_1_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_1_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.f823fap+0, 0x1.f823fap+0, 0x1.f823fap+0, 0x1.f823fap+0, 0x1.dcdc62p-2, 0x1.dcdc62p-2, 0x1.dcdc62p-2, 0x1.dcdc62p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_1_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_1_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21704832)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_1_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21706944)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_1_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21711104)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_1_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_1_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21715264)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_1_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21731712))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25926080))))[name = string("net_wavlm_encoder_layers_1_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_1_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_1_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25942528)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_1_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25946688))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30141056))))[name = string("net_wavlm_encoder_layers_1_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_2_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30145216)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_2_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30149376)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_2_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30153536)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_2_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30157696))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31206336))))[name = string("net_wavlm_encoder_layers_2_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_2_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31210496))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32259136))))[name = string("net_wavlm_encoder_layers_2_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_2_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(32263296))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33311936))))[name = string("net_wavlm_encoder_layers_2_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_2_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33316096))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34364736))))[name = string("net_wavlm_encoder_layers_2_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_2_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34368896)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_2_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34373056)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_2_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34377216)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_2_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_2_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34381376)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_2_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_2_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.a4eb54p+0, 0x1.a4eb54p+0, 0x1.a4eb54p+0, 0x1.a4eb54p+0, 0x1.6468bep-2, 0x1.6468bep-2, 0x1.6468bep-2, 0x1.6468bep-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_2_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_2_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34381504)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_2_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34383616)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_2_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34387776)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_2_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_2_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34391936)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_2_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(34408384))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38602752))))[name = string("net_wavlm_encoder_layers_2_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_2_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_2_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38619200)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_2_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38623360))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42817728))))[name = string("net_wavlm_encoder_layers_2_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_3_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42821888)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_3_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42826048)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_3_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42830208)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_3_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(42834368))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43883008))))[name = string("net_wavlm_encoder_layers_3_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_3_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43887168))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44935808))))[name = string("net_wavlm_encoder_layers_3_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_3_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(44939968))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45988608))))[name = string("net_wavlm_encoder_layers_3_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_3_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45992768))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47041408))))[name = string("net_wavlm_encoder_layers_3_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_3_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47045568)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_3_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47049728)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_3_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47053888)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_3_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_3_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47058048)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_3_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_3_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.e2fc34p-1, 0x1.e2fc34p-1, 0x1.e2fc34p-1, 0x1.e2fc34p-1, 0x1.bdeeeap-1, 0x1.bdeeeap-1, 0x1.bdeeeap-1, 0x1.bdeeeap-1])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_3_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_3_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47058176)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_3_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47060288)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_3_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47064448)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_3_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_3_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47068608)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_3_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47085056))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51279424))))[name = string("net_wavlm_encoder_layers_3_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_3_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_3_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51295872)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_3_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(51300032))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55494400))))[name = string("net_wavlm_encoder_layers_3_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_4_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55498560)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_4_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55502720)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_4_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55506880)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_4_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55511040))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56559680))))[name = string("net_wavlm_encoder_layers_4_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_4_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56563840))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57612480))))[name = string("net_wavlm_encoder_layers_4_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_4_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57616640))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58665280))))[name = string("net_wavlm_encoder_layers_4_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_4_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58669440))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59718080))))[name = string("net_wavlm_encoder_layers_4_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_4_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59722240)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_4_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59726400)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_4_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59730560)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_4_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_4_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59734720)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_4_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_4_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.ff6cacp-1, 0x1.ff6cacp-1, 0x1.ff6cacp-1, 0x1.ff6cacp-1, 0x1.38f1d6p-2, 0x1.38f1d6p-2, 0x1.38f1d6p-2, 0x1.38f1d6p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_4_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_4_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59734848)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_4_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59736960)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_4_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59741120)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_4_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_4_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59745280)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_4_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59761728))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63956096))))[name = string("net_wavlm_encoder_layers_4_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_4_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_4_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63972544)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_4_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63976704))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68171072))))[name = string("net_wavlm_encoder_layers_4_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_5_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68175232)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_5_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68179392)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_5_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68183552)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_5_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68187712))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69236352))))[name = string("net_wavlm_encoder_layers_5_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_5_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69240512))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70289152))))[name = string("net_wavlm_encoder_layers_5_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_5_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70293312))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71341952))))[name = string("net_wavlm_encoder_layers_5_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_5_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71346112))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72394752))))[name = string("net_wavlm_encoder_layers_5_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_5_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72398912)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_5_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72403072)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_5_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72407232)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_5_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_5_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72411392)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_5_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_5_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.370f5ap+0, 0x1.370f5ap+0, 0x1.370f5ap+0, 0x1.370f5ap+0, 0x1.f077e2p-1, 0x1.f077e2p-1, 0x1.f077e2p-1, 0x1.f077e2p-1])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_5_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_5_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72411520)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_5_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72413632)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_5_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72417792)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_5_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_5_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72421952)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_5_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(72438400))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76632768))))[name = string("net_wavlm_encoder_layers_5_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_5_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_5_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76649216)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_5_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76653376))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80847744))))[name = string("net_wavlm_encoder_layers_5_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_6_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80851904)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_6_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80856064)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_6_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80860224)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_6_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(80864384))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81913024))))[name = string("net_wavlm_encoder_layers_6_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_6_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81917184))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82965824))))[name = string("net_wavlm_encoder_layers_6_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_6_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(82969984))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84018624))))[name = string("net_wavlm_encoder_layers_6_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_6_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(84022784))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85071424))))[name = string("net_wavlm_encoder_layers_6_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_6_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85075584)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_6_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85079744)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_6_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85083904)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_6_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_6_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85088064)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_6_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_6_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.8c734p-1, 0x1.8c734p-1, 0x1.8c734p-1, 0x1.8c734p-1, 0x1.822ab6p-1, 0x1.822ab6p-1, 0x1.822ab6p-1, 0x1.822ab6p-1])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_6_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_6_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85088192)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_6_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85090304)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_6_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85094464)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_6_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_6_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85098624)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_6_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85115072))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89309440))))[name = string("net_wavlm_encoder_layers_6_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_6_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_6_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89325888)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_6_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(89330048))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93524416))))[name = string("net_wavlm_encoder_layers_6_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_7_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93528576)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_7_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93532736)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_7_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93536896)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_7_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(93541056))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94589696))))[name = string("net_wavlm_encoder_layers_7_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_7_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94593856))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95642496))))[name = string("net_wavlm_encoder_layers_7_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_7_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95646656))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96695296))))[name = string("net_wavlm_encoder_layers_7_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_7_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96699456))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97748096))))[name = string("net_wavlm_encoder_layers_7_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_7_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97752256)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_7_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97756416)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_7_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97760576)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_7_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_7_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97764736)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_7_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_7_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.fa3992p+0, 0x1.fa3992p+0, 0x1.fa3992p+0, 0x1.fa3992p+0, 0x1.f0cb04p-2, 0x1.f0cb04p-2, 0x1.f0cb04p-2, 0x1.f0cb04p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_7_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_7_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97764864)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_7_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97766976)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_7_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97771136)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_7_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_7_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97775296)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_7_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97791744))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(101986112))))[name = string("net_wavlm_encoder_layers_7_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_7_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_7_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102002560)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_7_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102006720))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106201088))))[name = string("net_wavlm_encoder_layers_7_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_8_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106205248)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_8_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106209408)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_8_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106213568)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_8_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106217728))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107266368))))[name = string("net_wavlm_encoder_layers_8_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_8_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107270528))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108319168))))[name = string("net_wavlm_encoder_layers_8_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_8_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108323328))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109371968))))[name = string("net_wavlm_encoder_layers_8_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_8_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109376128))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110424768))))[name = string("net_wavlm_encoder_layers_8_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_8_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110428928)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_8_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110433088)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_8_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110437248)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_8_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_8_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110441408)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_8_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_8_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.26712p+0, 0x1.26712p+0, 0x1.26712p+0, 0x1.26712p+0, 0x1.0c4728p-1, 0x1.0c4728p-1, 0x1.0c4728p-1, 0x1.0c4728p-1])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_8_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_8_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110441536)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_8_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110443648)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_8_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110447808)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_8_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_8_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110451968)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_8_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110468416))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114662784))))[name = string("net_wavlm_encoder_layers_8_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_8_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_8_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114679232)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_8_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114683392))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118877760))))[name = string("net_wavlm_encoder_layers_8_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_9_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118881920)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_9_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118886080)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_9_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118890240)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_9_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118894400))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119943040))))[name = string("net_wavlm_encoder_layers_9_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_9_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(119947200))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(120995840))))[name = string("net_wavlm_encoder_layers_9_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_9_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121000000))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122048640))))[name = string("net_wavlm_encoder_layers_9_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_9_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(122052800))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123101440))))[name = string("net_wavlm_encoder_layers_9_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_9_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123105600)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_9_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123109760)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_9_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123113920)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_9_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_9_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123118080)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_9_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_9_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.3cae84p+0, 0x1.3cae84p+0, 0x1.3cae84p+0, 0x1.3cae84p+0, 0x1.cfe13cp-2, 0x1.cfe13cp-2, 0x1.cfe13cp-2, 0x1.cfe13cp-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_9_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_9_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123118208)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_9_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123120320)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_9_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123124480)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_9_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_9_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123128640)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_9_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123145088))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127339456))))[name = string("net_wavlm_encoder_layers_9_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_9_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_9_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127355904)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_9_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127360064))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131554432))))[name = string("net_wavlm_encoder_layers_9_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_10_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131558592)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_10_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131562752)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_10_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131566912)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_10_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(131571072))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132619712))))[name = string("net_wavlm_encoder_layers_10_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_10_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132623872))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133672512))))[name = string("net_wavlm_encoder_layers_10_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_10_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133676672))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134725312))))[name = string("net_wavlm_encoder_layers_10_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_10_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134729472))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135778112))))[name = string("net_wavlm_encoder_layers_10_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_10_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135782272)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_10_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135786432)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_10_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135790592)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_10_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_10_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135794752)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_10_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_10_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.7b0804p+0, 0x1.7b0804p+0, 0x1.7b0804p+0, 0x1.7b0804p+0, 0x1.a58fdap-5, 0x1.a58fdap-5, 0x1.a58fdap-5, 0x1.a58fdap-5])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_10_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_10_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135794880)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_10_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135796992)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_10_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135801152)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_10_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_10_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135805312)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_10_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135821760))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140016128))))[name = string("net_wavlm_encoder_layers_10_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_10_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_10_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140032576)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_10_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140036736))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144231104))))[name = string("net_wavlm_encoder_layers_10_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_11_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144235264)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_11_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144239424)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_11_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144243584)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_11_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144247744))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145296384))))[name = string("net_wavlm_encoder_layers_11_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_11_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(145300544))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146349184))))[name = string("net_wavlm_encoder_layers_11_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_11_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(146353344))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147401984))))[name = string("net_wavlm_encoder_layers_11_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_11_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147406144))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148454784))))[name = string("net_wavlm_encoder_layers_11_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_11_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148458944)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_11_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148463104)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_11_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148467264)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_11_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_11_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148471424)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_11_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_11_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.3afaa4p+0, 0x1.3afaa4p+0, 0x1.3afaa4p+0, 0x1.3afaa4p+0, 0x1.b90a46p-2, 0x1.b90a46p-2, 0x1.b90a46p-2, 0x1.b90a46p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_11_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_11_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148471552)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_11_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148473664)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_11_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148477824)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_11_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_11_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148481984)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_11_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(148498432))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152692800))))[name = string("net_wavlm_encoder_layers_11_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_11_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_11_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152709248)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_11_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152713408))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156907776))))[name = string("net_wavlm_encoder_layers_11_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_12_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156911936)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_12_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156916096)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_12_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156920256)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_12_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(156924416))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157973056))))[name = string("net_wavlm_encoder_layers_12_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_12_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157977216))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159025856))))[name = string("net_wavlm_encoder_layers_12_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_12_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159030016))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160078656))))[name = string("net_wavlm_encoder_layers_12_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_12_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(160082816))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161131456))))[name = string("net_wavlm_encoder_layers_12_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_12_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161135616)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_12_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161139776)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_12_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161143936)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_12_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_12_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161148096)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_12_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_12_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.99ecf8p-1, 0x1.99ecf8p-1, 0x1.99ecf8p-1, 0x1.99ecf8p-1, 0x1.5cb076p-2, 0x1.5cb076p-2, 0x1.5cb076p-2, 0x1.5cb076p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_12_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_12_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161148224)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_12_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161150336)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_12_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161154496)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_12_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_12_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161158656)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_12_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161175104))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165369472))))[name = string("net_wavlm_encoder_layers_12_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_12_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_12_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165385920)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_12_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165390080))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169584448))))[name = string("net_wavlm_encoder_layers_12_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_13_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169588608)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_13_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169592768)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_13_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169596928)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_13_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(169601088))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170649728))))[name = string("net_wavlm_encoder_layers_13_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_13_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(170653888))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171702528))))[name = string("net_wavlm_encoder_layers_13_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_13_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(171706688))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172755328))))[name = string("net_wavlm_encoder_layers_13_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_13_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(172759488))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173808128))))[name = string("net_wavlm_encoder_layers_13_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_13_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173812288)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_13_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173816448)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_13_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173820608)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_13_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_13_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173824768)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_13_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_13_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.c762b4p+0, 0x1.c762b4p+0, 0x1.c762b4p+0, 0x1.c762b4p+0, 0x1.17e1cp-1, 0x1.17e1cp-1, 0x1.17e1cp-1, 0x1.17e1cp-1])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_13_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_13_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173824896)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_13_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173827008)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_13_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173831168)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_13_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_13_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173835328)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_13_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(173851776))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178046144))))[name = string("net_wavlm_encoder_layers_13_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_13_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_13_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178062592)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_13_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178066752))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182261120))))[name = string("net_wavlm_encoder_layers_13_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_14_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182265280)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_14_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182269440)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_14_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182273600)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_14_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182277760))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183326400))))[name = string("net_wavlm_encoder_layers_14_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_14_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(183330560))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184379200))))[name = string("net_wavlm_encoder_layers_14_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_14_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184383360))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185432000))))[name = string("net_wavlm_encoder_layers_14_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_14_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(185436160))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186484800))))[name = string("net_wavlm_encoder_layers_14_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_14_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186488960)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_14_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186493120)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_14_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186497280)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_14_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_14_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186501440)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_14_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_14_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.614bb6p+0, 0x1.614bb6p+0, 0x1.614bb6p+0, 0x1.614bb6p+0, 0x1.1ab1p-2, 0x1.1ab1p-2, 0x1.1ab1p-2, 0x1.1ab1p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_14_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_14_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186501568)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_14_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186503680)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_14_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186507840)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_14_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_14_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186512000)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_14_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(186528448))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190722816))))[name = string("net_wavlm_encoder_layers_14_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_14_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_14_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190739264)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_14_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(190743424))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194937792))))[name = string("net_wavlm_encoder_layers_14_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_15_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194941952)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_15_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194946112)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_15_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194950272)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_15_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(194954432))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196003072))))[name = string("net_wavlm_encoder_layers_15_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_15_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(196007232))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197055872))))[name = string("net_wavlm_encoder_layers_15_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_15_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197060032))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198108672))))[name = string("net_wavlm_encoder_layers_15_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_15_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(198112832))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199161472))))[name = string("net_wavlm_encoder_layers_15_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_15_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199165632)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_15_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199169792)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_15_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199173952)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_15_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_15_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199178112)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_15_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_15_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.f5f72cp-1, 0x1.f5f72cp-1, 0x1.f5f72cp-1, 0x1.f5f72cp-1, 0x1.784c66p-2, 0x1.784c66p-2, 0x1.784c66p-2, 0x1.784c66p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_15_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_15_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199178240)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_15_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199180352)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_15_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199184512)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_15_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_15_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199188672)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_15_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199205120))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203399488))))[name = string("net_wavlm_encoder_layers_15_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_15_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_15_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203415936)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_15_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203420096))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207614464))))[name = string("net_wavlm_encoder_layers_15_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_16_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207618624)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_16_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207622784)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_16_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207626944)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_16_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207631104))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208679744))))[name = string("net_wavlm_encoder_layers_16_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_16_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(208683904))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209732544))))[name = string("net_wavlm_encoder_layers_16_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_16_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209736704))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210785344))))[name = string("net_wavlm_encoder_layers_16_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_16_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(210789504))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211838144))))[name = string("net_wavlm_encoder_layers_16_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_16_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211842304)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_16_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211846464)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_16_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211850624)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_16_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_16_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211854784)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_16_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_16_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.567308p+0, 0x1.567308p+0, 0x1.567308p+0, 0x1.567308p+0, 0x1.45b52ep-2, 0x1.45b52ep-2, 0x1.45b52ep-2, 0x1.45b52ep-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_16_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_16_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211854912)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_16_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211857024)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_16_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211861184)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_16_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_16_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211865344)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_16_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(211881792))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216076160))))[name = string("net_wavlm_encoder_layers_16_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_16_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_16_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216092608)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_16_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(216096768))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220291136))))[name = string("net_wavlm_encoder_layers_16_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_17_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220295296)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_17_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220299456)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_17_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220303616)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_17_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(220307776))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221356416))))[name = string("net_wavlm_encoder_layers_17_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_17_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(221360576))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222409216))))[name = string("net_wavlm_encoder_layers_17_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_17_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(222413376))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223462016))))[name = string("net_wavlm_encoder_layers_17_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_17_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(223466176))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224514816))))[name = string("net_wavlm_encoder_layers_17_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_17_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224518976)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_17_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224523136)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_17_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224527296)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_17_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_17_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224531456)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_17_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_17_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.0efc32p+0, 0x1.0efc32p+0, 0x1.0efc32p+0, 0x1.0efc32p+0, 0x1.a6b59cp-3, 0x1.a6b59cp-3, 0x1.a6b59cp-3, 0x1.a6b59cp-3])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_17_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_17_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224531584)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_17_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224533696)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_17_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224537856)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_17_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_17_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224542016)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_17_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(224558464))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228752832))))[name = string("net_wavlm_encoder_layers_17_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_17_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_17_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228769280)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_17_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228773440))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232967808))))[name = string("net_wavlm_encoder_layers_17_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_18_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232971968)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_18_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232976128)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_18_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232980288)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_18_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232984448))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234033088))))[name = string("net_wavlm_encoder_layers_18_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_18_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234037248))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235085888))))[name = string("net_wavlm_encoder_layers_18_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_18_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235090048))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236138688))))[name = string("net_wavlm_encoder_layers_18_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_18_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(236142848))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237191488))))[name = string("net_wavlm_encoder_layers_18_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_18_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237195648)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_18_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237199808)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_18_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237203968)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_18_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_18_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237208128)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_18_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_18_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.eba49ep-1, 0x1.eba49ep-1, 0x1.eba49ep-1, 0x1.eba49ep-1, 0x1.9ed59ap-3, 0x1.9ed59ap-3, 0x1.9ed59ap-3, 0x1.9ed59ap-3])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_18_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_18_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237208256)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_18_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237210368)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_18_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237214528)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_18_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_18_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237218688)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_18_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(237235136))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241429504))))[name = string("net_wavlm_encoder_layers_18_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_18_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_18_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241445952)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_18_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(241450112))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245644480))))[name = string("net_wavlm_encoder_layers_18_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_19_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245648640)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_19_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245652800)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_19_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245656960)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_19_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(245661120))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246709760))))[name = string("net_wavlm_encoder_layers_19_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_19_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(246713920))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247762560))))[name = string("net_wavlm_encoder_layers_19_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_19_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(247766720))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248815360))))[name = string("net_wavlm_encoder_layers_19_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_19_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(248819520))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249868160))))[name = string("net_wavlm_encoder_layers_19_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_19_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249872320)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_19_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249876480)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_19_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249880640)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_19_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_19_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249884800)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_19_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_19_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.85e76p-6, 0x1.85e76p-6, 0x1.85e76p-6, 0x1.85e76p-6, 0x1.528ffp-1, 0x1.528ffp-1, 0x1.528ffp-1, 0x1.528ffp-1])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_19_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_19_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249884928)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_19_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249887040)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_19_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249891200)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_19_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_19_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249895360)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_19_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(249911808))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254106176))))[name = string("net_wavlm_encoder_layers_19_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_19_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_19_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254122624)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_19_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254126784))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258321152))))[name = string("net_wavlm_encoder_layers_19_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_20_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258325312)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_20_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258329472)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_20_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258333632)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_20_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258337792))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259386432))))[name = string("net_wavlm_encoder_layers_20_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_20_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(259390592))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260439232))))[name = string("net_wavlm_encoder_layers_20_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_20_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260443392))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261492032))))[name = string("net_wavlm_encoder_layers_20_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_20_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(261496192))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262544832))))[name = string("net_wavlm_encoder_layers_20_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_20_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262548992)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_20_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262553152)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_20_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262557312)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_20_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_20_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262561472)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_20_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_20_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.05953ep+0, 0x1.05953ep+0, 0x1.05953ep+0, 0x1.05953ep+0, 0x1.dec55p-4, 0x1.dec55p-4, 0x1.dec55p-4, 0x1.dec55p-4])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_20_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_20_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262561600)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_20_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262563712)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_20_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262567872)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_20_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_20_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262572032)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_20_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262588480))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266782848))))[name = string("net_wavlm_encoder_layers_20_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_20_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_20_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266799296)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_20_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266803456))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270997824))))[name = string("net_wavlm_encoder_layers_20_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_21_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271001984)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_21_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271006144)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_21_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271010304)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_21_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271014464))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272063104))))[name = string("net_wavlm_encoder_layers_21_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_21_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(272067264))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273115904))))[name = string("net_wavlm_encoder_layers_21_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_21_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273120064))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274168704))))[name = string("net_wavlm_encoder_layers_21_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_21_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(274172864))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275221504))))[name = string("net_wavlm_encoder_layers_21_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_21_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275225664)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_21_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275229824)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_21_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275233984)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_21_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_21_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275238144)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_21_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_21_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.ea7c62p+0, 0x1.ea7c62p+0, 0x1.ea7c62p+0, 0x1.ea7c62p+0, 0x1.824178p-3, 0x1.824178p-3, 0x1.824178p-3, 0x1.824178p-3])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_21_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_21_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275238272)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_21_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275240384)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_21_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275244544)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_21_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_21_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275248704)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_21_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(275265152))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279459520))))[name = string("net_wavlm_encoder_layers_21_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_21_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_21_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279475968)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_21_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279480128))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283674496))))[name = string("net_wavlm_encoder_layers_21_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_22_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283678656)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_22_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283682816)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_22_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283686976)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_22_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283691136))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284739776))))[name = string("net_wavlm_encoder_layers_22_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_22_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(284743936))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285792576))))[name = string("net_wavlm_encoder_layers_22_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_22_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285796736))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286845376))))[name = string("net_wavlm_encoder_layers_22_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_22_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(286849536))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287898176))))[name = string("net_wavlm_encoder_layers_22_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_22_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287902336)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_22_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287906496)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_22_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287910656)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_22_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_22_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287914816)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_22_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_22_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.0b0192p+0, 0x1.0b0192p+0, 0x1.0b0192p+0, 0x1.0b0192p+0, 0x1.e89ff6p-2, 0x1.e89ff6p-2, 0x1.e89ff6p-2, 0x1.e89ff6p-2])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_22_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_22_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287914944)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_22_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287917056)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_22_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287921216)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_22_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_22_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287925376)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_22_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(287941824))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292136192))))[name = string("net_wavlm_encoder_layers_22_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_22_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_22_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292152640)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_22_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292156800))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296351168))))[name = string("net_wavlm_encoder_layers_22_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_23_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296355328)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_23_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296359488)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_attention_out_proj_bias = const()[name = string("net_wavlm_encoder_layers_23_attention_out_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296363648)))]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_23_attention_out_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296367808))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297416448))))[name = string("net_wavlm_encoder_layers_23_attention_out_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_23_attention_v_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297420608))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298469248))))[name = string("net_wavlm_encoder_layers_23_attention_v_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_23_attention_k_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298473408))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299522048))))[name = string("net_wavlm_encoder_layers_23_attention_k_proj_weight_quantized")]; |
| tensor<fp32, [1024, 1024]> net_wavlm_encoder_layers_23_attention_q_proj_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299526208))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300574848))))[name = string("net_wavlm_encoder_layers_23_attention_q_proj_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_attention_v_proj_bias = const()[name = string("net_wavlm_encoder_layers_23_attention_v_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300579008)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_attention_k_proj_bias = const()[name = string("net_wavlm_encoder_layers_23_attention_k_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300583168)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_attention_q_proj_bias = const()[name = string("net_wavlm_encoder_layers_23_attention_q_proj_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300587328)))]; |
| tensor<fp32, [1, 16, 1, 1]> net_wavlm_encoder_layers_23_attention_gru_rel_pos_const = const()[name = string("net_wavlm_encoder_layers_23_attention_gru_rel_pos_const"), val = tensor<fp32, [1, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300591488)))]; |
| tensor<fp32, [8]> net_wavlm_encoder_layers_23_attention_gru_rel_pos_linear_bias = const()[name = string("net_wavlm_encoder_layers_23_attention_gru_rel_pos_linear_bias"), val = tensor<fp32, [8]>([0x1.c83ddp+0, 0x1.c83ddp+0, 0x1.c83ddp+0, 0x1.c83ddp+0, 0x1.3520bcp+0, 0x1.3520bcp+0, 0x1.3520bcp+0, 0x1.3520bcp+0])]; |
| tensor<fp32, [8, 64]> net_wavlm_encoder_layers_23_attention_gru_rel_pos_linear_weight = const()[name = string("net_wavlm_encoder_layers_23_attention_gru_rel_pos_linear_weight"), val = tensor<fp32, [8, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300591616)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_final_layer_norm_bias = const()[name = string("net_wavlm_encoder_layers_23_final_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300593728)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_final_layer_norm_weight = const()[name = string("net_wavlm_encoder_layers_23_final_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300597888)))]; |
| tensor<fp32, [4096]> net_wavlm_encoder_layers_23_feed_forward_intermediate_dense_bias = const()[name = string("net_wavlm_encoder_layers_23_feed_forward_intermediate_dense_bias"), val = tensor<fp32, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300602048)))]; |
| tensor<fp32, [4096, 1024]> net_wavlm_encoder_layers_23_feed_forward_intermediate_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(300618496))), scale = tensor<fp32, [4096, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304812864))))[name = string("net_wavlm_encoder_layers_23_feed_forward_intermediate_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layers_23_feed_forward_output_dense_bias = const()[name = string("net_wavlm_encoder_layers_23_feed_forward_output_dense_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304829312)))]; |
| tensor<fp32, [1024, 4096]> net_wavlm_encoder_layers_23_feed_forward_output_dense_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304833472))), scale = tensor<fp32, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309027840))))[name = string("net_wavlm_encoder_layers_23_feed_forward_output_dense_weight_quantized")]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layer_norm_bias = const()[name = string("net_wavlm_encoder_layer_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309032000)))]; |
| tensor<fp32, [1024]> net_wavlm_encoder_layer_norm_weight = const()[name = string("net_wavlm_encoder_layer_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309036160)))]; |
| tensor<fp32, [46]> net_lm_head_bias = const()[name = string("net_lm_head_bias"), val = tensor<fp32, [46]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309040320)))]; |
| tensor<fp32, [46, 1024]> net_lm_head_weight_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [46, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309040576))), scale = tensor<fp32, [46, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309087744))))[name = string("net_lm_head_weight_quantized")]; |
| int32 var_9 = const()[name = string("op_9"), val = int32(-1)]; |
| fp32 var_10 = const()[name = string("op_10"), val = fp32(0x1.4f8b58p-17)]; |
| int32 var_15 = const()[name = string("op_15"), val = int32(0)]; |
| int32 var_23 = const()[name = string("op_23"), val = int32(80)]; |
| tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp32, [1, 1, ?]> input_1 = expand_dims(axes = input_1_axes_0, x = audio)[name = string("input_1")]; |
| string hidden_states_1_pad_type_0 = const()[name = string("hidden_states_1_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> hidden_states_1_strides_0 = const()[name = string("hidden_states_1_strides_0"), val = tensor<int32, [1]>([5])]; |
| tensor<int32, [2]> hidden_states_1_pad_0 = const()[name = string("hidden_states_1_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> hidden_states_1_dilations_0 = const()[name = string("hidden_states_1_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 hidden_states_1_groups_0 = const()[name = string("hidden_states_1_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 512, ?]> hidden_states_1 = conv(dilations = hidden_states_1_dilations_0, groups = hidden_states_1_groups_0, pad = hidden_states_1_pad_0, pad_type = hidden_states_1_pad_type_0, strides = hidden_states_1_strides_0, weight = net_wavlm_feature_extractor_conv_layers_0_conv_weight_quantized, x = input_1)[name = string("hidden_states_1")]; |
| tensor<int32, [3]> input_3_perm_0 = const()[name = string("input_3_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<int32, [1]> hidden_states_3_axes_0 = const()[name = string("hidden_states_3_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_3 = transpose(perm = input_3_perm_0, x = hidden_states_1)[name = string("transpose_327")]; |
| tensor<fp32, [1, ?, 512]> hidden_states_3 = layer_norm(axes = hidden_states_3_axes_0, beta = net_wavlm_feature_extractor_conv_layers_0_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_extractor_conv_layers_0_layer_norm_weight, x = input_3)[name = string("hidden_states_3")]; |
| tensor<int32, [3]> input_5_perm_0 = const()[name = string("input_5_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| string input_7_mode_0 = const()[name = string("input_7_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, 512, ?]> input_5 = transpose(perm = input_5_perm_0, x = hidden_states_3)[name = string("transpose_326")]; |
| tensor<fp32, [1, 512, ?]> input_7 = gelu(mode = input_7_mode_0, x = input_5)[name = string("input_7")]; |
| string hidden_states_5_pad_type_0 = const()[name = string("hidden_states_5_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> hidden_states_5_strides_0 = const()[name = string("hidden_states_5_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> hidden_states_5_pad_0 = const()[name = string("hidden_states_5_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> hidden_states_5_dilations_0 = const()[name = string("hidden_states_5_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 hidden_states_5_groups_0 = const()[name = string("hidden_states_5_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 512, ?]> hidden_states_5 = conv(dilations = hidden_states_5_dilations_0, groups = hidden_states_5_groups_0, pad = hidden_states_5_pad_0, pad_type = hidden_states_5_pad_type_0, strides = hidden_states_5_strides_0, weight = net_wavlm_feature_extractor_conv_layers_1_conv_weight_quantized, x = input_7)[name = string("hidden_states_5")]; |
| tensor<int32, [3]> input_9_perm_0 = const()[name = string("input_9_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = string("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_9 = transpose(perm = input_9_perm_0, x = hidden_states_5)[name = string("transpose_325")]; |
| tensor<fp32, [1, ?, 512]> hidden_states_7 = layer_norm(axes = hidden_states_7_axes_0, beta = net_wavlm_feature_extractor_conv_layers_1_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_extractor_conv_layers_1_layer_norm_weight, x = input_9)[name = string("hidden_states_7")]; |
| tensor<int32, [3]> input_11_perm_0 = const()[name = string("input_11_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| string input_13_mode_0 = const()[name = string("input_13_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, 512, ?]> input_11 = transpose(perm = input_11_perm_0, x = hidden_states_7)[name = string("transpose_324")]; |
| tensor<fp32, [1, 512, ?]> input_13 = gelu(mode = input_13_mode_0, x = input_11)[name = string("input_13")]; |
| string hidden_states_9_pad_type_0 = const()[name = string("hidden_states_9_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> hidden_states_9_strides_0 = const()[name = string("hidden_states_9_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> hidden_states_9_pad_0 = const()[name = string("hidden_states_9_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> hidden_states_9_dilations_0 = const()[name = string("hidden_states_9_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 hidden_states_9_groups_0 = const()[name = string("hidden_states_9_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 512, ?]> hidden_states_9 = conv(dilations = hidden_states_9_dilations_0, groups = hidden_states_9_groups_0, pad = hidden_states_9_pad_0, pad_type = hidden_states_9_pad_type_0, strides = hidden_states_9_strides_0, weight = net_wavlm_feature_extractor_conv_layers_2_conv_weight_quantized, x = input_13)[name = string("hidden_states_9")]; |
| tensor<int32, [3]> input_15_perm_0 = const()[name = string("input_15_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<int32, [1]> hidden_states_11_axes_0 = const()[name = string("hidden_states_11_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_15 = transpose(perm = input_15_perm_0, x = hidden_states_9)[name = string("transpose_323")]; |
| tensor<fp32, [1, ?, 512]> hidden_states_11 = layer_norm(axes = hidden_states_11_axes_0, beta = net_wavlm_feature_extractor_conv_layers_2_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_extractor_conv_layers_2_layer_norm_weight, x = input_15)[name = string("hidden_states_11")]; |
| tensor<int32, [3]> input_17_perm_0 = const()[name = string("input_17_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, 512, ?]> input_17 = transpose(perm = input_17_perm_0, x = hidden_states_11)[name = string("transpose_322")]; |
| tensor<fp32, [1, 512, ?]> input_19 = gelu(mode = input_19_mode_0, x = input_17)[name = string("input_19")]; |
| string hidden_states_13_pad_type_0 = const()[name = string("hidden_states_13_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> hidden_states_13_strides_0 = const()[name = string("hidden_states_13_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> hidden_states_13_pad_0 = const()[name = string("hidden_states_13_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> hidden_states_13_dilations_0 = const()[name = string("hidden_states_13_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 hidden_states_13_groups_0 = const()[name = string("hidden_states_13_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 512, ?]> hidden_states_13 = conv(dilations = hidden_states_13_dilations_0, groups = hidden_states_13_groups_0, pad = hidden_states_13_pad_0, pad_type = hidden_states_13_pad_type_0, strides = hidden_states_13_strides_0, weight = net_wavlm_feature_extractor_conv_layers_3_conv_weight_quantized, x = input_19)[name = string("hidden_states_13")]; |
| tensor<int32, [3]> input_21_perm_0 = const()[name = string("input_21_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<int32, [1]> hidden_states_15_axes_0 = const()[name = string("hidden_states_15_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_21 = transpose(perm = input_21_perm_0, x = hidden_states_13)[name = string("transpose_321")]; |
| tensor<fp32, [1, ?, 512]> hidden_states_15 = layer_norm(axes = hidden_states_15_axes_0, beta = net_wavlm_feature_extractor_conv_layers_3_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_extractor_conv_layers_3_layer_norm_weight, x = input_21)[name = string("hidden_states_15")]; |
| tensor<int32, [3]> input_23_perm_0 = const()[name = string("input_23_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| string input_25_mode_0 = const()[name = string("input_25_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, 512, ?]> input_23 = transpose(perm = input_23_perm_0, x = hidden_states_15)[name = string("transpose_320")]; |
| tensor<fp32, [1, 512, ?]> input_25 = gelu(mode = input_25_mode_0, x = input_23)[name = string("input_25")]; |
| string hidden_states_17_pad_type_0 = const()[name = string("hidden_states_17_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> hidden_states_17_strides_0 = const()[name = string("hidden_states_17_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> hidden_states_17_pad_0 = const()[name = string("hidden_states_17_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> hidden_states_17_dilations_0 = const()[name = string("hidden_states_17_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 hidden_states_17_groups_0 = const()[name = string("hidden_states_17_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 512, ?]> hidden_states_17 = conv(dilations = hidden_states_17_dilations_0, groups = hidden_states_17_groups_0, pad = hidden_states_17_pad_0, pad_type = hidden_states_17_pad_type_0, strides = hidden_states_17_strides_0, weight = net_wavlm_feature_extractor_conv_layers_4_conv_weight_quantized, x = input_25)[name = string("hidden_states_17")]; |
| tensor<int32, [3]> input_27_perm_0 = const()[name = string("input_27_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = string("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_27 = transpose(perm = input_27_perm_0, x = hidden_states_17)[name = string("transpose_319")]; |
| tensor<fp32, [1, ?, 512]> hidden_states_19 = layer_norm(axes = hidden_states_19_axes_0, beta = net_wavlm_feature_extractor_conv_layers_4_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_extractor_conv_layers_4_layer_norm_weight, x = input_27)[name = string("hidden_states_19")]; |
| tensor<int32, [3]> input_29_perm_0 = const()[name = string("input_29_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| string input_31_mode_0 = const()[name = string("input_31_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, 512, ?]> input_29 = transpose(perm = input_29_perm_0, x = hidden_states_19)[name = string("transpose_318")]; |
| tensor<fp32, [1, 512, ?]> input_31 = gelu(mode = input_31_mode_0, x = input_29)[name = string("input_31")]; |
| string hidden_states_21_pad_type_0 = const()[name = string("hidden_states_21_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> hidden_states_21_strides_0 = const()[name = string("hidden_states_21_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> hidden_states_21_pad_0 = const()[name = string("hidden_states_21_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> hidden_states_21_dilations_0 = const()[name = string("hidden_states_21_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 hidden_states_21_groups_0 = const()[name = string("hidden_states_21_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 512, ?]> hidden_states_21 = conv(dilations = hidden_states_21_dilations_0, groups = hidden_states_21_groups_0, pad = hidden_states_21_pad_0, pad_type = hidden_states_21_pad_type_0, strides = hidden_states_21_strides_0, weight = net_wavlm_feature_extractor_conv_layers_5_conv_weight_quantized, x = input_31)[name = string("hidden_states_21")]; |
| tensor<int32, [3]> input_33_perm_0 = const()[name = string("input_33_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<int32, [1]> hidden_states_23_axes_0 = const()[name = string("hidden_states_23_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_33 = transpose(perm = input_33_perm_0, x = hidden_states_21)[name = string("transpose_317")]; |
| tensor<fp32, [1, ?, 512]> hidden_states_23 = layer_norm(axes = hidden_states_23_axes_0, beta = net_wavlm_feature_extractor_conv_layers_5_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_extractor_conv_layers_5_layer_norm_weight, x = input_33)[name = string("hidden_states_23")]; |
| tensor<int32, [3]> input_35_perm_0 = const()[name = string("input_35_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| string input_37_mode_0 = const()[name = string("input_37_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, 512, ?]> input_35 = transpose(perm = input_35_perm_0, x = hidden_states_23)[name = string("transpose_316")]; |
| tensor<fp32, [1, 512, ?]> input_37 = gelu(mode = input_37_mode_0, x = input_35)[name = string("input_37")]; |
| string hidden_states_25_pad_type_0 = const()[name = string("hidden_states_25_pad_type_0"), val = string("valid")]; |
| tensor<int32, [1]> hidden_states_25_strides_0 = const()[name = string("hidden_states_25_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [2]> hidden_states_25_pad_0 = const()[name = string("hidden_states_25_pad_0"), val = tensor<int32, [2]>([0, 0])]; |
| tensor<int32, [1]> hidden_states_25_dilations_0 = const()[name = string("hidden_states_25_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 hidden_states_25_groups_0 = const()[name = string("hidden_states_25_groups_0"), val = int32(1)]; |
| tensor<fp32, [1, 512, ?]> hidden_states_25 = conv(dilations = hidden_states_25_dilations_0, groups = hidden_states_25_groups_0, pad = hidden_states_25_pad_0, pad_type = hidden_states_25_pad_type_0, strides = hidden_states_25_strides_0, weight = net_wavlm_feature_extractor_conv_layers_6_conv_weight_quantized, x = input_37)[name = string("hidden_states_25")]; |
| tensor<int32, [3]> input_39_perm_0 = const()[name = string("input_39_perm_0"), val = tensor<int32, [3]>([0, -1, -2])]; |
| tensor<int32, [1]> hidden_states_27_axes_0 = const()[name = string("hidden_states_27_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_39 = transpose(perm = input_39_perm_0, x = hidden_states_25)[name = string("transpose_315")]; |
| tensor<fp32, [1, ?, 512]> hidden_states_27 = layer_norm(axes = hidden_states_27_axes_0, beta = net_wavlm_feature_extractor_conv_layers_6_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_extractor_conv_layers_6_layer_norm_weight, x = input_39)[name = string("hidden_states_27")]; |
| string extract_features_mode_0 = const()[name = string("extract_features_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 512]> extract_features = gelu(mode = extract_features_mode_0, x = hidden_states_27)[name = string("extract_features")]; |
| tensor<int32, [1]> input_45_axes_0 = const()[name = string("input_45_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 512]> input_45 = layer_norm(axes = input_45_axes_0, beta = net_wavlm_feature_projection_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_feature_projection_layer_norm_weight, x = extract_features)[name = string("input_45")]; |
| tensor<fp32, [1, ?, 1024]> input_47 = linear(bias = net_wavlm_feature_projection_projection_bias, weight = net_wavlm_feature_projection_projection_weight_quantized, x = input_45)[name = string("linear_0")]; |
| tensor<int32, [3]> input_49_perm_0 = const()[name = string("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1024, 64, 128]> weight_31_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 64, 128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309088000))), scale = tensor<fp32, [1024, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317476672))))[name = string("weight_31_quantized")]; |
| string hidden_states_31_pad_type_0 = const()[name = string("hidden_states_31_pad_type_0"), val = string("custom")]; |
| tensor<int32, [2]> hidden_states_31_pad_0 = const()[name = string("hidden_states_31_pad_0"), val = tensor<int32, [2]>([64, 64])]; |
| int32 hidden_states_31_groups_0 = const()[name = string("hidden_states_31_groups_0"), val = int32(16)]; |
| tensor<int32, [1]> hidden_states_31_strides_0 = const()[name = string("hidden_states_31_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [1]> hidden_states_31_dilations_0 = const()[name = string("hidden_states_31_dilations_0"), val = tensor<int32, [1]>([1])]; |
| tensor<fp32, [1, 1024, ?]> input_49 = transpose(perm = input_49_perm_0, x = input_47)[name = string("transpose_314")]; |
| tensor<fp32, [1, 1024, ?]> hidden_states_31 = conv(bias = net_wavlm_encoder_pos_conv_embed_conv_bias, dilations = hidden_states_31_dilations_0, groups = hidden_states_31_groups_0, pad = hidden_states_31_pad_0, pad_type = hidden_states_31_pad_type_0, strides = hidden_states_31_strides_0, weight = weight_31_quantized, x = input_49)[name = string("hidden_states_31")]; |
| tensor<int32, [3]> input_51_begin_0 = const()[name = string("input_51_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> input_51_end_0 = const()[name = string("input_51_end_0"), val = tensor<int32, [3]>([1, 1024, -1])]; |
| tensor<bool, [3]> input_51_end_mask_0 = const()[name = string("input_51_end_mask_0"), val = tensor<bool, [3]>([true, true, false])]; |
| tensor<fp32, [1, 1024, ?]> input_51 = slice_by_index(begin = input_51_begin_0, end = input_51_end_0, end_mask = input_51_end_mask_0, x = hidden_states_31)[name = string("input_51")]; |
| string hidden_states_33_mode_0 = const()[name = string("hidden_states_33_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, 1024, ?]> hidden_states_33 = gelu(mode = hidden_states_33_mode_0, x = input_51)[name = string("hidden_states_33")]; |
| tensor<int32, [3]> position_embeddings_perm_0 = const()[name = string("position_embeddings_perm_0"), val = tensor<int32, [3]>([0, 2, 1])]; |
| tensor<fp32, [1, ?, 1024]> position_embeddings = transpose(perm = position_embeddings_perm_0, x = hidden_states_33)[name = string("transpose_313")]; |
| tensor<fp32, [1, ?, 1024]> input_53 = add(x = input_47, y = position_embeddings)[name = string("input_53")]; |
| tensor<int32, [1]> hidden_states_35_axes_0 = const()[name = string("hidden_states_35_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_35 = layer_norm(axes = hidden_states_35_axes_0, beta = net_wavlm_encoder_layers_0_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_0_layer_norm_weight, x = input_53)[name = string("hidden_states_35")]; |
| tensor<int32, [3]> var_265_shape = shape(x = hidden_states_35)[name = string("op_265_shape")]; |
| int32 gather_1_batch_dims_0 = const()[name = string("gather_1_batch_dims_0"), val = int32(0)]; |
| bool gather_1_validate_indices_0 = const()[name = string("gather_1_validate_indices_0"), val = bool(false)]; |
| int32 select_0 = const()[name = string("select_0"), val = int32(1)]; |
| int32 gather_1_axis_0 = const()[name = string("gather_1_axis_0"), val = int32(0)]; |
| int32 gather_1 = gather(axis = gather_1_axis_0, batch_dims = gather_1_batch_dims_0, indices = select_0, validate_indices = gather_1_validate_indices_0, x = var_265_shape)[name = string("gather_1")]; |
| int32 const_0 = const()[name = string("const_0"), val = int32(0)]; |
| int32 const_1 = const()[name = string("const_1"), val = int32(1)]; |
| tensor<int32, [?]> var_266 = range_1d(end = gather_1, start = const_0, step = const_1)[name = string("op_266")]; |
| tensor<int32, [1]> context_position_axes_0 = const()[name = string("context_position_axes_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [?, 1]> context_position = expand_dims(axes = context_position_axes_0, x = var_266)[name = string("context_position")]; |
| tensor<int32, [1]> var_270_axes_0 = const()[name = string("op_270_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<int32, [1, ?]> var_270 = expand_dims(axes = var_270_axes_0, x = var_266)[name = string("op_270")]; |
| tensor<int32, [?, ?]> relative_positions_1 = sub(x = var_270, y = context_position)[name = string("relative_positions_1")]; |
| tensor<bool, [?, ?]> var_273 = greater(x = relative_positions_1, y = var_15)[name = string("op_273")]; |
| string var_274_dtype_0 = const()[name = string("op_274_dtype_0"), val = string("int32")]; |
| int32 var_275 = const()[name = string("op_275"), val = int32(160)]; |
| tensor<int32, [?, ?]> var_274 = cast(dtype = var_274_dtype_0, x = var_273)[name = string("cast_2")]; |
| tensor<int32, [?, ?]> relative_buckets = mul(x = var_274, y = var_275)[name = string("relative_buckets")]; |
| tensor<int32, [?, ?]> relative_positions = abs(x = relative_positions_1)[name = string("relative_positions")]; |
| tensor<bool, [?, ?]> is_small = less(x = relative_positions, y = var_23)[name = string("is_small")]; |
| string var_279_dtype_0 = const()[name = string("op_279_dtype_0"), val = string("fp32")]; |
| fp32 _inversed_281_y_0 = const()[name = string("_inversed_281_y_0"), val = fp32(0x1.99999ap-7)]; |
| tensor<fp32, [?, ?]> var_279 = cast(dtype = var_279_dtype_0, x = relative_positions)[name = string("cast_1")]; |
| tensor<fp32, [?, ?]> _inversed_281 = mul(x = var_279, y = _inversed_281_y_0)[name = string("_inversed_281")]; |
| fp32 relative_positions_if_large_1_epsilon_0 = const()[name = string("relative_positions_if_large_1_epsilon_0"), val = fp32(0x1p-149)]; |
| tensor<fp32, [?, ?]> relative_positions_if_large_1 = log(epsilon = relative_positions_if_large_1_epsilon_0, x = _inversed_281)[name = string("relative_positions_if_large_1")]; |
| fp32 _inversed_relative_positions_if_large_3_y_0 = const()[name = string("_inversed_relative_positions_if_large_3_y_0"), val = fp32(0x1.bcb7bp-2)]; |
| tensor<fp32, [?, ?]> _inversed_relative_positions_if_large_3 = mul(x = relative_positions_if_large_1, y = _inversed_relative_positions_if_large_3_y_0)[name = string("_inversed_relative_positions_if_large_3")]; |
| fp32 var_285_promoted = const()[name = string("op_285_promoted"), val = fp32(0x1.4p+6)]; |
| tensor<fp32, [?, ?]> relative_positions_if_large = mul(x = _inversed_relative_positions_if_large_3, y = var_285_promoted)[name = string("relative_positions_if_large")]; |
| fp32 var_287_promoted = const()[name = string("op_287_promoted"), val = fp32(0x1.4p+6)]; |
| tensor<fp32, [?, ?]> var_288 = add(x = relative_positions_if_large, y = var_287_promoted)[name = string("op_288")]; |
| string relative_position_if_large_1_dtype_0 = const()[name = string("relative_position_if_large_1_dtype_0"), val = string("int32")]; |
| int32 var_290_value_0 = const()[name = string("op_290_value_0"), val = int32(159)]; |
| tensor<int32, [?, ?]> relative_position_if_large_1 = cast(dtype = relative_position_if_large_1_dtype_0, x = var_288)[name = string("cast_0")]; |
| tensor<int32, [?, ?]> var_290 = fill_like(ref_tensor = relative_position_if_large_1, value = var_290_value_0)[name = string("op_290")]; |
| tensor<int32, [?, ?]> relative_position_if_large = minimum(x = relative_position_if_large_1, y = var_290)[name = string("relative_position_if_large")]; |
| tensor<int32, [?, ?]> var_292 = select(a = relative_positions, b = relative_position_if_large, cond = is_small)[name = string("op_292")]; |
| tensor<int32, [?, ?]> input_57 = add(x = relative_buckets, y = var_292)[name = string("input_57")]; |
| int32 values_batch_dims_0 = const()[name = string("values_batch_dims_0"), val = int32(0)]; |
| bool values_validate_indices_0 = const()[name = string("values_validate_indices_0"), val = bool(false)]; |
| int32 greater_equal_1_y_0 = const()[name = string("greater_equal_1_y_0"), val = int32(0)]; |
| tensor<bool, [?, ?]> greater_equal_1 = greater_equal(x = input_57, y = greater_equal_1_y_0)[name = string("greater_equal_1")]; |
| int32 slice_by_index_1 = const()[name = string("slice_by_index_1"), val = int32(320)]; |
| tensor<int32, [?, ?]> add_1 = add(x = input_57, y = slice_by_index_1)[name = string("add_1")]; |
| tensor<int32, [?, ?]> select_1 = select(a = input_57, b = add_1, cond = greater_equal_1)[name = string("select_1")]; |
| int32 greater_equal_1_y_0_1 = const()[name = string("greater_equal_1_y_0_1"), val = int32(0)]; |
| tensor<bool, [?, ?]> greater_equal_1_1 = greater_equal(x = select_1, y = greater_equal_1_y_0_1)[name = string("greater_equal_1_1")]; |
| int32 slice_by_index_1_1 = const()[name = string("slice_by_index_1_1"), val = int32(320)]; |
| tensor<int32, [?, ?]> add_1_1 = add(x = select_1, y = slice_by_index_1_1)[name = string("add_1_1")]; |
| tensor<int32, [?, ?]> select_1_1 = select(a = select_1, b = add_1_1, cond = greater_equal_1_1)[name = string("select_1_1")]; |
| int32 values_axis_0 = const()[name = string("values_axis_0"), val = int32(0)]; |
| tensor<fp32, [?, ?, 16]> values = gather(axis = values_axis_0, batch_dims = values_batch_dims_0, indices = select_1_1, validate_indices = values_validate_indices_0, x = net_wavlm_encoder_layers_0_attention_rel_attn_embed_weight_quantized)[name = string("values")]; |
| tensor<int32, [3]> var_296 = const()[name = string("op_296"), val = tensor<int32, [3]>([2, 0, 1])]; |
| tensor<int32, [1]> var_298_axes_0 = const()[name = string("op_298_axes_0"), val = tensor<int32, [1]>([0])]; |
| tensor<fp32, [16, ?, ?]> position_bias_1 = transpose(perm = var_296, x = values)[name = string("transpose_312")]; |
| tensor<fp32, [1, 16, ?, ?]> var_298 = expand_dims(axes = var_298_axes_0, x = position_bias_1)[name = string("op_298")]; |
| int32 var_303 = const()[name = string("op_303"), val = int32(16)]; |
| int32 concat_0_axis_0 = const()[name = string("concat_0_axis_0"), val = int32(0)]; |
| bool concat_0_interleave_0 = const()[name = string("concat_0_interleave_0"), val = bool(false)]; |
| tensor<int32, [3]> concat_0 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = (var_303, gather_1, gather_1))[name = string("concat_0")]; |
| tensor<fp32, [16, ?, ?]> position_bias = reshape(shape = concat_0, x = var_298)[name = string("position_bias")]; |
| tensor<int32, [4]> concat_1x = const()[name = string("concat_1x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_1 = reshape(shape = concat_1x, x = hidden_states_35)[name = string("gated_hidden_states_1")]; |
| tensor<int32, [4]> var_310 = const()[name = string("op_310"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_59 = transpose(perm = var_310, x = gated_hidden_states_1)[name = string("transpose_311")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_1 = linear(bias = net_wavlm_encoder_layers_0_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_0_attention_gru_rel_pos_linear_weight, x = input_59)[name = string("linear_1")]; |
| tensor<int32, [5]> concat_2x = const()[name = string("concat_2x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_319 = reshape(shape = concat_2x, x = relative_position_proj_1)[name = string("op_319")]; |
| tensor<int32, [1]> relative_position_proj_3_axes_0 = const()[name = string("relative_position_proj_3_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_3_keep_dims_0 = const()[name = string("relative_position_proj_3_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_3 = reduce_sum(axes = relative_position_proj_3_axes_0, keep_dims = relative_position_proj_3_keep_dims_0, x = var_319)[name = string("relative_position_proj_3")]; |
| tensor<fp32, [1, 16, ?, 2]> var_322 = sigmoid(x = relative_position_proj_3)[name = string("op_322")]; |
| tensor<int32, [2]> var_323_split_sizes_0 = const()[name = string("op_323_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_323_axis_0 = const()[name = string("op_323_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_323_0, tensor<fp32, [1, 16, ?, 1]> var_323_1 = split(axis = var_323_axis_0, split_sizes = var_323_split_sizes_0, x = var_322)[name = string("op_323")]; |
| tensor<fp32, [1, 16, ?, 1]> var_325 = mul(x = var_323_1, y = net_wavlm_encoder_layers_0_attention_gru_rel_pos_const)[name = string("op_325")]; |
| fp32 var_326 = const()[name = string("op_326"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_327 = sub(x = var_325, y = var_326)[name = string("op_327")]; |
| tensor<fp32, [1, 16, ?, 1]> var_328 = mul(x = var_323_0, y = var_327)[name = string("op_328")]; |
| fp32 var_329 = const()[name = string("op_329"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_1 = add(x = var_328, y = var_329)[name = string("gate_output_1")]; |
| tensor<int32, [3]> var_334 = const()[name = string("op_334"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_335 = reshape(shape = var_334, x = gate_output_1)[name = string("op_335")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_1 = mul(x = var_335, y = position_bias)[name = string("gated_position_bias_1")]; |
| tensor<int32, [3]> query_1_perm_0 = const()[name = string("query_1_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_1 = transpose(perm = query_1_perm_0, x = hidden_states_35)[name = string("transpose_310")]; |
| tensor<fp32, [?, 1, 1024]> q_1 = linear(bias = net_wavlm_encoder_layers_0_attention_q_proj_bias, weight = net_wavlm_encoder_layers_0_attention_q_proj_weight_quantized, x = query_1)[name = string("linear_2")]; |
| tensor<fp32, [?, 1, 1024]> k_1 = linear(bias = net_wavlm_encoder_layers_0_attention_k_proj_bias, weight = net_wavlm_encoder_layers_0_attention_k_proj_weight_quantized, x = query_1)[name = string("linear_3")]; |
| tensor<fp32, [?, 1, 1024]> v_1 = linear(bias = net_wavlm_encoder_layers_0_attention_v_proj_bias, weight = net_wavlm_encoder_layers_0_attention_v_proj_weight_quantized, x = query_1)[name = string("linear_4")]; |
| tensor<int32, [3]> concat_4x = const()[name = string("concat_4x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_362 = reshape(shape = concat_4x, x = q_1)[name = string("op_362")]; |
| tensor<int32, [3]> concat_5x = const()[name = string("concat_5x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_369 = reshape(shape = concat_5x, x = k_1)[name = string("op_369")]; |
| tensor<int32, [3]> concat_6x = const()[name = string("concat_6x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_376 = reshape(shape = concat_6x, x = v_1)[name = string("op_376")]; |
| tensor<int32, [3]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_378 = const()[name = string("op_378"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_1 = mul(x = var_362, y = var_378)[name = string("q_scaled_1")]; |
| bool input_61_bmm_transpose_x_0 = const()[name = string("input_61_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_61_bmm_transpose_y_0 = const()[name = string("input_61_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_97 = transpose(perm = transpose_97_perm_0, x = var_369)[name = string("transpose_308")]; |
| tensor<fp32, [16, ?, 64]> transpose_96 = transpose(perm = transpose_96_perm_0, x = q_scaled_1)[name = string("transpose_309")]; |
| tensor<fp32, [16, ?, ?]> input_61_bmm = matmul(transpose_x = input_61_bmm_transpose_x_0, transpose_y = input_61_bmm_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("input_61_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_61 = add(x = gated_position_bias_1, y = input_61_bmm)[name = string("input_61")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_1 = softmax(axis = var_9, x = input_61)[name = string("attn_output_weights_1")]; |
| bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)]; |
| bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = string("transpose_307")]; |
| tensor<fp32, [16, ?, 64]> attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_output_weights_1, y = v_3)[name = string("attn_output_1")]; |
| tensor<int32, [3]> var_384_perm_0 = const()[name = string("op_384_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_7x = const()[name = string("concat_7x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_384 = transpose(perm = var_384_perm_0, x = attn_output_1)[name = string("transpose_306")]; |
| tensor<fp32, [?, 1024]> attn_output_3 = reshape(shape = concat_7x, x = var_384)[name = string("attn_output_3")]; |
| tensor<fp32, [?, 1024]> attn_output_5 = linear(bias = net_wavlm_encoder_layers_0_attention_out_proj_bias, weight = net_wavlm_encoder_layers_0_attention_out_proj_weight_quantized, x = attn_output_3)[name = string("linear_5")]; |
| tensor<int32, [3]> concat_8x = const()[name = string("concat_8x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_7 = reshape(shape = concat_8x, x = attn_output_5)[name = string("attn_output_7")]; |
| tensor<int32, [3]> input_63_perm_0 = const()[name = string("input_63_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_63 = transpose(perm = input_63_perm_0, x = attn_output_7)[name = string("transpose_305")]; |
| tensor<fp32, [1, ?, 1024]> input_65 = add(x = input_53, y = input_63)[name = string("input_65")]; |
| tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_67 = layer_norm(axes = input_67_axes_0, beta = net_wavlm_encoder_layers_0_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_0_final_layer_norm_weight, x = input_65)[name = string("input_67")]; |
| tensor<fp32, [1, ?, 4096]> input_69 = linear(bias = net_wavlm_encoder_layers_0_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_0_feed_forward_intermediate_dense_weight_quantized, x = input_67)[name = string("linear_6")]; |
| string input_71_mode_0 = const()[name = string("input_71_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_71 = gelu(mode = input_71_mode_0, x = input_69)[name = string("input_71")]; |
| tensor<fp32, [1, ?, 1024]> input_75 = linear(bias = net_wavlm_encoder_layers_0_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_0_feed_forward_output_dense_weight_quantized, x = input_71)[name = string("linear_7")]; |
| tensor<fp32, [1, ?, 1024]> input_77 = add(x = input_65, y = input_75)[name = string("input_77")]; |
| tensor<int32, [1]> hidden_states_39_axes_0 = const()[name = string("hidden_states_39_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_39 = layer_norm(axes = hidden_states_39_axes_0, beta = net_wavlm_encoder_layers_1_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_1_layer_norm_weight, x = input_77)[name = string("hidden_states_39")]; |
| tensor<int32, [4]> concat_9x = const()[name = string("concat_9x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_3 = reshape(shape = concat_9x, x = hidden_states_39)[name = string("gated_hidden_states_3")]; |
| tensor<int32, [4]> var_451 = const()[name = string("op_451"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_79 = transpose(perm = var_451, x = gated_hidden_states_3)[name = string("transpose_304")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_5 = linear(bias = net_wavlm_encoder_layers_1_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_1_attention_gru_rel_pos_linear_weight, x = input_79)[name = string("linear_8")]; |
| tensor<int32, [5]> concat_10x = const()[name = string("concat_10x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_460 = reshape(shape = concat_10x, x = relative_position_proj_5)[name = string("op_460")]; |
| tensor<int32, [1]> relative_position_proj_7_axes_0 = const()[name = string("relative_position_proj_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_7_keep_dims_0 = const()[name = string("relative_position_proj_7_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_7 = reduce_sum(axes = relative_position_proj_7_axes_0, keep_dims = relative_position_proj_7_keep_dims_0, x = var_460)[name = string("relative_position_proj_7")]; |
| tensor<fp32, [1, 16, ?, 2]> var_463 = sigmoid(x = relative_position_proj_7)[name = string("op_463")]; |
| tensor<int32, [2]> var_464_split_sizes_0 = const()[name = string("op_464_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_464_axis_0 = const()[name = string("op_464_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_464_0, tensor<fp32, [1, 16, ?, 1]> var_464_1 = split(axis = var_464_axis_0, split_sizes = var_464_split_sizes_0, x = var_463)[name = string("op_464")]; |
| tensor<fp32, [1, 16, ?, 1]> var_466 = mul(x = var_464_1, y = net_wavlm_encoder_layers_1_attention_gru_rel_pos_const)[name = string("op_466")]; |
| fp32 var_467 = const()[name = string("op_467"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_468 = sub(x = var_466, y = var_467)[name = string("op_468")]; |
| tensor<fp32, [1, 16, ?, 1]> var_469 = mul(x = var_464_0, y = var_468)[name = string("op_469")]; |
| fp32 var_470 = const()[name = string("op_470"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_3 = add(x = var_469, y = var_470)[name = string("gate_output_3")]; |
| tensor<int32, [3]> var_475 = const()[name = string("op_475"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_476 = reshape(shape = var_475, x = gate_output_3)[name = string("op_476")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_3 = mul(x = var_476, y = position_bias)[name = string("gated_position_bias_3")]; |
| tensor<int32, [3]> query_3_perm_0 = const()[name = string("query_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_3 = transpose(perm = query_3_perm_0, x = hidden_states_39)[name = string("transpose_303")]; |
| tensor<fp32, [?, 1, 1024]> q_5 = linear(bias = net_wavlm_encoder_layers_1_attention_q_proj_bias, weight = net_wavlm_encoder_layers_1_attention_q_proj_weight_quantized, x = query_3)[name = string("linear_9")]; |
| tensor<fp32, [?, 1, 1024]> k_5 = linear(bias = net_wavlm_encoder_layers_1_attention_k_proj_bias, weight = net_wavlm_encoder_layers_1_attention_k_proj_weight_quantized, x = query_3)[name = string("linear_10")]; |
| tensor<fp32, [?, 1, 1024]> v_5 = linear(bias = net_wavlm_encoder_layers_1_attention_v_proj_bias, weight = net_wavlm_encoder_layers_1_attention_v_proj_weight_quantized, x = query_3)[name = string("linear_11")]; |
| tensor<int32, [3]> concat_12x = const()[name = string("concat_12x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_503 = reshape(shape = concat_12x, x = q_5)[name = string("op_503")]; |
| tensor<int32, [3]> concat_13x = const()[name = string("concat_13x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_510 = reshape(shape = concat_13x, x = k_5)[name = string("op_510")]; |
| tensor<int32, [3]> concat_14x = const()[name = string("concat_14x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_517 = reshape(shape = concat_14x, x = v_5)[name = string("op_517")]; |
| tensor<int32, [3]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_519 = const()[name = string("op_519"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_3 = mul(x = var_503, y = var_519)[name = string("q_scaled_3")]; |
| bool input_81_bmm_transpose_x_0 = const()[name = string("input_81_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_81_bmm_transpose_y_0 = const()[name = string("input_81_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_99 = transpose(perm = transpose_99_perm_0, x = var_510)[name = string("transpose_301")]; |
| tensor<fp32, [16, ?, 64]> transpose_98 = transpose(perm = transpose_98_perm_0, x = q_scaled_3)[name = string("transpose_302")]; |
| tensor<fp32, [16, ?, ?]> input_81_bmm = matmul(transpose_x = input_81_bmm_transpose_x_0, transpose_y = input_81_bmm_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("input_81_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_81 = add(x = gated_position_bias_3, y = input_81_bmm)[name = string("input_81")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_5 = softmax(axis = var_9, x = input_81)[name = string("attn_output_weights_5")]; |
| bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)]; |
| bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_7 = transpose(perm = v_7_perm_0, x = var_517)[name = string("transpose_300")]; |
| tensor<fp32, [16, ?, 64]> attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_output_weights_5, y = v_7)[name = string("attn_output_9")]; |
| tensor<int32, [3]> var_525_perm_0 = const()[name = string("op_525_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_15x = const()[name = string("concat_15x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_525 = transpose(perm = var_525_perm_0, x = attn_output_9)[name = string("transpose_299")]; |
| tensor<fp32, [?, 1024]> attn_output_11 = reshape(shape = concat_15x, x = var_525)[name = string("attn_output_11")]; |
| tensor<fp32, [?, 1024]> attn_output_13 = linear(bias = net_wavlm_encoder_layers_1_attention_out_proj_bias, weight = net_wavlm_encoder_layers_1_attention_out_proj_weight_quantized, x = attn_output_11)[name = string("linear_12")]; |
| tensor<int32, [3]> concat_16x = const()[name = string("concat_16x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_15 = reshape(shape = concat_16x, x = attn_output_13)[name = string("attn_output_15")]; |
| tensor<int32, [3]> input_83_perm_0 = const()[name = string("input_83_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_83 = transpose(perm = input_83_perm_0, x = attn_output_15)[name = string("transpose_298")]; |
| tensor<fp32, [1, ?, 1024]> input_85 = add(x = input_77, y = input_83)[name = string("input_85")]; |
| tensor<int32, [1]> input_87_axes_0 = const()[name = string("input_87_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_87 = layer_norm(axes = input_87_axes_0, beta = net_wavlm_encoder_layers_1_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_1_final_layer_norm_weight, x = input_85)[name = string("input_87")]; |
| tensor<fp32, [1, ?, 4096]> input_89 = linear(bias = net_wavlm_encoder_layers_1_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_1_feed_forward_intermediate_dense_weight_quantized, x = input_87)[name = string("linear_13")]; |
| string input_91_mode_0 = const()[name = string("input_91_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_91 = gelu(mode = input_91_mode_0, x = input_89)[name = string("input_91")]; |
| tensor<fp32, [1, ?, 1024]> input_95 = linear(bias = net_wavlm_encoder_layers_1_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_1_feed_forward_output_dense_weight_quantized, x = input_91)[name = string("linear_14")]; |
| tensor<fp32, [1, ?, 1024]> input_97 = add(x = input_85, y = input_95)[name = string("input_97")]; |
| tensor<int32, [1]> hidden_states_43_axes_0 = const()[name = string("hidden_states_43_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_43 = layer_norm(axes = hidden_states_43_axes_0, beta = net_wavlm_encoder_layers_2_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_2_layer_norm_weight, x = input_97)[name = string("hidden_states_43")]; |
| tensor<int32, [4]> concat_17x = const()[name = string("concat_17x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_5 = reshape(shape = concat_17x, x = hidden_states_43)[name = string("gated_hidden_states_5")]; |
| tensor<int32, [4]> var_586 = const()[name = string("op_586"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_99 = transpose(perm = var_586, x = gated_hidden_states_5)[name = string("transpose_297")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_9 = linear(bias = net_wavlm_encoder_layers_2_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_2_attention_gru_rel_pos_linear_weight, x = input_99)[name = string("linear_15")]; |
| tensor<int32, [5]> concat_18x = const()[name = string("concat_18x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_595 = reshape(shape = concat_18x, x = relative_position_proj_9)[name = string("op_595")]; |
| tensor<int32, [1]> relative_position_proj_11_axes_0 = const()[name = string("relative_position_proj_11_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_11_keep_dims_0 = const()[name = string("relative_position_proj_11_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_11 = reduce_sum(axes = relative_position_proj_11_axes_0, keep_dims = relative_position_proj_11_keep_dims_0, x = var_595)[name = string("relative_position_proj_11")]; |
| tensor<fp32, [1, 16, ?, 2]> var_598 = sigmoid(x = relative_position_proj_11)[name = string("op_598")]; |
| tensor<int32, [2]> var_599_split_sizes_0 = const()[name = string("op_599_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_599_axis_0 = const()[name = string("op_599_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_599_0, tensor<fp32, [1, 16, ?, 1]> var_599_1 = split(axis = var_599_axis_0, split_sizes = var_599_split_sizes_0, x = var_598)[name = string("op_599")]; |
| tensor<fp32, [1, 16, ?, 1]> var_601 = mul(x = var_599_1, y = net_wavlm_encoder_layers_2_attention_gru_rel_pos_const)[name = string("op_601")]; |
| fp32 var_602 = const()[name = string("op_602"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_603 = sub(x = var_601, y = var_602)[name = string("op_603")]; |
| tensor<fp32, [1, 16, ?, 1]> var_604 = mul(x = var_599_0, y = var_603)[name = string("op_604")]; |
| fp32 var_605 = const()[name = string("op_605"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_5 = add(x = var_604, y = var_605)[name = string("gate_output_5")]; |
| tensor<int32, [3]> var_610 = const()[name = string("op_610"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_611 = reshape(shape = var_610, x = gate_output_5)[name = string("op_611")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_5 = mul(x = var_611, y = position_bias)[name = string("gated_position_bias_5")]; |
| tensor<int32, [3]> query_5_perm_0 = const()[name = string("query_5_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_5 = transpose(perm = query_5_perm_0, x = hidden_states_43)[name = string("transpose_296")]; |
| tensor<fp32, [?, 1, 1024]> q_9 = linear(bias = net_wavlm_encoder_layers_2_attention_q_proj_bias, weight = net_wavlm_encoder_layers_2_attention_q_proj_weight_quantized, x = query_5)[name = string("linear_16")]; |
| tensor<fp32, [?, 1, 1024]> k_9 = linear(bias = net_wavlm_encoder_layers_2_attention_k_proj_bias, weight = net_wavlm_encoder_layers_2_attention_k_proj_weight_quantized, x = query_5)[name = string("linear_17")]; |
| tensor<fp32, [?, 1, 1024]> v_9 = linear(bias = net_wavlm_encoder_layers_2_attention_v_proj_bias, weight = net_wavlm_encoder_layers_2_attention_v_proj_weight_quantized, x = query_5)[name = string("linear_18")]; |
| tensor<int32, [3]> concat_20x = const()[name = string("concat_20x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_638 = reshape(shape = concat_20x, x = q_9)[name = string("op_638")]; |
| tensor<int32, [3]> concat_21x = const()[name = string("concat_21x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_645 = reshape(shape = concat_21x, x = k_9)[name = string("op_645")]; |
| tensor<int32, [3]> concat_22x = const()[name = string("concat_22x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_652 = reshape(shape = concat_22x, x = v_9)[name = string("op_652")]; |
| tensor<int32, [3]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_654 = const()[name = string("op_654"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_5 = mul(x = var_638, y = var_654)[name = string("q_scaled_5")]; |
| bool input_101_bmm_transpose_x_0 = const()[name = string("input_101_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_101_bmm_transpose_y_0 = const()[name = string("input_101_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_101 = transpose(perm = transpose_101_perm_0, x = var_645)[name = string("transpose_294")]; |
| tensor<fp32, [16, ?, 64]> transpose_100 = transpose(perm = transpose_100_perm_0, x = q_scaled_5)[name = string("transpose_295")]; |
| tensor<fp32, [16, ?, ?]> input_101_bmm = matmul(transpose_x = input_101_bmm_transpose_x_0, transpose_y = input_101_bmm_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("input_101_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_101 = add(x = gated_position_bias_5, y = input_101_bmm)[name = string("input_101")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_9 = softmax(axis = var_9, x = input_101)[name = string("attn_output_weights_9")]; |
| bool attn_output_17_transpose_x_0 = const()[name = string("attn_output_17_transpose_x_0"), val = bool(false)]; |
| bool attn_output_17_transpose_y_0 = const()[name = string("attn_output_17_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_11 = transpose(perm = v_11_perm_0, x = var_652)[name = string("transpose_293")]; |
| tensor<fp32, [16, ?, 64]> attn_output_17 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_output_weights_9, y = v_11)[name = string("attn_output_17")]; |
| tensor<int32, [3]> var_660_perm_0 = const()[name = string("op_660_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_23x = const()[name = string("concat_23x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_660 = transpose(perm = var_660_perm_0, x = attn_output_17)[name = string("transpose_292")]; |
| tensor<fp32, [?, 1024]> attn_output_19 = reshape(shape = concat_23x, x = var_660)[name = string("attn_output_19")]; |
| tensor<fp32, [?, 1024]> attn_output_21 = linear(bias = net_wavlm_encoder_layers_2_attention_out_proj_bias, weight = net_wavlm_encoder_layers_2_attention_out_proj_weight_quantized, x = attn_output_19)[name = string("linear_19")]; |
| tensor<int32, [3]> concat_24x = const()[name = string("concat_24x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_23 = reshape(shape = concat_24x, x = attn_output_21)[name = string("attn_output_23")]; |
| tensor<int32, [3]> input_103_perm_0 = const()[name = string("input_103_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_103 = transpose(perm = input_103_perm_0, x = attn_output_23)[name = string("transpose_291")]; |
| tensor<fp32, [1, ?, 1024]> input_105 = add(x = input_97, y = input_103)[name = string("input_105")]; |
| tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_107 = layer_norm(axes = input_107_axes_0, beta = net_wavlm_encoder_layers_2_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_2_final_layer_norm_weight, x = input_105)[name = string("input_107")]; |
| tensor<fp32, [1, ?, 4096]> input_109 = linear(bias = net_wavlm_encoder_layers_2_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_2_feed_forward_intermediate_dense_weight_quantized, x = input_107)[name = string("linear_20")]; |
| string input_111_mode_0 = const()[name = string("input_111_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_111 = gelu(mode = input_111_mode_0, x = input_109)[name = string("input_111")]; |
| tensor<fp32, [1, ?, 1024]> input_115 = linear(bias = net_wavlm_encoder_layers_2_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_2_feed_forward_output_dense_weight_quantized, x = input_111)[name = string("linear_21")]; |
| tensor<fp32, [1, ?, 1024]> input_117 = add(x = input_105, y = input_115)[name = string("input_117")]; |
| tensor<int32, [1]> hidden_states_47_axes_0 = const()[name = string("hidden_states_47_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_47 = layer_norm(axes = hidden_states_47_axes_0, beta = net_wavlm_encoder_layers_3_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_3_layer_norm_weight, x = input_117)[name = string("hidden_states_47")]; |
| tensor<int32, [4]> concat_25x = const()[name = string("concat_25x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_7 = reshape(shape = concat_25x, x = hidden_states_47)[name = string("gated_hidden_states_7")]; |
| tensor<int32, [4]> var_721 = const()[name = string("op_721"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_119 = transpose(perm = var_721, x = gated_hidden_states_7)[name = string("transpose_290")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_13 = linear(bias = net_wavlm_encoder_layers_3_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_3_attention_gru_rel_pos_linear_weight, x = input_119)[name = string("linear_22")]; |
| tensor<int32, [5]> concat_26x = const()[name = string("concat_26x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_730 = reshape(shape = concat_26x, x = relative_position_proj_13)[name = string("op_730")]; |
| tensor<int32, [1]> relative_position_proj_15_axes_0 = const()[name = string("relative_position_proj_15_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_15_keep_dims_0 = const()[name = string("relative_position_proj_15_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_15 = reduce_sum(axes = relative_position_proj_15_axes_0, keep_dims = relative_position_proj_15_keep_dims_0, x = var_730)[name = string("relative_position_proj_15")]; |
| tensor<fp32, [1, 16, ?, 2]> var_733 = sigmoid(x = relative_position_proj_15)[name = string("op_733")]; |
| tensor<int32, [2]> var_734_split_sizes_0 = const()[name = string("op_734_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_734_axis_0 = const()[name = string("op_734_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_734_0, tensor<fp32, [1, 16, ?, 1]> var_734_1 = split(axis = var_734_axis_0, split_sizes = var_734_split_sizes_0, x = var_733)[name = string("op_734")]; |
| tensor<fp32, [1, 16, ?, 1]> var_736 = mul(x = var_734_1, y = net_wavlm_encoder_layers_3_attention_gru_rel_pos_const)[name = string("op_736")]; |
| fp32 var_737 = const()[name = string("op_737"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_738 = sub(x = var_736, y = var_737)[name = string("op_738")]; |
| tensor<fp32, [1, 16, ?, 1]> var_739 = mul(x = var_734_0, y = var_738)[name = string("op_739")]; |
| fp32 var_740 = const()[name = string("op_740"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_7 = add(x = var_739, y = var_740)[name = string("gate_output_7")]; |
| tensor<int32, [3]> var_745 = const()[name = string("op_745"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_746 = reshape(shape = var_745, x = gate_output_7)[name = string("op_746")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_7 = mul(x = var_746, y = position_bias)[name = string("gated_position_bias_7")]; |
| tensor<int32, [3]> query_7_perm_0 = const()[name = string("query_7_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_7 = transpose(perm = query_7_perm_0, x = hidden_states_47)[name = string("transpose_289")]; |
| tensor<fp32, [?, 1, 1024]> q_13 = linear(bias = net_wavlm_encoder_layers_3_attention_q_proj_bias, weight = net_wavlm_encoder_layers_3_attention_q_proj_weight_quantized, x = query_7)[name = string("linear_23")]; |
| tensor<fp32, [?, 1, 1024]> k_13 = linear(bias = net_wavlm_encoder_layers_3_attention_k_proj_bias, weight = net_wavlm_encoder_layers_3_attention_k_proj_weight_quantized, x = query_7)[name = string("linear_24")]; |
| tensor<fp32, [?, 1, 1024]> v_13 = linear(bias = net_wavlm_encoder_layers_3_attention_v_proj_bias, weight = net_wavlm_encoder_layers_3_attention_v_proj_weight_quantized, x = query_7)[name = string("linear_25")]; |
| tensor<int32, [3]> concat_28x = const()[name = string("concat_28x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_773 = reshape(shape = concat_28x, x = q_13)[name = string("op_773")]; |
| tensor<int32, [3]> concat_29x = const()[name = string("concat_29x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_780 = reshape(shape = concat_29x, x = k_13)[name = string("op_780")]; |
| tensor<int32, [3]> concat_30x = const()[name = string("concat_30x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_787 = reshape(shape = concat_30x, x = v_13)[name = string("op_787")]; |
| tensor<int32, [3]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_789 = const()[name = string("op_789"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_7 = mul(x = var_773, y = var_789)[name = string("q_scaled_7")]; |
| bool input_121_bmm_transpose_x_0 = const()[name = string("input_121_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_121_bmm_transpose_y_0 = const()[name = string("input_121_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_103 = transpose(perm = transpose_103_perm_0, x = var_780)[name = string("transpose_287")]; |
| tensor<fp32, [16, ?, 64]> transpose_102 = transpose(perm = transpose_102_perm_0, x = q_scaled_7)[name = string("transpose_288")]; |
| tensor<fp32, [16, ?, ?]> input_121_bmm = matmul(transpose_x = input_121_bmm_transpose_x_0, transpose_y = input_121_bmm_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("input_121_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_121 = add(x = gated_position_bias_7, y = input_121_bmm)[name = string("input_121")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_13 = softmax(axis = var_9, x = input_121)[name = string("attn_output_weights_13")]; |
| bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)]; |
| bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_15 = transpose(perm = v_15_perm_0, x = var_787)[name = string("transpose_286")]; |
| tensor<fp32, [16, ?, 64]> attn_output_25 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_output_weights_13, y = v_15)[name = string("attn_output_25")]; |
| tensor<int32, [3]> var_795_perm_0 = const()[name = string("op_795_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_31x = const()[name = string("concat_31x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_795 = transpose(perm = var_795_perm_0, x = attn_output_25)[name = string("transpose_285")]; |
| tensor<fp32, [?, 1024]> attn_output_27 = reshape(shape = concat_31x, x = var_795)[name = string("attn_output_27")]; |
| tensor<fp32, [?, 1024]> attn_output_29 = linear(bias = net_wavlm_encoder_layers_3_attention_out_proj_bias, weight = net_wavlm_encoder_layers_3_attention_out_proj_weight_quantized, x = attn_output_27)[name = string("linear_26")]; |
| tensor<int32, [3]> concat_32x = const()[name = string("concat_32x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_31 = reshape(shape = concat_32x, x = attn_output_29)[name = string("attn_output_31")]; |
| tensor<int32, [3]> input_123_perm_0 = const()[name = string("input_123_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_123 = transpose(perm = input_123_perm_0, x = attn_output_31)[name = string("transpose_284")]; |
| tensor<fp32, [1, ?, 1024]> input_125 = add(x = input_117, y = input_123)[name = string("input_125")]; |
| tensor<int32, [1]> input_127_axes_0 = const()[name = string("input_127_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_127 = layer_norm(axes = input_127_axes_0, beta = net_wavlm_encoder_layers_3_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_3_final_layer_norm_weight, x = input_125)[name = string("input_127")]; |
| tensor<fp32, [1, ?, 4096]> input_129 = linear(bias = net_wavlm_encoder_layers_3_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_3_feed_forward_intermediate_dense_weight_quantized, x = input_127)[name = string("linear_27")]; |
| string input_131_mode_0 = const()[name = string("input_131_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_131 = gelu(mode = input_131_mode_0, x = input_129)[name = string("input_131")]; |
| tensor<fp32, [1, ?, 1024]> input_135 = linear(bias = net_wavlm_encoder_layers_3_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_3_feed_forward_output_dense_weight_quantized, x = input_131)[name = string("linear_28")]; |
| tensor<fp32, [1, ?, 1024]> input_137 = add(x = input_125, y = input_135)[name = string("input_137")]; |
| tensor<int32, [1]> hidden_states_51_axes_0 = const()[name = string("hidden_states_51_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_51 = layer_norm(axes = hidden_states_51_axes_0, beta = net_wavlm_encoder_layers_4_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_4_layer_norm_weight, x = input_137)[name = string("hidden_states_51")]; |
| tensor<int32, [4]> concat_33x = const()[name = string("concat_33x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_9 = reshape(shape = concat_33x, x = hidden_states_51)[name = string("gated_hidden_states_9")]; |
| tensor<int32, [4]> var_856 = const()[name = string("op_856"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_139 = transpose(perm = var_856, x = gated_hidden_states_9)[name = string("transpose_283")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_17 = linear(bias = net_wavlm_encoder_layers_4_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_4_attention_gru_rel_pos_linear_weight, x = input_139)[name = string("linear_29")]; |
| tensor<int32, [5]> concat_34x = const()[name = string("concat_34x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_865 = reshape(shape = concat_34x, x = relative_position_proj_17)[name = string("op_865")]; |
| tensor<int32, [1]> relative_position_proj_19_axes_0 = const()[name = string("relative_position_proj_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_19_keep_dims_0 = const()[name = string("relative_position_proj_19_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_19 = reduce_sum(axes = relative_position_proj_19_axes_0, keep_dims = relative_position_proj_19_keep_dims_0, x = var_865)[name = string("relative_position_proj_19")]; |
| tensor<fp32, [1, 16, ?, 2]> var_868 = sigmoid(x = relative_position_proj_19)[name = string("op_868")]; |
| tensor<int32, [2]> var_869_split_sizes_0 = const()[name = string("op_869_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_869_axis_0 = const()[name = string("op_869_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_869_0, tensor<fp32, [1, 16, ?, 1]> var_869_1 = split(axis = var_869_axis_0, split_sizes = var_869_split_sizes_0, x = var_868)[name = string("op_869")]; |
| tensor<fp32, [1, 16, ?, 1]> var_871 = mul(x = var_869_1, y = net_wavlm_encoder_layers_4_attention_gru_rel_pos_const)[name = string("op_871")]; |
| fp32 var_872 = const()[name = string("op_872"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_873 = sub(x = var_871, y = var_872)[name = string("op_873")]; |
| tensor<fp32, [1, 16, ?, 1]> var_874 = mul(x = var_869_0, y = var_873)[name = string("op_874")]; |
| fp32 var_875 = const()[name = string("op_875"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_9 = add(x = var_874, y = var_875)[name = string("gate_output_9")]; |
| tensor<int32, [3]> var_880 = const()[name = string("op_880"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_881 = reshape(shape = var_880, x = gate_output_9)[name = string("op_881")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_9 = mul(x = var_881, y = position_bias)[name = string("gated_position_bias_9")]; |
| tensor<int32, [3]> query_9_perm_0 = const()[name = string("query_9_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_9 = transpose(perm = query_9_perm_0, x = hidden_states_51)[name = string("transpose_282")]; |
| tensor<fp32, [?, 1, 1024]> q_17 = linear(bias = net_wavlm_encoder_layers_4_attention_q_proj_bias, weight = net_wavlm_encoder_layers_4_attention_q_proj_weight_quantized, x = query_9)[name = string("linear_30")]; |
| tensor<fp32, [?, 1, 1024]> k_17 = linear(bias = net_wavlm_encoder_layers_4_attention_k_proj_bias, weight = net_wavlm_encoder_layers_4_attention_k_proj_weight_quantized, x = query_9)[name = string("linear_31")]; |
| tensor<fp32, [?, 1, 1024]> v_17 = linear(bias = net_wavlm_encoder_layers_4_attention_v_proj_bias, weight = net_wavlm_encoder_layers_4_attention_v_proj_weight_quantized, x = query_9)[name = string("linear_32")]; |
| tensor<int32, [3]> concat_36x = const()[name = string("concat_36x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_908 = reshape(shape = concat_36x, x = q_17)[name = string("op_908")]; |
| tensor<int32, [3]> concat_37x = const()[name = string("concat_37x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_915 = reshape(shape = concat_37x, x = k_17)[name = string("op_915")]; |
| tensor<int32, [3]> concat_38x = const()[name = string("concat_38x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_922 = reshape(shape = concat_38x, x = v_17)[name = string("op_922")]; |
| tensor<int32, [3]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_924 = const()[name = string("op_924"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_9 = mul(x = var_908, y = var_924)[name = string("q_scaled_9")]; |
| bool input_141_bmm_transpose_x_0 = const()[name = string("input_141_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_141_bmm_transpose_y_0 = const()[name = string("input_141_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_105 = transpose(perm = transpose_105_perm_0, x = var_915)[name = string("transpose_280")]; |
| tensor<fp32, [16, ?, 64]> transpose_104 = transpose(perm = transpose_104_perm_0, x = q_scaled_9)[name = string("transpose_281")]; |
| tensor<fp32, [16, ?, ?]> input_141_bmm = matmul(transpose_x = input_141_bmm_transpose_x_0, transpose_y = input_141_bmm_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("input_141_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_141 = add(x = gated_position_bias_9, y = input_141_bmm)[name = string("input_141")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_17 = softmax(axis = var_9, x = input_141)[name = string("attn_output_weights_17")]; |
| bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)]; |
| bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_19 = transpose(perm = v_19_perm_0, x = var_922)[name = string("transpose_279")]; |
| tensor<fp32, [16, ?, 64]> attn_output_33 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_output_weights_17, y = v_19)[name = string("attn_output_33")]; |
| tensor<int32, [3]> var_930_perm_0 = const()[name = string("op_930_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_39x = const()[name = string("concat_39x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_930 = transpose(perm = var_930_perm_0, x = attn_output_33)[name = string("transpose_278")]; |
| tensor<fp32, [?, 1024]> attn_output_35 = reshape(shape = concat_39x, x = var_930)[name = string("attn_output_35")]; |
| tensor<fp32, [?, 1024]> attn_output_37 = linear(bias = net_wavlm_encoder_layers_4_attention_out_proj_bias, weight = net_wavlm_encoder_layers_4_attention_out_proj_weight_quantized, x = attn_output_35)[name = string("linear_33")]; |
| tensor<int32, [3]> concat_40x = const()[name = string("concat_40x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_39 = reshape(shape = concat_40x, x = attn_output_37)[name = string("attn_output_39")]; |
| tensor<int32, [3]> input_143_perm_0 = const()[name = string("input_143_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_143 = transpose(perm = input_143_perm_0, x = attn_output_39)[name = string("transpose_277")]; |
| tensor<fp32, [1, ?, 1024]> input_145 = add(x = input_137, y = input_143)[name = string("input_145")]; |
| tensor<int32, [1]> input_147_axes_0 = const()[name = string("input_147_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_147 = layer_norm(axes = input_147_axes_0, beta = net_wavlm_encoder_layers_4_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_4_final_layer_norm_weight, x = input_145)[name = string("input_147")]; |
| tensor<fp32, [1, ?, 4096]> input_149 = linear(bias = net_wavlm_encoder_layers_4_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_4_feed_forward_intermediate_dense_weight_quantized, x = input_147)[name = string("linear_34")]; |
| string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_151 = gelu(mode = input_151_mode_0, x = input_149)[name = string("input_151")]; |
| tensor<fp32, [1, ?, 1024]> input_155 = linear(bias = net_wavlm_encoder_layers_4_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_4_feed_forward_output_dense_weight_quantized, x = input_151)[name = string("linear_35")]; |
| tensor<fp32, [1, ?, 1024]> input_157 = add(x = input_145, y = input_155)[name = string("input_157")]; |
| tensor<int32, [1]> hidden_states_55_axes_0 = const()[name = string("hidden_states_55_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_55 = layer_norm(axes = hidden_states_55_axes_0, beta = net_wavlm_encoder_layers_5_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_5_layer_norm_weight, x = input_157)[name = string("hidden_states_55")]; |
| tensor<int32, [4]> concat_41x = const()[name = string("concat_41x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_11 = reshape(shape = concat_41x, x = hidden_states_55)[name = string("gated_hidden_states_11")]; |
| tensor<int32, [4]> var_991 = const()[name = string("op_991"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_159 = transpose(perm = var_991, x = gated_hidden_states_11)[name = string("transpose_276")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_21 = linear(bias = net_wavlm_encoder_layers_5_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_5_attention_gru_rel_pos_linear_weight, x = input_159)[name = string("linear_36")]; |
| tensor<int32, [5]> concat_42x = const()[name = string("concat_42x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1000 = reshape(shape = concat_42x, x = relative_position_proj_21)[name = string("op_1000")]; |
| tensor<int32, [1]> relative_position_proj_23_axes_0 = const()[name = string("relative_position_proj_23_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_23_keep_dims_0 = const()[name = string("relative_position_proj_23_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_23 = reduce_sum(axes = relative_position_proj_23_axes_0, keep_dims = relative_position_proj_23_keep_dims_0, x = var_1000)[name = string("relative_position_proj_23")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1003 = sigmoid(x = relative_position_proj_23)[name = string("op_1003")]; |
| tensor<int32, [2]> var_1004_split_sizes_0 = const()[name = string("op_1004_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1004_axis_0 = const()[name = string("op_1004_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1004_0, tensor<fp32, [1, 16, ?, 1]> var_1004_1 = split(axis = var_1004_axis_0, split_sizes = var_1004_split_sizes_0, x = var_1003)[name = string("op_1004")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1006 = mul(x = var_1004_1, y = net_wavlm_encoder_layers_5_attention_gru_rel_pos_const)[name = string("op_1006")]; |
| fp32 var_1007 = const()[name = string("op_1007"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1008 = sub(x = var_1006, y = var_1007)[name = string("op_1008")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1009 = mul(x = var_1004_0, y = var_1008)[name = string("op_1009")]; |
| fp32 var_1010 = const()[name = string("op_1010"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_11 = add(x = var_1009, y = var_1010)[name = string("gate_output_11")]; |
| tensor<int32, [3]> var_1015 = const()[name = string("op_1015"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1016 = reshape(shape = var_1015, x = gate_output_11)[name = string("op_1016")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_11 = mul(x = var_1016, y = position_bias)[name = string("gated_position_bias_11")]; |
| tensor<int32, [3]> query_11_perm_0 = const()[name = string("query_11_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_11 = transpose(perm = query_11_perm_0, x = hidden_states_55)[name = string("transpose_275")]; |
| tensor<fp32, [?, 1, 1024]> q_21 = linear(bias = net_wavlm_encoder_layers_5_attention_q_proj_bias, weight = net_wavlm_encoder_layers_5_attention_q_proj_weight_quantized, x = query_11)[name = string("linear_37")]; |
| tensor<fp32, [?, 1, 1024]> k_21 = linear(bias = net_wavlm_encoder_layers_5_attention_k_proj_bias, weight = net_wavlm_encoder_layers_5_attention_k_proj_weight_quantized, x = query_11)[name = string("linear_38")]; |
| tensor<fp32, [?, 1, 1024]> v_21 = linear(bias = net_wavlm_encoder_layers_5_attention_v_proj_bias, weight = net_wavlm_encoder_layers_5_attention_v_proj_weight_quantized, x = query_11)[name = string("linear_39")]; |
| tensor<int32, [3]> concat_44x = const()[name = string("concat_44x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1043 = reshape(shape = concat_44x, x = q_21)[name = string("op_1043")]; |
| tensor<int32, [3]> concat_45x = const()[name = string("concat_45x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1050 = reshape(shape = concat_45x, x = k_21)[name = string("op_1050")]; |
| tensor<int32, [3]> concat_46x = const()[name = string("concat_46x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1057 = reshape(shape = concat_46x, x = v_21)[name = string("op_1057")]; |
| tensor<int32, [3]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_1059 = const()[name = string("op_1059"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_11 = mul(x = var_1043, y = var_1059)[name = string("q_scaled_11")]; |
| bool input_161_bmm_transpose_x_0 = const()[name = string("input_161_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_161_bmm_transpose_y_0 = const()[name = string("input_161_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_107 = transpose(perm = transpose_107_perm_0, x = var_1050)[name = string("transpose_273")]; |
| tensor<fp32, [16, ?, 64]> transpose_106 = transpose(perm = transpose_106_perm_0, x = q_scaled_11)[name = string("transpose_274")]; |
| tensor<fp32, [16, ?, ?]> input_161_bmm = matmul(transpose_x = input_161_bmm_transpose_x_0, transpose_y = input_161_bmm_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("input_161_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_161 = add(x = gated_position_bias_11, y = input_161_bmm)[name = string("input_161")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_21 = softmax(axis = var_9, x = input_161)[name = string("attn_output_weights_21")]; |
| bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)]; |
| bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_23 = transpose(perm = v_23_perm_0, x = var_1057)[name = string("transpose_272")]; |
| tensor<fp32, [16, ?, 64]> attn_output_41 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_output_weights_21, y = v_23)[name = string("attn_output_41")]; |
| tensor<int32, [3]> var_1065_perm_0 = const()[name = string("op_1065_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_47x = const()[name = string("concat_47x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_1065 = transpose(perm = var_1065_perm_0, x = attn_output_41)[name = string("transpose_271")]; |
| tensor<fp32, [?, 1024]> attn_output_43 = reshape(shape = concat_47x, x = var_1065)[name = string("attn_output_43")]; |
| tensor<fp32, [?, 1024]> attn_output_45 = linear(bias = net_wavlm_encoder_layers_5_attention_out_proj_bias, weight = net_wavlm_encoder_layers_5_attention_out_proj_weight_quantized, x = attn_output_43)[name = string("linear_40")]; |
| tensor<int32, [3]> concat_48x = const()[name = string("concat_48x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_47 = reshape(shape = concat_48x, x = attn_output_45)[name = string("attn_output_47")]; |
| tensor<int32, [3]> input_163_perm_0 = const()[name = string("input_163_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_163 = transpose(perm = input_163_perm_0, x = attn_output_47)[name = string("transpose_270")]; |
| tensor<fp32, [1, ?, 1024]> input_165 = add(x = input_157, y = input_163)[name = string("input_165")]; |
| tensor<int32, [1]> input_167_axes_0 = const()[name = string("input_167_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_167 = layer_norm(axes = input_167_axes_0, beta = net_wavlm_encoder_layers_5_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_5_final_layer_norm_weight, x = input_165)[name = string("input_167")]; |
| tensor<fp32, [1, ?, 4096]> input_169 = linear(bias = net_wavlm_encoder_layers_5_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_5_feed_forward_intermediate_dense_weight_quantized, x = input_167)[name = string("linear_41")]; |
| string input_171_mode_0 = const()[name = string("input_171_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_171 = gelu(mode = input_171_mode_0, x = input_169)[name = string("input_171")]; |
| tensor<fp32, [1, ?, 1024]> input_175 = linear(bias = net_wavlm_encoder_layers_5_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_5_feed_forward_output_dense_weight_quantized, x = input_171)[name = string("linear_42")]; |
| tensor<fp32, [1, ?, 1024]> input_177 = add(x = input_165, y = input_175)[name = string("input_177")]; |
| tensor<int32, [1]> hidden_states_59_axes_0 = const()[name = string("hidden_states_59_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_59 = layer_norm(axes = hidden_states_59_axes_0, beta = net_wavlm_encoder_layers_6_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_6_layer_norm_weight, x = input_177)[name = string("hidden_states_59")]; |
| tensor<int32, [4]> concat_49x = const()[name = string("concat_49x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_13 = reshape(shape = concat_49x, x = hidden_states_59)[name = string("gated_hidden_states_13")]; |
| tensor<int32, [4]> var_1126 = const()[name = string("op_1126"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_179 = transpose(perm = var_1126, x = gated_hidden_states_13)[name = string("transpose_269")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_25 = linear(bias = net_wavlm_encoder_layers_6_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_6_attention_gru_rel_pos_linear_weight, x = input_179)[name = string("linear_43")]; |
| tensor<int32, [5]> concat_50x = const()[name = string("concat_50x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1135 = reshape(shape = concat_50x, x = relative_position_proj_25)[name = string("op_1135")]; |
| tensor<int32, [1]> relative_position_proj_27_axes_0 = const()[name = string("relative_position_proj_27_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_27_keep_dims_0 = const()[name = string("relative_position_proj_27_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_27 = reduce_sum(axes = relative_position_proj_27_axes_0, keep_dims = relative_position_proj_27_keep_dims_0, x = var_1135)[name = string("relative_position_proj_27")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1138 = sigmoid(x = relative_position_proj_27)[name = string("op_1138")]; |
| tensor<int32, [2]> var_1139_split_sizes_0 = const()[name = string("op_1139_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1139_axis_0 = const()[name = string("op_1139_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1139_0, tensor<fp32, [1, 16, ?, 1]> var_1139_1 = split(axis = var_1139_axis_0, split_sizes = var_1139_split_sizes_0, x = var_1138)[name = string("op_1139")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1141 = mul(x = var_1139_1, y = net_wavlm_encoder_layers_6_attention_gru_rel_pos_const)[name = string("op_1141")]; |
| fp32 var_1142 = const()[name = string("op_1142"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1143 = sub(x = var_1141, y = var_1142)[name = string("op_1143")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1144 = mul(x = var_1139_0, y = var_1143)[name = string("op_1144")]; |
| fp32 var_1145 = const()[name = string("op_1145"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_13 = add(x = var_1144, y = var_1145)[name = string("gate_output_13")]; |
| tensor<int32, [3]> var_1150 = const()[name = string("op_1150"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1151 = reshape(shape = var_1150, x = gate_output_13)[name = string("op_1151")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_13 = mul(x = var_1151, y = position_bias)[name = string("gated_position_bias_13")]; |
| tensor<int32, [3]> query_13_perm_0 = const()[name = string("query_13_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_13 = transpose(perm = query_13_perm_0, x = hidden_states_59)[name = string("transpose_268")]; |
| tensor<fp32, [?, 1, 1024]> q_25 = linear(bias = net_wavlm_encoder_layers_6_attention_q_proj_bias, weight = net_wavlm_encoder_layers_6_attention_q_proj_weight_quantized, x = query_13)[name = string("linear_44")]; |
| tensor<fp32, [?, 1, 1024]> k_25 = linear(bias = net_wavlm_encoder_layers_6_attention_k_proj_bias, weight = net_wavlm_encoder_layers_6_attention_k_proj_weight_quantized, x = query_13)[name = string("linear_45")]; |
| tensor<fp32, [?, 1, 1024]> v_25 = linear(bias = net_wavlm_encoder_layers_6_attention_v_proj_bias, weight = net_wavlm_encoder_layers_6_attention_v_proj_weight_quantized, x = query_13)[name = string("linear_46")]; |
| tensor<int32, [3]> concat_52x = const()[name = string("concat_52x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1178 = reshape(shape = concat_52x, x = q_25)[name = string("op_1178")]; |
| tensor<int32, [3]> concat_53x = const()[name = string("concat_53x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1185 = reshape(shape = concat_53x, x = k_25)[name = string("op_1185")]; |
| tensor<int32, [3]> concat_54x = const()[name = string("concat_54x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1192 = reshape(shape = concat_54x, x = v_25)[name = string("op_1192")]; |
| tensor<int32, [3]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_1194 = const()[name = string("op_1194"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_13 = mul(x = var_1178, y = var_1194)[name = string("q_scaled_13")]; |
| bool input_181_bmm_transpose_x_0 = const()[name = string("input_181_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_181_bmm_transpose_y_0 = const()[name = string("input_181_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_109 = transpose(perm = transpose_109_perm_0, x = var_1185)[name = string("transpose_266")]; |
| tensor<fp32, [16, ?, 64]> transpose_108 = transpose(perm = transpose_108_perm_0, x = q_scaled_13)[name = string("transpose_267")]; |
| tensor<fp32, [16, ?, ?]> input_181_bmm = matmul(transpose_x = input_181_bmm_transpose_x_0, transpose_y = input_181_bmm_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("input_181_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_181 = add(x = gated_position_bias_13, y = input_181_bmm)[name = string("input_181")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_25 = softmax(axis = var_9, x = input_181)[name = string("attn_output_weights_25")]; |
| bool attn_output_49_transpose_x_0 = const()[name = string("attn_output_49_transpose_x_0"), val = bool(false)]; |
| bool attn_output_49_transpose_y_0 = const()[name = string("attn_output_49_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_27 = transpose(perm = v_27_perm_0, x = var_1192)[name = string("transpose_265")]; |
| tensor<fp32, [16, ?, 64]> attn_output_49 = matmul(transpose_x = attn_output_49_transpose_x_0, transpose_y = attn_output_49_transpose_y_0, x = attn_output_weights_25, y = v_27)[name = string("attn_output_49")]; |
| tensor<int32, [3]> var_1200_perm_0 = const()[name = string("op_1200_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_55x = const()[name = string("concat_55x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_1200 = transpose(perm = var_1200_perm_0, x = attn_output_49)[name = string("transpose_264")]; |
| tensor<fp32, [?, 1024]> attn_output_51 = reshape(shape = concat_55x, x = var_1200)[name = string("attn_output_51")]; |
| tensor<fp32, [?, 1024]> attn_output_53 = linear(bias = net_wavlm_encoder_layers_6_attention_out_proj_bias, weight = net_wavlm_encoder_layers_6_attention_out_proj_weight_quantized, x = attn_output_51)[name = string("linear_47")]; |
| tensor<int32, [3]> concat_56x = const()[name = string("concat_56x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_55 = reshape(shape = concat_56x, x = attn_output_53)[name = string("attn_output_55")]; |
| tensor<int32, [3]> input_183_perm_0 = const()[name = string("input_183_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_183 = transpose(perm = input_183_perm_0, x = attn_output_55)[name = string("transpose_263")]; |
| tensor<fp32, [1, ?, 1024]> input_185 = add(x = input_177, y = input_183)[name = string("input_185")]; |
| tensor<int32, [1]> input_187_axes_0 = const()[name = string("input_187_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_187 = layer_norm(axes = input_187_axes_0, beta = net_wavlm_encoder_layers_6_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_6_final_layer_norm_weight, x = input_185)[name = string("input_187")]; |
| tensor<fp32, [1, ?, 4096]> input_189 = linear(bias = net_wavlm_encoder_layers_6_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_6_feed_forward_intermediate_dense_weight_quantized, x = input_187)[name = string("linear_48")]; |
| string input_191_mode_0 = const()[name = string("input_191_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_191 = gelu(mode = input_191_mode_0, x = input_189)[name = string("input_191")]; |
| tensor<fp32, [1, ?, 1024]> input_195 = linear(bias = net_wavlm_encoder_layers_6_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_6_feed_forward_output_dense_weight_quantized, x = input_191)[name = string("linear_49")]; |
| tensor<fp32, [1, ?, 1024]> input_197 = add(x = input_185, y = input_195)[name = string("input_197")]; |
| tensor<int32, [1]> hidden_states_63_axes_0 = const()[name = string("hidden_states_63_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_63 = layer_norm(axes = hidden_states_63_axes_0, beta = net_wavlm_encoder_layers_7_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_7_layer_norm_weight, x = input_197)[name = string("hidden_states_63")]; |
| tensor<int32, [4]> concat_57x = const()[name = string("concat_57x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_15 = reshape(shape = concat_57x, x = hidden_states_63)[name = string("gated_hidden_states_15")]; |
| tensor<int32, [4]> var_1261 = const()[name = string("op_1261"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_199 = transpose(perm = var_1261, x = gated_hidden_states_15)[name = string("transpose_262")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_29 = linear(bias = net_wavlm_encoder_layers_7_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_7_attention_gru_rel_pos_linear_weight, x = input_199)[name = string("linear_50")]; |
| tensor<int32, [5]> concat_58x = const()[name = string("concat_58x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1270 = reshape(shape = concat_58x, x = relative_position_proj_29)[name = string("op_1270")]; |
| tensor<int32, [1]> relative_position_proj_31_axes_0 = const()[name = string("relative_position_proj_31_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_31_keep_dims_0 = const()[name = string("relative_position_proj_31_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_31 = reduce_sum(axes = relative_position_proj_31_axes_0, keep_dims = relative_position_proj_31_keep_dims_0, x = var_1270)[name = string("relative_position_proj_31")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1273 = sigmoid(x = relative_position_proj_31)[name = string("op_1273")]; |
| tensor<int32, [2]> var_1274_split_sizes_0 = const()[name = string("op_1274_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1274_axis_0 = const()[name = string("op_1274_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1274_0, tensor<fp32, [1, 16, ?, 1]> var_1274_1 = split(axis = var_1274_axis_0, split_sizes = var_1274_split_sizes_0, x = var_1273)[name = string("op_1274")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1276 = mul(x = var_1274_1, y = net_wavlm_encoder_layers_7_attention_gru_rel_pos_const)[name = string("op_1276")]; |
| fp32 var_1277 = const()[name = string("op_1277"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1278 = sub(x = var_1276, y = var_1277)[name = string("op_1278")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1279 = mul(x = var_1274_0, y = var_1278)[name = string("op_1279")]; |
| fp32 var_1280 = const()[name = string("op_1280"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_15 = add(x = var_1279, y = var_1280)[name = string("gate_output_15")]; |
| tensor<int32, [3]> var_1285 = const()[name = string("op_1285"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1286 = reshape(shape = var_1285, x = gate_output_15)[name = string("op_1286")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_15 = mul(x = var_1286, y = position_bias)[name = string("gated_position_bias_15")]; |
| tensor<int32, [3]> query_15_perm_0 = const()[name = string("query_15_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_15 = transpose(perm = query_15_perm_0, x = hidden_states_63)[name = string("transpose_261")]; |
| tensor<fp32, [?, 1, 1024]> q_29 = linear(bias = net_wavlm_encoder_layers_7_attention_q_proj_bias, weight = net_wavlm_encoder_layers_7_attention_q_proj_weight_quantized, x = query_15)[name = string("linear_51")]; |
| tensor<fp32, [?, 1, 1024]> k_29 = linear(bias = net_wavlm_encoder_layers_7_attention_k_proj_bias, weight = net_wavlm_encoder_layers_7_attention_k_proj_weight_quantized, x = query_15)[name = string("linear_52")]; |
| tensor<fp32, [?, 1, 1024]> v_29 = linear(bias = net_wavlm_encoder_layers_7_attention_v_proj_bias, weight = net_wavlm_encoder_layers_7_attention_v_proj_weight_quantized, x = query_15)[name = string("linear_53")]; |
| tensor<int32, [3]> concat_60x = const()[name = string("concat_60x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1313 = reshape(shape = concat_60x, x = q_29)[name = string("op_1313")]; |
| tensor<int32, [3]> concat_61x = const()[name = string("concat_61x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1320 = reshape(shape = concat_61x, x = k_29)[name = string("op_1320")]; |
| tensor<int32, [3]> concat_62x = const()[name = string("concat_62x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1327 = reshape(shape = concat_62x, x = v_29)[name = string("op_1327")]; |
| tensor<int32, [3]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_1329 = const()[name = string("op_1329"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_15 = mul(x = var_1313, y = var_1329)[name = string("q_scaled_15")]; |
| bool input_201_bmm_transpose_x_0 = const()[name = string("input_201_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_201_bmm_transpose_y_0 = const()[name = string("input_201_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_111 = transpose(perm = transpose_111_perm_0, x = var_1320)[name = string("transpose_259")]; |
| tensor<fp32, [16, ?, 64]> transpose_110 = transpose(perm = transpose_110_perm_0, x = q_scaled_15)[name = string("transpose_260")]; |
| tensor<fp32, [16, ?, ?]> input_201_bmm = matmul(transpose_x = input_201_bmm_transpose_x_0, transpose_y = input_201_bmm_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("input_201_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_201 = add(x = gated_position_bias_15, y = input_201_bmm)[name = string("input_201")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_29 = softmax(axis = var_9, x = input_201)[name = string("attn_output_weights_29")]; |
| bool attn_output_57_transpose_x_0 = const()[name = string("attn_output_57_transpose_x_0"), val = bool(false)]; |
| bool attn_output_57_transpose_y_0 = const()[name = string("attn_output_57_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_31 = transpose(perm = v_31_perm_0, x = var_1327)[name = string("transpose_258")]; |
| tensor<fp32, [16, ?, 64]> attn_output_57 = matmul(transpose_x = attn_output_57_transpose_x_0, transpose_y = attn_output_57_transpose_y_0, x = attn_output_weights_29, y = v_31)[name = string("attn_output_57")]; |
| tensor<int32, [3]> var_1335_perm_0 = const()[name = string("op_1335_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_63x = const()[name = string("concat_63x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_1335 = transpose(perm = var_1335_perm_0, x = attn_output_57)[name = string("transpose_257")]; |
| tensor<fp32, [?, 1024]> attn_output_59 = reshape(shape = concat_63x, x = var_1335)[name = string("attn_output_59")]; |
| tensor<fp32, [?, 1024]> attn_output_61 = linear(bias = net_wavlm_encoder_layers_7_attention_out_proj_bias, weight = net_wavlm_encoder_layers_7_attention_out_proj_weight_quantized, x = attn_output_59)[name = string("linear_54")]; |
| tensor<int32, [3]> concat_64x = const()[name = string("concat_64x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_63 = reshape(shape = concat_64x, x = attn_output_61)[name = string("attn_output_63")]; |
| tensor<int32, [3]> input_203_perm_0 = const()[name = string("input_203_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_203 = transpose(perm = input_203_perm_0, x = attn_output_63)[name = string("transpose_256")]; |
| tensor<fp32, [1, ?, 1024]> input_205 = add(x = input_197, y = input_203)[name = string("input_205")]; |
| tensor<int32, [1]> input_207_axes_0 = const()[name = string("input_207_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_207 = layer_norm(axes = input_207_axes_0, beta = net_wavlm_encoder_layers_7_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_7_final_layer_norm_weight, x = input_205)[name = string("input_207")]; |
| tensor<fp32, [1, ?, 4096]> input_209 = linear(bias = net_wavlm_encoder_layers_7_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_7_feed_forward_intermediate_dense_weight_quantized, x = input_207)[name = string("linear_55")]; |
| string input_211_mode_0 = const()[name = string("input_211_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_211 = gelu(mode = input_211_mode_0, x = input_209)[name = string("input_211")]; |
| tensor<fp32, [1, ?, 1024]> input_215 = linear(bias = net_wavlm_encoder_layers_7_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_7_feed_forward_output_dense_weight_quantized, x = input_211)[name = string("linear_56")]; |
| tensor<fp32, [1, ?, 1024]> input_217 = add(x = input_205, y = input_215)[name = string("input_217")]; |
| tensor<int32, [1]> hidden_states_67_axes_0 = const()[name = string("hidden_states_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_67 = layer_norm(axes = hidden_states_67_axes_0, beta = net_wavlm_encoder_layers_8_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_8_layer_norm_weight, x = input_217)[name = string("hidden_states_67")]; |
| tensor<int32, [4]> concat_65x = const()[name = string("concat_65x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_17 = reshape(shape = concat_65x, x = hidden_states_67)[name = string("gated_hidden_states_17")]; |
| tensor<int32, [4]> var_1396 = const()[name = string("op_1396"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_219 = transpose(perm = var_1396, x = gated_hidden_states_17)[name = string("transpose_255")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_33 = linear(bias = net_wavlm_encoder_layers_8_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_8_attention_gru_rel_pos_linear_weight, x = input_219)[name = string("linear_57")]; |
| tensor<int32, [5]> concat_66x = const()[name = string("concat_66x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1405 = reshape(shape = concat_66x, x = relative_position_proj_33)[name = string("op_1405")]; |
| tensor<int32, [1]> relative_position_proj_35_axes_0 = const()[name = string("relative_position_proj_35_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_35_keep_dims_0 = const()[name = string("relative_position_proj_35_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_35 = reduce_sum(axes = relative_position_proj_35_axes_0, keep_dims = relative_position_proj_35_keep_dims_0, x = var_1405)[name = string("relative_position_proj_35")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1408 = sigmoid(x = relative_position_proj_35)[name = string("op_1408")]; |
| tensor<int32, [2]> var_1409_split_sizes_0 = const()[name = string("op_1409_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1409_axis_0 = const()[name = string("op_1409_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1409_0, tensor<fp32, [1, 16, ?, 1]> var_1409_1 = split(axis = var_1409_axis_0, split_sizes = var_1409_split_sizes_0, x = var_1408)[name = string("op_1409")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1411 = mul(x = var_1409_1, y = net_wavlm_encoder_layers_8_attention_gru_rel_pos_const)[name = string("op_1411")]; |
| fp32 var_1412 = const()[name = string("op_1412"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1413 = sub(x = var_1411, y = var_1412)[name = string("op_1413")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1414 = mul(x = var_1409_0, y = var_1413)[name = string("op_1414")]; |
| fp32 var_1415 = const()[name = string("op_1415"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_17 = add(x = var_1414, y = var_1415)[name = string("gate_output_17")]; |
| tensor<int32, [3]> var_1420 = const()[name = string("op_1420"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1421 = reshape(shape = var_1420, x = gate_output_17)[name = string("op_1421")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_17 = mul(x = var_1421, y = position_bias)[name = string("gated_position_bias_17")]; |
| tensor<int32, [3]> query_17_perm_0 = const()[name = string("query_17_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_17 = transpose(perm = query_17_perm_0, x = hidden_states_67)[name = string("transpose_254")]; |
| tensor<fp32, [?, 1, 1024]> q_33 = linear(bias = net_wavlm_encoder_layers_8_attention_q_proj_bias, weight = net_wavlm_encoder_layers_8_attention_q_proj_weight_quantized, x = query_17)[name = string("linear_58")]; |
| tensor<fp32, [?, 1, 1024]> k_33 = linear(bias = net_wavlm_encoder_layers_8_attention_k_proj_bias, weight = net_wavlm_encoder_layers_8_attention_k_proj_weight_quantized, x = query_17)[name = string("linear_59")]; |
| tensor<fp32, [?, 1, 1024]> v_33 = linear(bias = net_wavlm_encoder_layers_8_attention_v_proj_bias, weight = net_wavlm_encoder_layers_8_attention_v_proj_weight_quantized, x = query_17)[name = string("linear_60")]; |
| tensor<int32, [3]> concat_68x = const()[name = string("concat_68x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1448 = reshape(shape = concat_68x, x = q_33)[name = string("op_1448")]; |
| tensor<int32, [3]> concat_69x = const()[name = string("concat_69x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1455 = reshape(shape = concat_69x, x = k_33)[name = string("op_1455")]; |
| tensor<int32, [3]> concat_70x = const()[name = string("concat_70x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1462 = reshape(shape = concat_70x, x = v_33)[name = string("op_1462")]; |
| tensor<int32, [3]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_1464 = const()[name = string("op_1464"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_17 = mul(x = var_1448, y = var_1464)[name = string("q_scaled_17")]; |
| bool input_221_bmm_transpose_x_0 = const()[name = string("input_221_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_221_bmm_transpose_y_0 = const()[name = string("input_221_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_113 = transpose(perm = transpose_113_perm_0, x = var_1455)[name = string("transpose_252")]; |
| tensor<fp32, [16, ?, 64]> transpose_112 = transpose(perm = transpose_112_perm_0, x = q_scaled_17)[name = string("transpose_253")]; |
| tensor<fp32, [16, ?, ?]> input_221_bmm = matmul(transpose_x = input_221_bmm_transpose_x_0, transpose_y = input_221_bmm_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("input_221_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_221 = add(x = gated_position_bias_17, y = input_221_bmm)[name = string("input_221")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_33 = softmax(axis = var_9, x = input_221)[name = string("attn_output_weights_33")]; |
| bool attn_output_65_transpose_x_0 = const()[name = string("attn_output_65_transpose_x_0"), val = bool(false)]; |
| bool attn_output_65_transpose_y_0 = const()[name = string("attn_output_65_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_35 = transpose(perm = v_35_perm_0, x = var_1462)[name = string("transpose_251")]; |
| tensor<fp32, [16, ?, 64]> attn_output_65 = matmul(transpose_x = attn_output_65_transpose_x_0, transpose_y = attn_output_65_transpose_y_0, x = attn_output_weights_33, y = v_35)[name = string("attn_output_65")]; |
| tensor<int32, [3]> var_1470_perm_0 = const()[name = string("op_1470_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_71x = const()[name = string("concat_71x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_1470 = transpose(perm = var_1470_perm_0, x = attn_output_65)[name = string("transpose_250")]; |
| tensor<fp32, [?, 1024]> attn_output_67 = reshape(shape = concat_71x, x = var_1470)[name = string("attn_output_67")]; |
| tensor<fp32, [?, 1024]> attn_output_69 = linear(bias = net_wavlm_encoder_layers_8_attention_out_proj_bias, weight = net_wavlm_encoder_layers_8_attention_out_proj_weight_quantized, x = attn_output_67)[name = string("linear_61")]; |
| tensor<int32, [3]> concat_72x = const()[name = string("concat_72x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_71 = reshape(shape = concat_72x, x = attn_output_69)[name = string("attn_output_71")]; |
| tensor<int32, [3]> input_223_perm_0 = const()[name = string("input_223_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_223 = transpose(perm = input_223_perm_0, x = attn_output_71)[name = string("transpose_249")]; |
| tensor<fp32, [1, ?, 1024]> input_225 = add(x = input_217, y = input_223)[name = string("input_225")]; |
| tensor<int32, [1]> input_227_axes_0 = const()[name = string("input_227_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_227 = layer_norm(axes = input_227_axes_0, beta = net_wavlm_encoder_layers_8_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_8_final_layer_norm_weight, x = input_225)[name = string("input_227")]; |
| tensor<fp32, [1, ?, 4096]> input_229 = linear(bias = net_wavlm_encoder_layers_8_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_8_feed_forward_intermediate_dense_weight_quantized, x = input_227)[name = string("linear_62")]; |
| string input_231_mode_0 = const()[name = string("input_231_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_231 = gelu(mode = input_231_mode_0, x = input_229)[name = string("input_231")]; |
| tensor<fp32, [1, ?, 1024]> input_235 = linear(bias = net_wavlm_encoder_layers_8_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_8_feed_forward_output_dense_weight_quantized, x = input_231)[name = string("linear_63")]; |
| tensor<fp32, [1, ?, 1024]> input_237 = add(x = input_225, y = input_235)[name = string("input_237")]; |
| tensor<int32, [1]> hidden_states_71_axes_0 = const()[name = string("hidden_states_71_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_71 = layer_norm(axes = hidden_states_71_axes_0, beta = net_wavlm_encoder_layers_9_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_9_layer_norm_weight, x = input_237)[name = string("hidden_states_71")]; |
| tensor<int32, [4]> concat_73x = const()[name = string("concat_73x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_19 = reshape(shape = concat_73x, x = hidden_states_71)[name = string("gated_hidden_states_19")]; |
| tensor<int32, [4]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_239 = transpose(perm = var_1531, x = gated_hidden_states_19)[name = string("transpose_248")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_37 = linear(bias = net_wavlm_encoder_layers_9_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_9_attention_gru_rel_pos_linear_weight, x = input_239)[name = string("linear_64")]; |
| tensor<int32, [5]> concat_74x = const()[name = string("concat_74x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1540 = reshape(shape = concat_74x, x = relative_position_proj_37)[name = string("op_1540")]; |
| tensor<int32, [1]> relative_position_proj_39_axes_0 = const()[name = string("relative_position_proj_39_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_39_keep_dims_0 = const()[name = string("relative_position_proj_39_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_39 = reduce_sum(axes = relative_position_proj_39_axes_0, keep_dims = relative_position_proj_39_keep_dims_0, x = var_1540)[name = string("relative_position_proj_39")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1543 = sigmoid(x = relative_position_proj_39)[name = string("op_1543")]; |
| tensor<int32, [2]> var_1544_split_sizes_0 = const()[name = string("op_1544_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1544_axis_0 = const()[name = string("op_1544_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1544_0, tensor<fp32, [1, 16, ?, 1]> var_1544_1 = split(axis = var_1544_axis_0, split_sizes = var_1544_split_sizes_0, x = var_1543)[name = string("op_1544")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1546 = mul(x = var_1544_1, y = net_wavlm_encoder_layers_9_attention_gru_rel_pos_const)[name = string("op_1546")]; |
| fp32 var_1547 = const()[name = string("op_1547"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1548 = sub(x = var_1546, y = var_1547)[name = string("op_1548")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1549 = mul(x = var_1544_0, y = var_1548)[name = string("op_1549")]; |
| fp32 var_1550 = const()[name = string("op_1550"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_19 = add(x = var_1549, y = var_1550)[name = string("gate_output_19")]; |
| tensor<int32, [3]> var_1555 = const()[name = string("op_1555"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1556 = reshape(shape = var_1555, x = gate_output_19)[name = string("op_1556")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_19 = mul(x = var_1556, y = position_bias)[name = string("gated_position_bias_19")]; |
| tensor<int32, [3]> query_19_perm_0 = const()[name = string("query_19_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_19 = transpose(perm = query_19_perm_0, x = hidden_states_71)[name = string("transpose_247")]; |
| tensor<fp32, [?, 1, 1024]> q_37 = linear(bias = net_wavlm_encoder_layers_9_attention_q_proj_bias, weight = net_wavlm_encoder_layers_9_attention_q_proj_weight_quantized, x = query_19)[name = string("linear_65")]; |
| tensor<fp32, [?, 1, 1024]> k_37 = linear(bias = net_wavlm_encoder_layers_9_attention_k_proj_bias, weight = net_wavlm_encoder_layers_9_attention_k_proj_weight_quantized, x = query_19)[name = string("linear_66")]; |
| tensor<fp32, [?, 1, 1024]> v_37 = linear(bias = net_wavlm_encoder_layers_9_attention_v_proj_bias, weight = net_wavlm_encoder_layers_9_attention_v_proj_weight_quantized, x = query_19)[name = string("linear_67")]; |
| tensor<int32, [3]> concat_76x = const()[name = string("concat_76x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1583 = reshape(shape = concat_76x, x = q_37)[name = string("op_1583")]; |
| tensor<int32, [3]> concat_77x = const()[name = string("concat_77x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1590 = reshape(shape = concat_77x, x = k_37)[name = string("op_1590")]; |
| tensor<int32, [3]> concat_78x = const()[name = string("concat_78x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1597 = reshape(shape = concat_78x, x = v_37)[name = string("op_1597")]; |
| tensor<int32, [3]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_1599 = const()[name = string("op_1599"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_19 = mul(x = var_1583, y = var_1599)[name = string("q_scaled_19")]; |
| bool input_241_bmm_transpose_x_0 = const()[name = string("input_241_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_241_bmm_transpose_y_0 = const()[name = string("input_241_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_115 = transpose(perm = transpose_115_perm_0, x = var_1590)[name = string("transpose_245")]; |
| tensor<fp32, [16, ?, 64]> transpose_114 = transpose(perm = transpose_114_perm_0, x = q_scaled_19)[name = string("transpose_246")]; |
| tensor<fp32, [16, ?, ?]> input_241_bmm = matmul(transpose_x = input_241_bmm_transpose_x_0, transpose_y = input_241_bmm_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("input_241_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_241 = add(x = gated_position_bias_19, y = input_241_bmm)[name = string("input_241")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_37 = softmax(axis = var_9, x = input_241)[name = string("attn_output_weights_37")]; |
| bool attn_output_73_transpose_x_0 = const()[name = string("attn_output_73_transpose_x_0"), val = bool(false)]; |
| bool attn_output_73_transpose_y_0 = const()[name = string("attn_output_73_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_39 = transpose(perm = v_39_perm_0, x = var_1597)[name = string("transpose_244")]; |
| tensor<fp32, [16, ?, 64]> attn_output_73 = matmul(transpose_x = attn_output_73_transpose_x_0, transpose_y = attn_output_73_transpose_y_0, x = attn_output_weights_37, y = v_39)[name = string("attn_output_73")]; |
| tensor<int32, [3]> var_1605_perm_0 = const()[name = string("op_1605_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_79x = const()[name = string("concat_79x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_1605 = transpose(perm = var_1605_perm_0, x = attn_output_73)[name = string("transpose_243")]; |
| tensor<fp32, [?, 1024]> attn_output_75 = reshape(shape = concat_79x, x = var_1605)[name = string("attn_output_75")]; |
| tensor<fp32, [?, 1024]> attn_output_77 = linear(bias = net_wavlm_encoder_layers_9_attention_out_proj_bias, weight = net_wavlm_encoder_layers_9_attention_out_proj_weight_quantized, x = attn_output_75)[name = string("linear_68")]; |
| tensor<int32, [3]> concat_80x = const()[name = string("concat_80x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_79 = reshape(shape = concat_80x, x = attn_output_77)[name = string("attn_output_79")]; |
| tensor<int32, [3]> input_243_perm_0 = const()[name = string("input_243_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_243 = transpose(perm = input_243_perm_0, x = attn_output_79)[name = string("transpose_242")]; |
| tensor<fp32, [1, ?, 1024]> input_245 = add(x = input_237, y = input_243)[name = string("input_245")]; |
| tensor<int32, [1]> input_247_axes_0 = const()[name = string("input_247_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_247 = layer_norm(axes = input_247_axes_0, beta = net_wavlm_encoder_layers_9_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_9_final_layer_norm_weight, x = input_245)[name = string("input_247")]; |
| tensor<fp32, [1, ?, 4096]> input_249 = linear(bias = net_wavlm_encoder_layers_9_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_9_feed_forward_intermediate_dense_weight_quantized, x = input_247)[name = string("linear_69")]; |
| string input_251_mode_0 = const()[name = string("input_251_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_251 = gelu(mode = input_251_mode_0, x = input_249)[name = string("input_251")]; |
| tensor<fp32, [1, ?, 1024]> input_255 = linear(bias = net_wavlm_encoder_layers_9_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_9_feed_forward_output_dense_weight_quantized, x = input_251)[name = string("linear_70")]; |
| tensor<fp32, [1, ?, 1024]> input_257 = add(x = input_245, y = input_255)[name = string("input_257")]; |
| tensor<int32, [1]> hidden_states_75_axes_0 = const()[name = string("hidden_states_75_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_75 = layer_norm(axes = hidden_states_75_axes_0, beta = net_wavlm_encoder_layers_10_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_10_layer_norm_weight, x = input_257)[name = string("hidden_states_75")]; |
| tensor<int32, [4]> concat_81x = const()[name = string("concat_81x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_21 = reshape(shape = concat_81x, x = hidden_states_75)[name = string("gated_hidden_states_21")]; |
| tensor<int32, [4]> var_1666 = const()[name = string("op_1666"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_259 = transpose(perm = var_1666, x = gated_hidden_states_21)[name = string("transpose_241")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_41 = linear(bias = net_wavlm_encoder_layers_10_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_10_attention_gru_rel_pos_linear_weight, x = input_259)[name = string("linear_71")]; |
| tensor<int32, [5]> concat_82x = const()[name = string("concat_82x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1675 = reshape(shape = concat_82x, x = relative_position_proj_41)[name = string("op_1675")]; |
| tensor<int32, [1]> relative_position_proj_43_axes_0 = const()[name = string("relative_position_proj_43_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_43_keep_dims_0 = const()[name = string("relative_position_proj_43_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_43 = reduce_sum(axes = relative_position_proj_43_axes_0, keep_dims = relative_position_proj_43_keep_dims_0, x = var_1675)[name = string("relative_position_proj_43")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1678 = sigmoid(x = relative_position_proj_43)[name = string("op_1678")]; |
| tensor<int32, [2]> var_1679_split_sizes_0 = const()[name = string("op_1679_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1679_axis_0 = const()[name = string("op_1679_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1679_0, tensor<fp32, [1, 16, ?, 1]> var_1679_1 = split(axis = var_1679_axis_0, split_sizes = var_1679_split_sizes_0, x = var_1678)[name = string("op_1679")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1681 = mul(x = var_1679_1, y = net_wavlm_encoder_layers_10_attention_gru_rel_pos_const)[name = string("op_1681")]; |
| fp32 var_1682 = const()[name = string("op_1682"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1683 = sub(x = var_1681, y = var_1682)[name = string("op_1683")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1684 = mul(x = var_1679_0, y = var_1683)[name = string("op_1684")]; |
| fp32 var_1685 = const()[name = string("op_1685"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_21 = add(x = var_1684, y = var_1685)[name = string("gate_output_21")]; |
| tensor<int32, [3]> var_1690 = const()[name = string("op_1690"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1691 = reshape(shape = var_1690, x = gate_output_21)[name = string("op_1691")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_21 = mul(x = var_1691, y = position_bias)[name = string("gated_position_bias_21")]; |
| tensor<int32, [3]> query_21_perm_0 = const()[name = string("query_21_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_21 = transpose(perm = query_21_perm_0, x = hidden_states_75)[name = string("transpose_240")]; |
| tensor<fp32, [?, 1, 1024]> q_41 = linear(bias = net_wavlm_encoder_layers_10_attention_q_proj_bias, weight = net_wavlm_encoder_layers_10_attention_q_proj_weight_quantized, x = query_21)[name = string("linear_72")]; |
| tensor<fp32, [?, 1, 1024]> k_41 = linear(bias = net_wavlm_encoder_layers_10_attention_k_proj_bias, weight = net_wavlm_encoder_layers_10_attention_k_proj_weight_quantized, x = query_21)[name = string("linear_73")]; |
| tensor<fp32, [?, 1, 1024]> v_41 = linear(bias = net_wavlm_encoder_layers_10_attention_v_proj_bias, weight = net_wavlm_encoder_layers_10_attention_v_proj_weight_quantized, x = query_21)[name = string("linear_74")]; |
| tensor<int32, [3]> concat_84x = const()[name = string("concat_84x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1718 = reshape(shape = concat_84x, x = q_41)[name = string("op_1718")]; |
| tensor<int32, [3]> concat_85x = const()[name = string("concat_85x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1725 = reshape(shape = concat_85x, x = k_41)[name = string("op_1725")]; |
| tensor<int32, [3]> concat_86x = const()[name = string("concat_86x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1732 = reshape(shape = concat_86x, x = v_41)[name = string("op_1732")]; |
| tensor<int32, [3]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_1734 = const()[name = string("op_1734"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_21 = mul(x = var_1718, y = var_1734)[name = string("q_scaled_21")]; |
| bool input_261_bmm_transpose_x_0 = const()[name = string("input_261_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_261_bmm_transpose_y_0 = const()[name = string("input_261_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_117 = transpose(perm = transpose_117_perm_0, x = var_1725)[name = string("transpose_238")]; |
| tensor<fp32, [16, ?, 64]> transpose_116 = transpose(perm = transpose_116_perm_0, x = q_scaled_21)[name = string("transpose_239")]; |
| tensor<fp32, [16, ?, ?]> input_261_bmm = matmul(transpose_x = input_261_bmm_transpose_x_0, transpose_y = input_261_bmm_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("input_261_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_261 = add(x = gated_position_bias_21, y = input_261_bmm)[name = string("input_261")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_41 = softmax(axis = var_9, x = input_261)[name = string("attn_output_weights_41")]; |
| bool attn_output_81_transpose_x_0 = const()[name = string("attn_output_81_transpose_x_0"), val = bool(false)]; |
| bool attn_output_81_transpose_y_0 = const()[name = string("attn_output_81_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_43 = transpose(perm = v_43_perm_0, x = var_1732)[name = string("transpose_237")]; |
| tensor<fp32, [16, ?, 64]> attn_output_81 = matmul(transpose_x = attn_output_81_transpose_x_0, transpose_y = attn_output_81_transpose_y_0, x = attn_output_weights_41, y = v_43)[name = string("attn_output_81")]; |
| tensor<int32, [3]> var_1740_perm_0 = const()[name = string("op_1740_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_87x = const()[name = string("concat_87x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_1740 = transpose(perm = var_1740_perm_0, x = attn_output_81)[name = string("transpose_236")]; |
| tensor<fp32, [?, 1024]> attn_output_83 = reshape(shape = concat_87x, x = var_1740)[name = string("attn_output_83")]; |
| tensor<fp32, [?, 1024]> attn_output_85 = linear(bias = net_wavlm_encoder_layers_10_attention_out_proj_bias, weight = net_wavlm_encoder_layers_10_attention_out_proj_weight_quantized, x = attn_output_83)[name = string("linear_75")]; |
| tensor<int32, [3]> concat_88x = const()[name = string("concat_88x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_87 = reshape(shape = concat_88x, x = attn_output_85)[name = string("attn_output_87")]; |
| tensor<int32, [3]> input_263_perm_0 = const()[name = string("input_263_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_263 = transpose(perm = input_263_perm_0, x = attn_output_87)[name = string("transpose_235")]; |
| tensor<fp32, [1, ?, 1024]> input_265 = add(x = input_257, y = input_263)[name = string("input_265")]; |
| tensor<int32, [1]> input_267_axes_0 = const()[name = string("input_267_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_267 = layer_norm(axes = input_267_axes_0, beta = net_wavlm_encoder_layers_10_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_10_final_layer_norm_weight, x = input_265)[name = string("input_267")]; |
| tensor<fp32, [1, ?, 4096]> input_269 = linear(bias = net_wavlm_encoder_layers_10_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_10_feed_forward_intermediate_dense_weight_quantized, x = input_267)[name = string("linear_76")]; |
| string input_271_mode_0 = const()[name = string("input_271_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_271 = gelu(mode = input_271_mode_0, x = input_269)[name = string("input_271")]; |
| tensor<fp32, [1, ?, 1024]> input_275 = linear(bias = net_wavlm_encoder_layers_10_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_10_feed_forward_output_dense_weight_quantized, x = input_271)[name = string("linear_77")]; |
| tensor<fp32, [1, ?, 1024]> input_277 = add(x = input_265, y = input_275)[name = string("input_277")]; |
| tensor<int32, [1]> hidden_states_79_axes_0 = const()[name = string("hidden_states_79_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_79 = layer_norm(axes = hidden_states_79_axes_0, beta = net_wavlm_encoder_layers_11_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_11_layer_norm_weight, x = input_277)[name = string("hidden_states_79")]; |
| tensor<int32, [4]> concat_89x = const()[name = string("concat_89x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_23 = reshape(shape = concat_89x, x = hidden_states_79)[name = string("gated_hidden_states_23")]; |
| tensor<int32, [4]> var_1801 = const()[name = string("op_1801"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_279 = transpose(perm = var_1801, x = gated_hidden_states_23)[name = string("transpose_234")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_45 = linear(bias = net_wavlm_encoder_layers_11_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_11_attention_gru_rel_pos_linear_weight, x = input_279)[name = string("linear_78")]; |
| tensor<int32, [5]> concat_90x = const()[name = string("concat_90x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1810 = reshape(shape = concat_90x, x = relative_position_proj_45)[name = string("op_1810")]; |
| tensor<int32, [1]> relative_position_proj_47_axes_0 = const()[name = string("relative_position_proj_47_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_47_keep_dims_0 = const()[name = string("relative_position_proj_47_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_47 = reduce_sum(axes = relative_position_proj_47_axes_0, keep_dims = relative_position_proj_47_keep_dims_0, x = var_1810)[name = string("relative_position_proj_47")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1813 = sigmoid(x = relative_position_proj_47)[name = string("op_1813")]; |
| tensor<int32, [2]> var_1814_split_sizes_0 = const()[name = string("op_1814_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1814_axis_0 = const()[name = string("op_1814_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1814_0, tensor<fp32, [1, 16, ?, 1]> var_1814_1 = split(axis = var_1814_axis_0, split_sizes = var_1814_split_sizes_0, x = var_1813)[name = string("op_1814")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1816 = mul(x = var_1814_1, y = net_wavlm_encoder_layers_11_attention_gru_rel_pos_const)[name = string("op_1816")]; |
| fp32 var_1817 = const()[name = string("op_1817"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1818 = sub(x = var_1816, y = var_1817)[name = string("op_1818")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1819 = mul(x = var_1814_0, y = var_1818)[name = string("op_1819")]; |
| fp32 var_1820 = const()[name = string("op_1820"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_23 = add(x = var_1819, y = var_1820)[name = string("gate_output_23")]; |
| tensor<int32, [3]> var_1825 = const()[name = string("op_1825"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1826 = reshape(shape = var_1825, x = gate_output_23)[name = string("op_1826")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_23 = mul(x = var_1826, y = position_bias)[name = string("gated_position_bias_23")]; |
| tensor<int32, [3]> query_23_perm_0 = const()[name = string("query_23_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_23 = transpose(perm = query_23_perm_0, x = hidden_states_79)[name = string("transpose_233")]; |
| tensor<fp32, [?, 1, 1024]> q_45 = linear(bias = net_wavlm_encoder_layers_11_attention_q_proj_bias, weight = net_wavlm_encoder_layers_11_attention_q_proj_weight_quantized, x = query_23)[name = string("linear_79")]; |
| tensor<fp32, [?, 1, 1024]> k_45 = linear(bias = net_wavlm_encoder_layers_11_attention_k_proj_bias, weight = net_wavlm_encoder_layers_11_attention_k_proj_weight_quantized, x = query_23)[name = string("linear_80")]; |
| tensor<fp32, [?, 1, 1024]> v_45 = linear(bias = net_wavlm_encoder_layers_11_attention_v_proj_bias, weight = net_wavlm_encoder_layers_11_attention_v_proj_weight_quantized, x = query_23)[name = string("linear_81")]; |
| tensor<int32, [3]> concat_92x = const()[name = string("concat_92x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1853 = reshape(shape = concat_92x, x = q_45)[name = string("op_1853")]; |
| tensor<int32, [3]> concat_93x = const()[name = string("concat_93x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1860 = reshape(shape = concat_93x, x = k_45)[name = string("op_1860")]; |
| tensor<int32, [3]> concat_94x = const()[name = string("concat_94x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1867 = reshape(shape = concat_94x, x = v_45)[name = string("op_1867")]; |
| tensor<int32, [3]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_1869 = const()[name = string("op_1869"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_23 = mul(x = var_1853, y = var_1869)[name = string("q_scaled_23")]; |
| bool input_281_bmm_transpose_x_0 = const()[name = string("input_281_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_281_bmm_transpose_y_0 = const()[name = string("input_281_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_119 = transpose(perm = transpose_119_perm_0, x = var_1860)[name = string("transpose_231")]; |
| tensor<fp32, [16, ?, 64]> transpose_118 = transpose(perm = transpose_118_perm_0, x = q_scaled_23)[name = string("transpose_232")]; |
| tensor<fp32, [16, ?, ?]> input_281_bmm = matmul(transpose_x = input_281_bmm_transpose_x_0, transpose_y = input_281_bmm_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("input_281_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_281 = add(x = gated_position_bias_23, y = input_281_bmm)[name = string("input_281")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_45 = softmax(axis = var_9, x = input_281)[name = string("attn_output_weights_45")]; |
| bool attn_output_89_transpose_x_0 = const()[name = string("attn_output_89_transpose_x_0"), val = bool(false)]; |
| bool attn_output_89_transpose_y_0 = const()[name = string("attn_output_89_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_47 = transpose(perm = v_47_perm_0, x = var_1867)[name = string("transpose_230")]; |
| tensor<fp32, [16, ?, 64]> attn_output_89 = matmul(transpose_x = attn_output_89_transpose_x_0, transpose_y = attn_output_89_transpose_y_0, x = attn_output_weights_45, y = v_47)[name = string("attn_output_89")]; |
| tensor<int32, [3]> var_1875_perm_0 = const()[name = string("op_1875_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_95x = const()[name = string("concat_95x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_1875 = transpose(perm = var_1875_perm_0, x = attn_output_89)[name = string("transpose_229")]; |
| tensor<fp32, [?, 1024]> attn_output_91 = reshape(shape = concat_95x, x = var_1875)[name = string("attn_output_91")]; |
| tensor<fp32, [?, 1024]> attn_output_93 = linear(bias = net_wavlm_encoder_layers_11_attention_out_proj_bias, weight = net_wavlm_encoder_layers_11_attention_out_proj_weight_quantized, x = attn_output_91)[name = string("linear_82")]; |
| tensor<int32, [3]> concat_96x = const()[name = string("concat_96x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_95 = reshape(shape = concat_96x, x = attn_output_93)[name = string("attn_output_95")]; |
| tensor<int32, [3]> input_283_perm_0 = const()[name = string("input_283_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_283 = transpose(perm = input_283_perm_0, x = attn_output_95)[name = string("transpose_228")]; |
| tensor<fp32, [1, ?, 1024]> input_285 = add(x = input_277, y = input_283)[name = string("input_285")]; |
| tensor<int32, [1]> input_287_axes_0 = const()[name = string("input_287_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_287 = layer_norm(axes = input_287_axes_0, beta = net_wavlm_encoder_layers_11_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_11_final_layer_norm_weight, x = input_285)[name = string("input_287")]; |
| tensor<fp32, [1, ?, 4096]> input_289 = linear(bias = net_wavlm_encoder_layers_11_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_11_feed_forward_intermediate_dense_weight_quantized, x = input_287)[name = string("linear_83")]; |
| string input_291_mode_0 = const()[name = string("input_291_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_291 = gelu(mode = input_291_mode_0, x = input_289)[name = string("input_291")]; |
| tensor<fp32, [1, ?, 1024]> input_295 = linear(bias = net_wavlm_encoder_layers_11_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_11_feed_forward_output_dense_weight_quantized, x = input_291)[name = string("linear_84")]; |
| tensor<fp32, [1, ?, 1024]> input_297 = add(x = input_285, y = input_295)[name = string("input_297")]; |
| tensor<int32, [1]> hidden_states_83_axes_0 = const()[name = string("hidden_states_83_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_83 = layer_norm(axes = hidden_states_83_axes_0, beta = net_wavlm_encoder_layers_12_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_12_layer_norm_weight, x = input_297)[name = string("hidden_states_83")]; |
| tensor<int32, [4]> concat_97x = const()[name = string("concat_97x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_25 = reshape(shape = concat_97x, x = hidden_states_83)[name = string("gated_hidden_states_25")]; |
| tensor<int32, [4]> var_1936 = const()[name = string("op_1936"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_299 = transpose(perm = var_1936, x = gated_hidden_states_25)[name = string("transpose_227")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_49 = linear(bias = net_wavlm_encoder_layers_12_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_12_attention_gru_rel_pos_linear_weight, x = input_299)[name = string("linear_85")]; |
| tensor<int32, [5]> concat_98x = const()[name = string("concat_98x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_1945 = reshape(shape = concat_98x, x = relative_position_proj_49)[name = string("op_1945")]; |
| tensor<int32, [1]> relative_position_proj_51_axes_0 = const()[name = string("relative_position_proj_51_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_51_keep_dims_0 = const()[name = string("relative_position_proj_51_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_51 = reduce_sum(axes = relative_position_proj_51_axes_0, keep_dims = relative_position_proj_51_keep_dims_0, x = var_1945)[name = string("relative_position_proj_51")]; |
| tensor<fp32, [1, 16, ?, 2]> var_1948 = sigmoid(x = relative_position_proj_51)[name = string("op_1948")]; |
| tensor<int32, [2]> var_1949_split_sizes_0 = const()[name = string("op_1949_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_1949_axis_0 = const()[name = string("op_1949_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1949_0, tensor<fp32, [1, 16, ?, 1]> var_1949_1 = split(axis = var_1949_axis_0, split_sizes = var_1949_split_sizes_0, x = var_1948)[name = string("op_1949")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1951 = mul(x = var_1949_1, y = net_wavlm_encoder_layers_12_attention_gru_rel_pos_const)[name = string("op_1951")]; |
| fp32 var_1952 = const()[name = string("op_1952"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_1953 = sub(x = var_1951, y = var_1952)[name = string("op_1953")]; |
| tensor<fp32, [1, 16, ?, 1]> var_1954 = mul(x = var_1949_0, y = var_1953)[name = string("op_1954")]; |
| fp32 var_1955 = const()[name = string("op_1955"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_25 = add(x = var_1954, y = var_1955)[name = string("gate_output_25")]; |
| tensor<int32, [3]> var_1960 = const()[name = string("op_1960"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_1961 = reshape(shape = var_1960, x = gate_output_25)[name = string("op_1961")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_25 = mul(x = var_1961, y = position_bias)[name = string("gated_position_bias_25")]; |
| tensor<int32, [3]> query_25_perm_0 = const()[name = string("query_25_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_25 = transpose(perm = query_25_perm_0, x = hidden_states_83)[name = string("transpose_226")]; |
| tensor<fp32, [?, 1, 1024]> q_49 = linear(bias = net_wavlm_encoder_layers_12_attention_q_proj_bias, weight = net_wavlm_encoder_layers_12_attention_q_proj_weight_quantized, x = query_25)[name = string("linear_86")]; |
| tensor<fp32, [?, 1, 1024]> k_49 = linear(bias = net_wavlm_encoder_layers_12_attention_k_proj_bias, weight = net_wavlm_encoder_layers_12_attention_k_proj_weight_quantized, x = query_25)[name = string("linear_87")]; |
| tensor<fp32, [?, 1, 1024]> v_49 = linear(bias = net_wavlm_encoder_layers_12_attention_v_proj_bias, weight = net_wavlm_encoder_layers_12_attention_v_proj_weight_quantized, x = query_25)[name = string("linear_88")]; |
| tensor<int32, [3]> concat_100x = const()[name = string("concat_100x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1988 = reshape(shape = concat_100x, x = q_49)[name = string("op_1988")]; |
| tensor<int32, [3]> concat_101x = const()[name = string("concat_101x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_1995 = reshape(shape = concat_101x, x = k_49)[name = string("op_1995")]; |
| tensor<int32, [3]> concat_102x = const()[name = string("concat_102x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2002 = reshape(shape = concat_102x, x = v_49)[name = string("op_2002")]; |
| tensor<int32, [3]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2004 = const()[name = string("op_2004"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_25 = mul(x = var_1988, y = var_2004)[name = string("q_scaled_25")]; |
| bool input_301_bmm_transpose_x_0 = const()[name = string("input_301_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_301_bmm_transpose_y_0 = const()[name = string("input_301_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_120_perm_0 = const()[name = string("transpose_120_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_121_perm_0 = const()[name = string("transpose_121_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_121 = transpose(perm = transpose_121_perm_0, x = var_1995)[name = string("transpose_224")]; |
| tensor<fp32, [16, ?, 64]> transpose_120 = transpose(perm = transpose_120_perm_0, x = q_scaled_25)[name = string("transpose_225")]; |
| tensor<fp32, [16, ?, ?]> input_301_bmm = matmul(transpose_x = input_301_bmm_transpose_x_0, transpose_y = input_301_bmm_transpose_y_0, x = transpose_120, y = transpose_121)[name = string("input_301_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_301 = add(x = gated_position_bias_25, y = input_301_bmm)[name = string("input_301")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_49 = softmax(axis = var_9, x = input_301)[name = string("attn_output_weights_49")]; |
| bool attn_output_97_transpose_x_0 = const()[name = string("attn_output_97_transpose_x_0"), val = bool(false)]; |
| bool attn_output_97_transpose_y_0 = const()[name = string("attn_output_97_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_51 = transpose(perm = v_51_perm_0, x = var_2002)[name = string("transpose_223")]; |
| tensor<fp32, [16, ?, 64]> attn_output_97 = matmul(transpose_x = attn_output_97_transpose_x_0, transpose_y = attn_output_97_transpose_y_0, x = attn_output_weights_49, y = v_51)[name = string("attn_output_97")]; |
| tensor<int32, [3]> var_2010_perm_0 = const()[name = string("op_2010_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_103x = const()[name = string("concat_103x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2010 = transpose(perm = var_2010_perm_0, x = attn_output_97)[name = string("transpose_222")]; |
| tensor<fp32, [?, 1024]> attn_output_99 = reshape(shape = concat_103x, x = var_2010)[name = string("attn_output_99")]; |
| tensor<fp32, [?, 1024]> attn_output_101 = linear(bias = net_wavlm_encoder_layers_12_attention_out_proj_bias, weight = net_wavlm_encoder_layers_12_attention_out_proj_weight_quantized, x = attn_output_99)[name = string("linear_89")]; |
| tensor<int32, [3]> concat_104x = const()[name = string("concat_104x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_103 = reshape(shape = concat_104x, x = attn_output_101)[name = string("attn_output_103")]; |
| tensor<int32, [3]> input_303_perm_0 = const()[name = string("input_303_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_303 = transpose(perm = input_303_perm_0, x = attn_output_103)[name = string("transpose_221")]; |
| tensor<fp32, [1, ?, 1024]> input_305 = add(x = input_297, y = input_303)[name = string("input_305")]; |
| tensor<int32, [1]> input_307_axes_0 = const()[name = string("input_307_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_307 = layer_norm(axes = input_307_axes_0, beta = net_wavlm_encoder_layers_12_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_12_final_layer_norm_weight, x = input_305)[name = string("input_307")]; |
| tensor<fp32, [1, ?, 4096]> input_309 = linear(bias = net_wavlm_encoder_layers_12_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_12_feed_forward_intermediate_dense_weight_quantized, x = input_307)[name = string("linear_90")]; |
| string input_311_mode_0 = const()[name = string("input_311_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_311 = gelu(mode = input_311_mode_0, x = input_309)[name = string("input_311")]; |
| tensor<fp32, [1, ?, 1024]> input_315 = linear(bias = net_wavlm_encoder_layers_12_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_12_feed_forward_output_dense_weight_quantized, x = input_311)[name = string("linear_91")]; |
| tensor<fp32, [1, ?, 1024]> input_317 = add(x = input_305, y = input_315)[name = string("input_317")]; |
| tensor<int32, [1]> hidden_states_87_axes_0 = const()[name = string("hidden_states_87_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_87 = layer_norm(axes = hidden_states_87_axes_0, beta = net_wavlm_encoder_layers_13_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_13_layer_norm_weight, x = input_317)[name = string("hidden_states_87")]; |
| tensor<int32, [4]> concat_105x = const()[name = string("concat_105x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_27 = reshape(shape = concat_105x, x = hidden_states_87)[name = string("gated_hidden_states_27")]; |
| tensor<int32, [4]> var_2071 = const()[name = string("op_2071"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_319 = transpose(perm = var_2071, x = gated_hidden_states_27)[name = string("transpose_220")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_53 = linear(bias = net_wavlm_encoder_layers_13_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_13_attention_gru_rel_pos_linear_weight, x = input_319)[name = string("linear_92")]; |
| tensor<int32, [5]> concat_106x = const()[name = string("concat_106x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_2080 = reshape(shape = concat_106x, x = relative_position_proj_53)[name = string("op_2080")]; |
| tensor<int32, [1]> relative_position_proj_55_axes_0 = const()[name = string("relative_position_proj_55_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_55_keep_dims_0 = const()[name = string("relative_position_proj_55_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_55 = reduce_sum(axes = relative_position_proj_55_axes_0, keep_dims = relative_position_proj_55_keep_dims_0, x = var_2080)[name = string("relative_position_proj_55")]; |
| tensor<fp32, [1, 16, ?, 2]> var_2083 = sigmoid(x = relative_position_proj_55)[name = string("op_2083")]; |
| tensor<int32, [2]> var_2084_split_sizes_0 = const()[name = string("op_2084_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_2084_axis_0 = const()[name = string("op_2084_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2084_0, tensor<fp32, [1, 16, ?, 1]> var_2084_1 = split(axis = var_2084_axis_0, split_sizes = var_2084_split_sizes_0, x = var_2083)[name = string("op_2084")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2086 = mul(x = var_2084_1, y = net_wavlm_encoder_layers_13_attention_gru_rel_pos_const)[name = string("op_2086")]; |
| fp32 var_2087 = const()[name = string("op_2087"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2088 = sub(x = var_2086, y = var_2087)[name = string("op_2088")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2089 = mul(x = var_2084_0, y = var_2088)[name = string("op_2089")]; |
| fp32 var_2090 = const()[name = string("op_2090"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_27 = add(x = var_2089, y = var_2090)[name = string("gate_output_27")]; |
| tensor<int32, [3]> var_2095 = const()[name = string("op_2095"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_2096 = reshape(shape = var_2095, x = gate_output_27)[name = string("op_2096")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_27 = mul(x = var_2096, y = position_bias)[name = string("gated_position_bias_27")]; |
| tensor<int32, [3]> query_27_perm_0 = const()[name = string("query_27_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_27 = transpose(perm = query_27_perm_0, x = hidden_states_87)[name = string("transpose_219")]; |
| tensor<fp32, [?, 1, 1024]> q_53 = linear(bias = net_wavlm_encoder_layers_13_attention_q_proj_bias, weight = net_wavlm_encoder_layers_13_attention_q_proj_weight_quantized, x = query_27)[name = string("linear_93")]; |
| tensor<fp32, [?, 1, 1024]> k_53 = linear(bias = net_wavlm_encoder_layers_13_attention_k_proj_bias, weight = net_wavlm_encoder_layers_13_attention_k_proj_weight_quantized, x = query_27)[name = string("linear_94")]; |
| tensor<fp32, [?, 1, 1024]> v_53 = linear(bias = net_wavlm_encoder_layers_13_attention_v_proj_bias, weight = net_wavlm_encoder_layers_13_attention_v_proj_weight_quantized, x = query_27)[name = string("linear_95")]; |
| tensor<int32, [3]> concat_108x = const()[name = string("concat_108x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2123 = reshape(shape = concat_108x, x = q_53)[name = string("op_2123")]; |
| tensor<int32, [3]> concat_109x = const()[name = string("concat_109x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2130 = reshape(shape = concat_109x, x = k_53)[name = string("op_2130")]; |
| tensor<int32, [3]> concat_110x = const()[name = string("concat_110x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2137 = reshape(shape = concat_110x, x = v_53)[name = string("op_2137")]; |
| tensor<int32, [3]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2139 = const()[name = string("op_2139"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_27 = mul(x = var_2123, y = var_2139)[name = string("q_scaled_27")]; |
| bool input_321_bmm_transpose_x_0 = const()[name = string("input_321_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_321_bmm_transpose_y_0 = const()[name = string("input_321_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_122_perm_0 = const()[name = string("transpose_122_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_123_perm_0 = const()[name = string("transpose_123_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_123 = transpose(perm = transpose_123_perm_0, x = var_2130)[name = string("transpose_217")]; |
| tensor<fp32, [16, ?, 64]> transpose_122 = transpose(perm = transpose_122_perm_0, x = q_scaled_27)[name = string("transpose_218")]; |
| tensor<fp32, [16, ?, ?]> input_321_bmm = matmul(transpose_x = input_321_bmm_transpose_x_0, transpose_y = input_321_bmm_transpose_y_0, x = transpose_122, y = transpose_123)[name = string("input_321_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_321 = add(x = gated_position_bias_27, y = input_321_bmm)[name = string("input_321")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_53 = softmax(axis = var_9, x = input_321)[name = string("attn_output_weights_53")]; |
| bool attn_output_105_transpose_x_0 = const()[name = string("attn_output_105_transpose_x_0"), val = bool(false)]; |
| bool attn_output_105_transpose_y_0 = const()[name = string("attn_output_105_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_55 = transpose(perm = v_55_perm_0, x = var_2137)[name = string("transpose_216")]; |
| tensor<fp32, [16, ?, 64]> attn_output_105 = matmul(transpose_x = attn_output_105_transpose_x_0, transpose_y = attn_output_105_transpose_y_0, x = attn_output_weights_53, y = v_55)[name = string("attn_output_105")]; |
| tensor<int32, [3]> var_2145_perm_0 = const()[name = string("op_2145_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_111x = const()[name = string("concat_111x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2145 = transpose(perm = var_2145_perm_0, x = attn_output_105)[name = string("transpose_215")]; |
| tensor<fp32, [?, 1024]> attn_output_107 = reshape(shape = concat_111x, x = var_2145)[name = string("attn_output_107")]; |
| tensor<fp32, [?, 1024]> attn_output_109 = linear(bias = net_wavlm_encoder_layers_13_attention_out_proj_bias, weight = net_wavlm_encoder_layers_13_attention_out_proj_weight_quantized, x = attn_output_107)[name = string("linear_96")]; |
| tensor<int32, [3]> concat_112x = const()[name = string("concat_112x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_111 = reshape(shape = concat_112x, x = attn_output_109)[name = string("attn_output_111")]; |
| tensor<int32, [3]> input_323_perm_0 = const()[name = string("input_323_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_323 = transpose(perm = input_323_perm_0, x = attn_output_111)[name = string("transpose_214")]; |
| tensor<fp32, [1, ?, 1024]> input_325 = add(x = input_317, y = input_323)[name = string("input_325")]; |
| tensor<int32, [1]> input_327_axes_0 = const()[name = string("input_327_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_327 = layer_norm(axes = input_327_axes_0, beta = net_wavlm_encoder_layers_13_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_13_final_layer_norm_weight, x = input_325)[name = string("input_327")]; |
| tensor<fp32, [1, ?, 4096]> input_329 = linear(bias = net_wavlm_encoder_layers_13_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_13_feed_forward_intermediate_dense_weight_quantized, x = input_327)[name = string("linear_97")]; |
| string input_331_mode_0 = const()[name = string("input_331_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_331 = gelu(mode = input_331_mode_0, x = input_329)[name = string("input_331")]; |
| tensor<fp32, [1, ?, 1024]> input_335 = linear(bias = net_wavlm_encoder_layers_13_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_13_feed_forward_output_dense_weight_quantized, x = input_331)[name = string("linear_98")]; |
| tensor<fp32, [1, ?, 1024]> input_337 = add(x = input_325, y = input_335)[name = string("input_337")]; |
| tensor<int32, [1]> hidden_states_91_axes_0 = const()[name = string("hidden_states_91_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_91 = layer_norm(axes = hidden_states_91_axes_0, beta = net_wavlm_encoder_layers_14_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_14_layer_norm_weight, x = input_337)[name = string("hidden_states_91")]; |
| tensor<int32, [4]> concat_113x = const()[name = string("concat_113x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_29 = reshape(shape = concat_113x, x = hidden_states_91)[name = string("gated_hidden_states_29")]; |
| tensor<int32, [4]> var_2206 = const()[name = string("op_2206"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_339 = transpose(perm = var_2206, x = gated_hidden_states_29)[name = string("transpose_213")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_57 = linear(bias = net_wavlm_encoder_layers_14_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_14_attention_gru_rel_pos_linear_weight, x = input_339)[name = string("linear_99")]; |
| tensor<int32, [5]> concat_114x = const()[name = string("concat_114x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_2215 = reshape(shape = concat_114x, x = relative_position_proj_57)[name = string("op_2215")]; |
| tensor<int32, [1]> relative_position_proj_59_axes_0 = const()[name = string("relative_position_proj_59_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_59_keep_dims_0 = const()[name = string("relative_position_proj_59_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_59 = reduce_sum(axes = relative_position_proj_59_axes_0, keep_dims = relative_position_proj_59_keep_dims_0, x = var_2215)[name = string("relative_position_proj_59")]; |
| tensor<fp32, [1, 16, ?, 2]> var_2218 = sigmoid(x = relative_position_proj_59)[name = string("op_2218")]; |
| tensor<int32, [2]> var_2219_split_sizes_0 = const()[name = string("op_2219_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_2219_axis_0 = const()[name = string("op_2219_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2219_0, tensor<fp32, [1, 16, ?, 1]> var_2219_1 = split(axis = var_2219_axis_0, split_sizes = var_2219_split_sizes_0, x = var_2218)[name = string("op_2219")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2221 = mul(x = var_2219_1, y = net_wavlm_encoder_layers_14_attention_gru_rel_pos_const)[name = string("op_2221")]; |
| fp32 var_2222 = const()[name = string("op_2222"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2223 = sub(x = var_2221, y = var_2222)[name = string("op_2223")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2224 = mul(x = var_2219_0, y = var_2223)[name = string("op_2224")]; |
| fp32 var_2225 = const()[name = string("op_2225"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_29 = add(x = var_2224, y = var_2225)[name = string("gate_output_29")]; |
| tensor<int32, [3]> var_2230 = const()[name = string("op_2230"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_2231 = reshape(shape = var_2230, x = gate_output_29)[name = string("op_2231")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_29 = mul(x = var_2231, y = position_bias)[name = string("gated_position_bias_29")]; |
| tensor<int32, [3]> query_29_perm_0 = const()[name = string("query_29_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_29 = transpose(perm = query_29_perm_0, x = hidden_states_91)[name = string("transpose_212")]; |
| tensor<fp32, [?, 1, 1024]> q_57 = linear(bias = net_wavlm_encoder_layers_14_attention_q_proj_bias, weight = net_wavlm_encoder_layers_14_attention_q_proj_weight_quantized, x = query_29)[name = string("linear_100")]; |
| tensor<fp32, [?, 1, 1024]> k_57 = linear(bias = net_wavlm_encoder_layers_14_attention_k_proj_bias, weight = net_wavlm_encoder_layers_14_attention_k_proj_weight_quantized, x = query_29)[name = string("linear_101")]; |
| tensor<fp32, [?, 1, 1024]> v_57 = linear(bias = net_wavlm_encoder_layers_14_attention_v_proj_bias, weight = net_wavlm_encoder_layers_14_attention_v_proj_weight_quantized, x = query_29)[name = string("linear_102")]; |
| tensor<int32, [3]> concat_116x = const()[name = string("concat_116x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2258 = reshape(shape = concat_116x, x = q_57)[name = string("op_2258")]; |
| tensor<int32, [3]> concat_117x = const()[name = string("concat_117x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2265 = reshape(shape = concat_117x, x = k_57)[name = string("op_2265")]; |
| tensor<int32, [3]> concat_118x = const()[name = string("concat_118x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2272 = reshape(shape = concat_118x, x = v_57)[name = string("op_2272")]; |
| tensor<int32, [3]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2274 = const()[name = string("op_2274"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_29 = mul(x = var_2258, y = var_2274)[name = string("q_scaled_29")]; |
| bool input_341_bmm_transpose_x_0 = const()[name = string("input_341_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_341_bmm_transpose_y_0 = const()[name = string("input_341_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_124_perm_0 = const()[name = string("transpose_124_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_125_perm_0 = const()[name = string("transpose_125_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_125 = transpose(perm = transpose_125_perm_0, x = var_2265)[name = string("transpose_210")]; |
| tensor<fp32, [16, ?, 64]> transpose_124 = transpose(perm = transpose_124_perm_0, x = q_scaled_29)[name = string("transpose_211")]; |
| tensor<fp32, [16, ?, ?]> input_341_bmm = matmul(transpose_x = input_341_bmm_transpose_x_0, transpose_y = input_341_bmm_transpose_y_0, x = transpose_124, y = transpose_125)[name = string("input_341_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_341 = add(x = gated_position_bias_29, y = input_341_bmm)[name = string("input_341")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_57 = softmax(axis = var_9, x = input_341)[name = string("attn_output_weights_57")]; |
| bool attn_output_113_transpose_x_0 = const()[name = string("attn_output_113_transpose_x_0"), val = bool(false)]; |
| bool attn_output_113_transpose_y_0 = const()[name = string("attn_output_113_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_59 = transpose(perm = v_59_perm_0, x = var_2272)[name = string("transpose_209")]; |
| tensor<fp32, [16, ?, 64]> attn_output_113 = matmul(transpose_x = attn_output_113_transpose_x_0, transpose_y = attn_output_113_transpose_y_0, x = attn_output_weights_57, y = v_59)[name = string("attn_output_113")]; |
| tensor<int32, [3]> var_2280_perm_0 = const()[name = string("op_2280_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_119x = const()[name = string("concat_119x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2280 = transpose(perm = var_2280_perm_0, x = attn_output_113)[name = string("transpose_208")]; |
| tensor<fp32, [?, 1024]> attn_output_115 = reshape(shape = concat_119x, x = var_2280)[name = string("attn_output_115")]; |
| tensor<fp32, [?, 1024]> attn_output_117 = linear(bias = net_wavlm_encoder_layers_14_attention_out_proj_bias, weight = net_wavlm_encoder_layers_14_attention_out_proj_weight_quantized, x = attn_output_115)[name = string("linear_103")]; |
| tensor<int32, [3]> concat_120x = const()[name = string("concat_120x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_119 = reshape(shape = concat_120x, x = attn_output_117)[name = string("attn_output_119")]; |
| tensor<int32, [3]> input_343_perm_0 = const()[name = string("input_343_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_343 = transpose(perm = input_343_perm_0, x = attn_output_119)[name = string("transpose_207")]; |
| tensor<fp32, [1, ?, 1024]> input_345 = add(x = input_337, y = input_343)[name = string("input_345")]; |
| tensor<int32, [1]> input_347_axes_0 = const()[name = string("input_347_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_347 = layer_norm(axes = input_347_axes_0, beta = net_wavlm_encoder_layers_14_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_14_final_layer_norm_weight, x = input_345)[name = string("input_347")]; |
| tensor<fp32, [1, ?, 4096]> input_349 = linear(bias = net_wavlm_encoder_layers_14_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_14_feed_forward_intermediate_dense_weight_quantized, x = input_347)[name = string("linear_104")]; |
| string input_351_mode_0 = const()[name = string("input_351_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_351 = gelu(mode = input_351_mode_0, x = input_349)[name = string("input_351")]; |
| tensor<fp32, [1, ?, 1024]> input_355 = linear(bias = net_wavlm_encoder_layers_14_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_14_feed_forward_output_dense_weight_quantized, x = input_351)[name = string("linear_105")]; |
| tensor<fp32, [1, ?, 1024]> input_357 = add(x = input_345, y = input_355)[name = string("input_357")]; |
| tensor<int32, [1]> hidden_states_95_axes_0 = const()[name = string("hidden_states_95_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_95 = layer_norm(axes = hidden_states_95_axes_0, beta = net_wavlm_encoder_layers_15_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_15_layer_norm_weight, x = input_357)[name = string("hidden_states_95")]; |
| tensor<int32, [4]> concat_121x = const()[name = string("concat_121x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_31 = reshape(shape = concat_121x, x = hidden_states_95)[name = string("gated_hidden_states_31")]; |
| tensor<int32, [4]> var_2341 = const()[name = string("op_2341"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_359 = transpose(perm = var_2341, x = gated_hidden_states_31)[name = string("transpose_206")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_61 = linear(bias = net_wavlm_encoder_layers_15_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_15_attention_gru_rel_pos_linear_weight, x = input_359)[name = string("linear_106")]; |
| tensor<int32, [5]> concat_122x = const()[name = string("concat_122x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_2350 = reshape(shape = concat_122x, x = relative_position_proj_61)[name = string("op_2350")]; |
| tensor<int32, [1]> relative_position_proj_63_axes_0 = const()[name = string("relative_position_proj_63_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_63_keep_dims_0 = const()[name = string("relative_position_proj_63_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_63 = reduce_sum(axes = relative_position_proj_63_axes_0, keep_dims = relative_position_proj_63_keep_dims_0, x = var_2350)[name = string("relative_position_proj_63")]; |
| tensor<fp32, [1, 16, ?, 2]> var_2353 = sigmoid(x = relative_position_proj_63)[name = string("op_2353")]; |
| tensor<int32, [2]> var_2354_split_sizes_0 = const()[name = string("op_2354_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_2354_axis_0 = const()[name = string("op_2354_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2354_0, tensor<fp32, [1, 16, ?, 1]> var_2354_1 = split(axis = var_2354_axis_0, split_sizes = var_2354_split_sizes_0, x = var_2353)[name = string("op_2354")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2356 = mul(x = var_2354_1, y = net_wavlm_encoder_layers_15_attention_gru_rel_pos_const)[name = string("op_2356")]; |
| fp32 var_2357 = const()[name = string("op_2357"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2358 = sub(x = var_2356, y = var_2357)[name = string("op_2358")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2359 = mul(x = var_2354_0, y = var_2358)[name = string("op_2359")]; |
| fp32 var_2360 = const()[name = string("op_2360"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_31 = add(x = var_2359, y = var_2360)[name = string("gate_output_31")]; |
| tensor<int32, [3]> var_2365 = const()[name = string("op_2365"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_2366 = reshape(shape = var_2365, x = gate_output_31)[name = string("op_2366")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_31 = mul(x = var_2366, y = position_bias)[name = string("gated_position_bias_31")]; |
| tensor<int32, [3]> query_31_perm_0 = const()[name = string("query_31_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_31 = transpose(perm = query_31_perm_0, x = hidden_states_95)[name = string("transpose_205")]; |
| tensor<fp32, [?, 1, 1024]> q_61 = linear(bias = net_wavlm_encoder_layers_15_attention_q_proj_bias, weight = net_wavlm_encoder_layers_15_attention_q_proj_weight_quantized, x = query_31)[name = string("linear_107")]; |
| tensor<fp32, [?, 1, 1024]> k_61 = linear(bias = net_wavlm_encoder_layers_15_attention_k_proj_bias, weight = net_wavlm_encoder_layers_15_attention_k_proj_weight_quantized, x = query_31)[name = string("linear_108")]; |
| tensor<fp32, [?, 1, 1024]> v_61 = linear(bias = net_wavlm_encoder_layers_15_attention_v_proj_bias, weight = net_wavlm_encoder_layers_15_attention_v_proj_weight_quantized, x = query_31)[name = string("linear_109")]; |
| tensor<int32, [3]> concat_124x = const()[name = string("concat_124x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2393 = reshape(shape = concat_124x, x = q_61)[name = string("op_2393")]; |
| tensor<int32, [3]> concat_125x = const()[name = string("concat_125x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2400 = reshape(shape = concat_125x, x = k_61)[name = string("op_2400")]; |
| tensor<int32, [3]> concat_126x = const()[name = string("concat_126x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2407 = reshape(shape = concat_126x, x = v_61)[name = string("op_2407")]; |
| tensor<int32, [3]> v_63_perm_0 = const()[name = string("v_63_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2409 = const()[name = string("op_2409"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_31 = mul(x = var_2393, y = var_2409)[name = string("q_scaled_31")]; |
| bool input_361_bmm_transpose_x_0 = const()[name = string("input_361_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_361_bmm_transpose_y_0 = const()[name = string("input_361_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_126_perm_0 = const()[name = string("transpose_126_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_127_perm_0 = const()[name = string("transpose_127_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_127 = transpose(perm = transpose_127_perm_0, x = var_2400)[name = string("transpose_203")]; |
| tensor<fp32, [16, ?, 64]> transpose_126 = transpose(perm = transpose_126_perm_0, x = q_scaled_31)[name = string("transpose_204")]; |
| tensor<fp32, [16, ?, ?]> input_361_bmm = matmul(transpose_x = input_361_bmm_transpose_x_0, transpose_y = input_361_bmm_transpose_y_0, x = transpose_126, y = transpose_127)[name = string("input_361_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_361 = add(x = gated_position_bias_31, y = input_361_bmm)[name = string("input_361")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_61 = softmax(axis = var_9, x = input_361)[name = string("attn_output_weights_61")]; |
| bool attn_output_121_transpose_x_0 = const()[name = string("attn_output_121_transpose_x_0"), val = bool(false)]; |
| bool attn_output_121_transpose_y_0 = const()[name = string("attn_output_121_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_63 = transpose(perm = v_63_perm_0, x = var_2407)[name = string("transpose_202")]; |
| tensor<fp32, [16, ?, 64]> attn_output_121 = matmul(transpose_x = attn_output_121_transpose_x_0, transpose_y = attn_output_121_transpose_y_0, x = attn_output_weights_61, y = v_63)[name = string("attn_output_121")]; |
| tensor<int32, [3]> var_2415_perm_0 = const()[name = string("op_2415_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_127x = const()[name = string("concat_127x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2415 = transpose(perm = var_2415_perm_0, x = attn_output_121)[name = string("transpose_201")]; |
| tensor<fp32, [?, 1024]> attn_output_123 = reshape(shape = concat_127x, x = var_2415)[name = string("attn_output_123")]; |
| tensor<fp32, [?, 1024]> attn_output_125 = linear(bias = net_wavlm_encoder_layers_15_attention_out_proj_bias, weight = net_wavlm_encoder_layers_15_attention_out_proj_weight_quantized, x = attn_output_123)[name = string("linear_110")]; |
| tensor<int32, [3]> concat_128x = const()[name = string("concat_128x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_127 = reshape(shape = concat_128x, x = attn_output_125)[name = string("attn_output_127")]; |
| tensor<int32, [3]> input_363_perm_0 = const()[name = string("input_363_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_363 = transpose(perm = input_363_perm_0, x = attn_output_127)[name = string("transpose_200")]; |
| tensor<fp32, [1, ?, 1024]> input_365 = add(x = input_357, y = input_363)[name = string("input_365")]; |
| tensor<int32, [1]> input_367_axes_0 = const()[name = string("input_367_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_367 = layer_norm(axes = input_367_axes_0, beta = net_wavlm_encoder_layers_15_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_15_final_layer_norm_weight, x = input_365)[name = string("input_367")]; |
| tensor<fp32, [1, ?, 4096]> input_369 = linear(bias = net_wavlm_encoder_layers_15_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_15_feed_forward_intermediate_dense_weight_quantized, x = input_367)[name = string("linear_111")]; |
| string input_371_mode_0 = const()[name = string("input_371_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_371 = gelu(mode = input_371_mode_0, x = input_369)[name = string("input_371")]; |
| tensor<fp32, [1, ?, 1024]> input_375 = linear(bias = net_wavlm_encoder_layers_15_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_15_feed_forward_output_dense_weight_quantized, x = input_371)[name = string("linear_112")]; |
| tensor<fp32, [1, ?, 1024]> input_377 = add(x = input_365, y = input_375)[name = string("input_377")]; |
| tensor<int32, [1]> hidden_states_99_axes_0 = const()[name = string("hidden_states_99_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_99 = layer_norm(axes = hidden_states_99_axes_0, beta = net_wavlm_encoder_layers_16_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_16_layer_norm_weight, x = input_377)[name = string("hidden_states_99")]; |
| tensor<int32, [4]> concat_129x = const()[name = string("concat_129x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_33 = reshape(shape = concat_129x, x = hidden_states_99)[name = string("gated_hidden_states_33")]; |
| tensor<int32, [4]> var_2476 = const()[name = string("op_2476"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_379 = transpose(perm = var_2476, x = gated_hidden_states_33)[name = string("transpose_199")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_65 = linear(bias = net_wavlm_encoder_layers_16_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_16_attention_gru_rel_pos_linear_weight, x = input_379)[name = string("linear_113")]; |
| tensor<int32, [5]> concat_130x = const()[name = string("concat_130x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_2485 = reshape(shape = concat_130x, x = relative_position_proj_65)[name = string("op_2485")]; |
| tensor<int32, [1]> relative_position_proj_67_axes_0 = const()[name = string("relative_position_proj_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_67_keep_dims_0 = const()[name = string("relative_position_proj_67_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_67 = reduce_sum(axes = relative_position_proj_67_axes_0, keep_dims = relative_position_proj_67_keep_dims_0, x = var_2485)[name = string("relative_position_proj_67")]; |
| tensor<fp32, [1, 16, ?, 2]> var_2488 = sigmoid(x = relative_position_proj_67)[name = string("op_2488")]; |
| tensor<int32, [2]> var_2489_split_sizes_0 = const()[name = string("op_2489_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_2489_axis_0 = const()[name = string("op_2489_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2489_0, tensor<fp32, [1, 16, ?, 1]> var_2489_1 = split(axis = var_2489_axis_0, split_sizes = var_2489_split_sizes_0, x = var_2488)[name = string("op_2489")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2491 = mul(x = var_2489_1, y = net_wavlm_encoder_layers_16_attention_gru_rel_pos_const)[name = string("op_2491")]; |
| fp32 var_2492 = const()[name = string("op_2492"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2493 = sub(x = var_2491, y = var_2492)[name = string("op_2493")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2494 = mul(x = var_2489_0, y = var_2493)[name = string("op_2494")]; |
| fp32 var_2495 = const()[name = string("op_2495"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_33 = add(x = var_2494, y = var_2495)[name = string("gate_output_33")]; |
| tensor<int32, [3]> var_2500 = const()[name = string("op_2500"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_2501 = reshape(shape = var_2500, x = gate_output_33)[name = string("op_2501")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_33 = mul(x = var_2501, y = position_bias)[name = string("gated_position_bias_33")]; |
| tensor<int32, [3]> query_33_perm_0 = const()[name = string("query_33_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_33 = transpose(perm = query_33_perm_0, x = hidden_states_99)[name = string("transpose_198")]; |
| tensor<fp32, [?, 1, 1024]> q_65 = linear(bias = net_wavlm_encoder_layers_16_attention_q_proj_bias, weight = net_wavlm_encoder_layers_16_attention_q_proj_weight_quantized, x = query_33)[name = string("linear_114")]; |
| tensor<fp32, [?, 1, 1024]> k_65 = linear(bias = net_wavlm_encoder_layers_16_attention_k_proj_bias, weight = net_wavlm_encoder_layers_16_attention_k_proj_weight_quantized, x = query_33)[name = string("linear_115")]; |
| tensor<fp32, [?, 1, 1024]> v_65 = linear(bias = net_wavlm_encoder_layers_16_attention_v_proj_bias, weight = net_wavlm_encoder_layers_16_attention_v_proj_weight_quantized, x = query_33)[name = string("linear_116")]; |
| tensor<int32, [3]> concat_132x = const()[name = string("concat_132x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2528 = reshape(shape = concat_132x, x = q_65)[name = string("op_2528")]; |
| tensor<int32, [3]> concat_133x = const()[name = string("concat_133x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2535 = reshape(shape = concat_133x, x = k_65)[name = string("op_2535")]; |
| tensor<int32, [3]> concat_134x = const()[name = string("concat_134x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2542 = reshape(shape = concat_134x, x = v_65)[name = string("op_2542")]; |
| tensor<int32, [3]> v_67_perm_0 = const()[name = string("v_67_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2544 = const()[name = string("op_2544"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_33 = mul(x = var_2528, y = var_2544)[name = string("q_scaled_33")]; |
| bool input_381_bmm_transpose_x_0 = const()[name = string("input_381_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_381_bmm_transpose_y_0 = const()[name = string("input_381_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_128_perm_0 = const()[name = string("transpose_128_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_129_perm_0 = const()[name = string("transpose_129_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_129 = transpose(perm = transpose_129_perm_0, x = var_2535)[name = string("transpose_196")]; |
| tensor<fp32, [16, ?, 64]> transpose_128 = transpose(perm = transpose_128_perm_0, x = q_scaled_33)[name = string("transpose_197")]; |
| tensor<fp32, [16, ?, ?]> input_381_bmm = matmul(transpose_x = input_381_bmm_transpose_x_0, transpose_y = input_381_bmm_transpose_y_0, x = transpose_128, y = transpose_129)[name = string("input_381_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_381 = add(x = gated_position_bias_33, y = input_381_bmm)[name = string("input_381")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_65 = softmax(axis = var_9, x = input_381)[name = string("attn_output_weights_65")]; |
| bool attn_output_129_transpose_x_0 = const()[name = string("attn_output_129_transpose_x_0"), val = bool(false)]; |
| bool attn_output_129_transpose_y_0 = const()[name = string("attn_output_129_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_67 = transpose(perm = v_67_perm_0, x = var_2542)[name = string("transpose_195")]; |
| tensor<fp32, [16, ?, 64]> attn_output_129 = matmul(transpose_x = attn_output_129_transpose_x_0, transpose_y = attn_output_129_transpose_y_0, x = attn_output_weights_65, y = v_67)[name = string("attn_output_129")]; |
| tensor<int32, [3]> var_2550_perm_0 = const()[name = string("op_2550_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_135x = const()[name = string("concat_135x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2550 = transpose(perm = var_2550_perm_0, x = attn_output_129)[name = string("transpose_194")]; |
| tensor<fp32, [?, 1024]> attn_output_131 = reshape(shape = concat_135x, x = var_2550)[name = string("attn_output_131")]; |
| tensor<fp32, [?, 1024]> attn_output_133 = linear(bias = net_wavlm_encoder_layers_16_attention_out_proj_bias, weight = net_wavlm_encoder_layers_16_attention_out_proj_weight_quantized, x = attn_output_131)[name = string("linear_117")]; |
| tensor<int32, [3]> concat_136x = const()[name = string("concat_136x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_135 = reshape(shape = concat_136x, x = attn_output_133)[name = string("attn_output_135")]; |
| tensor<int32, [3]> input_383_perm_0 = const()[name = string("input_383_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_383 = transpose(perm = input_383_perm_0, x = attn_output_135)[name = string("transpose_193")]; |
| tensor<fp32, [1, ?, 1024]> input_385 = add(x = input_377, y = input_383)[name = string("input_385")]; |
| tensor<int32, [1]> input_387_axes_0 = const()[name = string("input_387_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_387 = layer_norm(axes = input_387_axes_0, beta = net_wavlm_encoder_layers_16_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_16_final_layer_norm_weight, x = input_385)[name = string("input_387")]; |
| tensor<fp32, [1, ?, 4096]> input_389 = linear(bias = net_wavlm_encoder_layers_16_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_16_feed_forward_intermediate_dense_weight_quantized, x = input_387)[name = string("linear_118")]; |
| string input_391_mode_0 = const()[name = string("input_391_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_391 = gelu(mode = input_391_mode_0, x = input_389)[name = string("input_391")]; |
| tensor<fp32, [1, ?, 1024]> input_395 = linear(bias = net_wavlm_encoder_layers_16_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_16_feed_forward_output_dense_weight_quantized, x = input_391)[name = string("linear_119")]; |
| tensor<fp32, [1, ?, 1024]> input_397 = add(x = input_385, y = input_395)[name = string("input_397")]; |
| tensor<int32, [1]> hidden_states_103_axes_0 = const()[name = string("hidden_states_103_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_103 = layer_norm(axes = hidden_states_103_axes_0, beta = net_wavlm_encoder_layers_17_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_17_layer_norm_weight, x = input_397)[name = string("hidden_states_103")]; |
| tensor<int32, [4]> concat_137x = const()[name = string("concat_137x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_35 = reshape(shape = concat_137x, x = hidden_states_103)[name = string("gated_hidden_states_35")]; |
| tensor<int32, [4]> var_2611 = const()[name = string("op_2611"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_399 = transpose(perm = var_2611, x = gated_hidden_states_35)[name = string("transpose_192")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_69 = linear(bias = net_wavlm_encoder_layers_17_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_17_attention_gru_rel_pos_linear_weight, x = input_399)[name = string("linear_120")]; |
| tensor<int32, [5]> concat_138x = const()[name = string("concat_138x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_2620 = reshape(shape = concat_138x, x = relative_position_proj_69)[name = string("op_2620")]; |
| tensor<int32, [1]> relative_position_proj_71_axes_0 = const()[name = string("relative_position_proj_71_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_71_keep_dims_0 = const()[name = string("relative_position_proj_71_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_71 = reduce_sum(axes = relative_position_proj_71_axes_0, keep_dims = relative_position_proj_71_keep_dims_0, x = var_2620)[name = string("relative_position_proj_71")]; |
| tensor<fp32, [1, 16, ?, 2]> var_2623 = sigmoid(x = relative_position_proj_71)[name = string("op_2623")]; |
| tensor<int32, [2]> var_2624_split_sizes_0 = const()[name = string("op_2624_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_2624_axis_0 = const()[name = string("op_2624_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2624_0, tensor<fp32, [1, 16, ?, 1]> var_2624_1 = split(axis = var_2624_axis_0, split_sizes = var_2624_split_sizes_0, x = var_2623)[name = string("op_2624")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2626 = mul(x = var_2624_1, y = net_wavlm_encoder_layers_17_attention_gru_rel_pos_const)[name = string("op_2626")]; |
| fp32 var_2627 = const()[name = string("op_2627"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2628 = sub(x = var_2626, y = var_2627)[name = string("op_2628")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2629 = mul(x = var_2624_0, y = var_2628)[name = string("op_2629")]; |
| fp32 var_2630 = const()[name = string("op_2630"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_35 = add(x = var_2629, y = var_2630)[name = string("gate_output_35")]; |
| tensor<int32, [3]> var_2635 = const()[name = string("op_2635"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_2636 = reshape(shape = var_2635, x = gate_output_35)[name = string("op_2636")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_35 = mul(x = var_2636, y = position_bias)[name = string("gated_position_bias_35")]; |
| tensor<int32, [3]> query_35_perm_0 = const()[name = string("query_35_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_35 = transpose(perm = query_35_perm_0, x = hidden_states_103)[name = string("transpose_191")]; |
| tensor<fp32, [?, 1, 1024]> q_69 = linear(bias = net_wavlm_encoder_layers_17_attention_q_proj_bias, weight = net_wavlm_encoder_layers_17_attention_q_proj_weight_quantized, x = query_35)[name = string("linear_121")]; |
| tensor<fp32, [?, 1, 1024]> k_69 = linear(bias = net_wavlm_encoder_layers_17_attention_k_proj_bias, weight = net_wavlm_encoder_layers_17_attention_k_proj_weight_quantized, x = query_35)[name = string("linear_122")]; |
| tensor<fp32, [?, 1, 1024]> v_69 = linear(bias = net_wavlm_encoder_layers_17_attention_v_proj_bias, weight = net_wavlm_encoder_layers_17_attention_v_proj_weight_quantized, x = query_35)[name = string("linear_123")]; |
| tensor<int32, [3]> concat_140x = const()[name = string("concat_140x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2663 = reshape(shape = concat_140x, x = q_69)[name = string("op_2663")]; |
| tensor<int32, [3]> concat_141x = const()[name = string("concat_141x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2670 = reshape(shape = concat_141x, x = k_69)[name = string("op_2670")]; |
| tensor<int32, [3]> concat_142x = const()[name = string("concat_142x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2677 = reshape(shape = concat_142x, x = v_69)[name = string("op_2677")]; |
| tensor<int32, [3]> v_71_perm_0 = const()[name = string("v_71_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2679 = const()[name = string("op_2679"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_35 = mul(x = var_2663, y = var_2679)[name = string("q_scaled_35")]; |
| bool input_401_bmm_transpose_x_0 = const()[name = string("input_401_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_401_bmm_transpose_y_0 = const()[name = string("input_401_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_130_perm_0 = const()[name = string("transpose_130_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_131_perm_0 = const()[name = string("transpose_131_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_131 = transpose(perm = transpose_131_perm_0, x = var_2670)[name = string("transpose_189")]; |
| tensor<fp32, [16, ?, 64]> transpose_130 = transpose(perm = transpose_130_perm_0, x = q_scaled_35)[name = string("transpose_190")]; |
| tensor<fp32, [16, ?, ?]> input_401_bmm = matmul(transpose_x = input_401_bmm_transpose_x_0, transpose_y = input_401_bmm_transpose_y_0, x = transpose_130, y = transpose_131)[name = string("input_401_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_401 = add(x = gated_position_bias_35, y = input_401_bmm)[name = string("input_401")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_69 = softmax(axis = var_9, x = input_401)[name = string("attn_output_weights_69")]; |
| bool attn_output_137_transpose_x_0 = const()[name = string("attn_output_137_transpose_x_0"), val = bool(false)]; |
| bool attn_output_137_transpose_y_0 = const()[name = string("attn_output_137_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_71 = transpose(perm = v_71_perm_0, x = var_2677)[name = string("transpose_188")]; |
| tensor<fp32, [16, ?, 64]> attn_output_137 = matmul(transpose_x = attn_output_137_transpose_x_0, transpose_y = attn_output_137_transpose_y_0, x = attn_output_weights_69, y = v_71)[name = string("attn_output_137")]; |
| tensor<int32, [3]> var_2685_perm_0 = const()[name = string("op_2685_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_143x = const()[name = string("concat_143x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2685 = transpose(perm = var_2685_perm_0, x = attn_output_137)[name = string("transpose_187")]; |
| tensor<fp32, [?, 1024]> attn_output_139 = reshape(shape = concat_143x, x = var_2685)[name = string("attn_output_139")]; |
| tensor<fp32, [?, 1024]> attn_output_141 = linear(bias = net_wavlm_encoder_layers_17_attention_out_proj_bias, weight = net_wavlm_encoder_layers_17_attention_out_proj_weight_quantized, x = attn_output_139)[name = string("linear_124")]; |
| tensor<int32, [3]> concat_144x = const()[name = string("concat_144x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_143 = reshape(shape = concat_144x, x = attn_output_141)[name = string("attn_output_143")]; |
| tensor<int32, [3]> input_403_perm_0 = const()[name = string("input_403_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_403 = transpose(perm = input_403_perm_0, x = attn_output_143)[name = string("transpose_186")]; |
| tensor<fp32, [1, ?, 1024]> input_405 = add(x = input_397, y = input_403)[name = string("input_405")]; |
| tensor<int32, [1]> input_407_axes_0 = const()[name = string("input_407_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_407 = layer_norm(axes = input_407_axes_0, beta = net_wavlm_encoder_layers_17_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_17_final_layer_norm_weight, x = input_405)[name = string("input_407")]; |
| tensor<fp32, [1, ?, 4096]> input_409 = linear(bias = net_wavlm_encoder_layers_17_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_17_feed_forward_intermediate_dense_weight_quantized, x = input_407)[name = string("linear_125")]; |
| string input_411_mode_0 = const()[name = string("input_411_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_411 = gelu(mode = input_411_mode_0, x = input_409)[name = string("input_411")]; |
| tensor<fp32, [1, ?, 1024]> input_415 = linear(bias = net_wavlm_encoder_layers_17_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_17_feed_forward_output_dense_weight_quantized, x = input_411)[name = string("linear_126")]; |
| tensor<fp32, [1, ?, 1024]> input_417 = add(x = input_405, y = input_415)[name = string("input_417")]; |
| tensor<int32, [1]> hidden_states_107_axes_0 = const()[name = string("hidden_states_107_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_107 = layer_norm(axes = hidden_states_107_axes_0, beta = net_wavlm_encoder_layers_18_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_18_layer_norm_weight, x = input_417)[name = string("hidden_states_107")]; |
| tensor<int32, [4]> concat_145x = const()[name = string("concat_145x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_37 = reshape(shape = concat_145x, x = hidden_states_107)[name = string("gated_hidden_states_37")]; |
| tensor<int32, [4]> var_2746 = const()[name = string("op_2746"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_419 = transpose(perm = var_2746, x = gated_hidden_states_37)[name = string("transpose_185")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_73 = linear(bias = net_wavlm_encoder_layers_18_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_18_attention_gru_rel_pos_linear_weight, x = input_419)[name = string("linear_127")]; |
| tensor<int32, [5]> concat_146x = const()[name = string("concat_146x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_2755 = reshape(shape = concat_146x, x = relative_position_proj_73)[name = string("op_2755")]; |
| tensor<int32, [1]> relative_position_proj_75_axes_0 = const()[name = string("relative_position_proj_75_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_75_keep_dims_0 = const()[name = string("relative_position_proj_75_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_75 = reduce_sum(axes = relative_position_proj_75_axes_0, keep_dims = relative_position_proj_75_keep_dims_0, x = var_2755)[name = string("relative_position_proj_75")]; |
| tensor<fp32, [1, 16, ?, 2]> var_2758 = sigmoid(x = relative_position_proj_75)[name = string("op_2758")]; |
| tensor<int32, [2]> var_2759_split_sizes_0 = const()[name = string("op_2759_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_2759_axis_0 = const()[name = string("op_2759_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2759_0, tensor<fp32, [1, 16, ?, 1]> var_2759_1 = split(axis = var_2759_axis_0, split_sizes = var_2759_split_sizes_0, x = var_2758)[name = string("op_2759")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2761 = mul(x = var_2759_1, y = net_wavlm_encoder_layers_18_attention_gru_rel_pos_const)[name = string("op_2761")]; |
| fp32 var_2762 = const()[name = string("op_2762"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2763 = sub(x = var_2761, y = var_2762)[name = string("op_2763")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2764 = mul(x = var_2759_0, y = var_2763)[name = string("op_2764")]; |
| fp32 var_2765 = const()[name = string("op_2765"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_37 = add(x = var_2764, y = var_2765)[name = string("gate_output_37")]; |
| tensor<int32, [3]> var_2770 = const()[name = string("op_2770"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_2771 = reshape(shape = var_2770, x = gate_output_37)[name = string("op_2771")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_37 = mul(x = var_2771, y = position_bias)[name = string("gated_position_bias_37")]; |
| tensor<int32, [3]> query_37_perm_0 = const()[name = string("query_37_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_37 = transpose(perm = query_37_perm_0, x = hidden_states_107)[name = string("transpose_184")]; |
| tensor<fp32, [?, 1, 1024]> q_73 = linear(bias = net_wavlm_encoder_layers_18_attention_q_proj_bias, weight = net_wavlm_encoder_layers_18_attention_q_proj_weight_quantized, x = query_37)[name = string("linear_128")]; |
| tensor<fp32, [?, 1, 1024]> k_73 = linear(bias = net_wavlm_encoder_layers_18_attention_k_proj_bias, weight = net_wavlm_encoder_layers_18_attention_k_proj_weight_quantized, x = query_37)[name = string("linear_129")]; |
| tensor<fp32, [?, 1, 1024]> v_73 = linear(bias = net_wavlm_encoder_layers_18_attention_v_proj_bias, weight = net_wavlm_encoder_layers_18_attention_v_proj_weight_quantized, x = query_37)[name = string("linear_130")]; |
| tensor<int32, [3]> concat_148x = const()[name = string("concat_148x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2798 = reshape(shape = concat_148x, x = q_73)[name = string("op_2798")]; |
| tensor<int32, [3]> concat_149x = const()[name = string("concat_149x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2805 = reshape(shape = concat_149x, x = k_73)[name = string("op_2805")]; |
| tensor<int32, [3]> concat_150x = const()[name = string("concat_150x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2812 = reshape(shape = concat_150x, x = v_73)[name = string("op_2812")]; |
| tensor<int32, [3]> v_75_perm_0 = const()[name = string("v_75_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2814 = const()[name = string("op_2814"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_37 = mul(x = var_2798, y = var_2814)[name = string("q_scaled_37")]; |
| bool input_421_bmm_transpose_x_0 = const()[name = string("input_421_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_421_bmm_transpose_y_0 = const()[name = string("input_421_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_132_perm_0 = const()[name = string("transpose_132_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_133_perm_0 = const()[name = string("transpose_133_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_133 = transpose(perm = transpose_133_perm_0, x = var_2805)[name = string("transpose_182")]; |
| tensor<fp32, [16, ?, 64]> transpose_132 = transpose(perm = transpose_132_perm_0, x = q_scaled_37)[name = string("transpose_183")]; |
| tensor<fp32, [16, ?, ?]> input_421_bmm = matmul(transpose_x = input_421_bmm_transpose_x_0, transpose_y = input_421_bmm_transpose_y_0, x = transpose_132, y = transpose_133)[name = string("input_421_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_421 = add(x = gated_position_bias_37, y = input_421_bmm)[name = string("input_421")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_73 = softmax(axis = var_9, x = input_421)[name = string("attn_output_weights_73")]; |
| bool attn_output_145_transpose_x_0 = const()[name = string("attn_output_145_transpose_x_0"), val = bool(false)]; |
| bool attn_output_145_transpose_y_0 = const()[name = string("attn_output_145_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_75 = transpose(perm = v_75_perm_0, x = var_2812)[name = string("transpose_181")]; |
| tensor<fp32, [16, ?, 64]> attn_output_145 = matmul(transpose_x = attn_output_145_transpose_x_0, transpose_y = attn_output_145_transpose_y_0, x = attn_output_weights_73, y = v_75)[name = string("attn_output_145")]; |
| tensor<int32, [3]> var_2820_perm_0 = const()[name = string("op_2820_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_151x = const()[name = string("concat_151x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2820 = transpose(perm = var_2820_perm_0, x = attn_output_145)[name = string("transpose_180")]; |
| tensor<fp32, [?, 1024]> attn_output_147 = reshape(shape = concat_151x, x = var_2820)[name = string("attn_output_147")]; |
| tensor<fp32, [?, 1024]> attn_output_149 = linear(bias = net_wavlm_encoder_layers_18_attention_out_proj_bias, weight = net_wavlm_encoder_layers_18_attention_out_proj_weight_quantized, x = attn_output_147)[name = string("linear_131")]; |
| tensor<int32, [3]> concat_152x = const()[name = string("concat_152x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_151 = reshape(shape = concat_152x, x = attn_output_149)[name = string("attn_output_151")]; |
| tensor<int32, [3]> input_423_perm_0 = const()[name = string("input_423_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_423 = transpose(perm = input_423_perm_0, x = attn_output_151)[name = string("transpose_179")]; |
| tensor<fp32, [1, ?, 1024]> input_425 = add(x = input_417, y = input_423)[name = string("input_425")]; |
| tensor<int32, [1]> input_427_axes_0 = const()[name = string("input_427_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_427 = layer_norm(axes = input_427_axes_0, beta = net_wavlm_encoder_layers_18_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_18_final_layer_norm_weight, x = input_425)[name = string("input_427")]; |
| tensor<fp32, [1, ?, 4096]> input_429 = linear(bias = net_wavlm_encoder_layers_18_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_18_feed_forward_intermediate_dense_weight_quantized, x = input_427)[name = string("linear_132")]; |
| string input_431_mode_0 = const()[name = string("input_431_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_431 = gelu(mode = input_431_mode_0, x = input_429)[name = string("input_431")]; |
| tensor<fp32, [1, ?, 1024]> input_435 = linear(bias = net_wavlm_encoder_layers_18_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_18_feed_forward_output_dense_weight_quantized, x = input_431)[name = string("linear_133")]; |
| tensor<fp32, [1, ?, 1024]> input_437 = add(x = input_425, y = input_435)[name = string("input_437")]; |
| tensor<int32, [1]> hidden_states_111_axes_0 = const()[name = string("hidden_states_111_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_111 = layer_norm(axes = hidden_states_111_axes_0, beta = net_wavlm_encoder_layers_19_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_19_layer_norm_weight, x = input_437)[name = string("hidden_states_111")]; |
| tensor<int32, [4]> concat_153x = const()[name = string("concat_153x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_39 = reshape(shape = concat_153x, x = hidden_states_111)[name = string("gated_hidden_states_39")]; |
| tensor<int32, [4]> var_2881 = const()[name = string("op_2881"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_439 = transpose(perm = var_2881, x = gated_hidden_states_39)[name = string("transpose_178")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_77 = linear(bias = net_wavlm_encoder_layers_19_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_19_attention_gru_rel_pos_linear_weight, x = input_439)[name = string("linear_134")]; |
| tensor<int32, [5]> concat_154x = const()[name = string("concat_154x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_2890 = reshape(shape = concat_154x, x = relative_position_proj_77)[name = string("op_2890")]; |
| tensor<int32, [1]> relative_position_proj_79_axes_0 = const()[name = string("relative_position_proj_79_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_79_keep_dims_0 = const()[name = string("relative_position_proj_79_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_79 = reduce_sum(axes = relative_position_proj_79_axes_0, keep_dims = relative_position_proj_79_keep_dims_0, x = var_2890)[name = string("relative_position_proj_79")]; |
| tensor<fp32, [1, 16, ?, 2]> var_2893 = sigmoid(x = relative_position_proj_79)[name = string("op_2893")]; |
| tensor<int32, [2]> var_2894_split_sizes_0 = const()[name = string("op_2894_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_2894_axis_0 = const()[name = string("op_2894_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2894_0, tensor<fp32, [1, 16, ?, 1]> var_2894_1 = split(axis = var_2894_axis_0, split_sizes = var_2894_split_sizes_0, x = var_2893)[name = string("op_2894")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2896 = mul(x = var_2894_1, y = net_wavlm_encoder_layers_19_attention_gru_rel_pos_const)[name = string("op_2896")]; |
| fp32 var_2897 = const()[name = string("op_2897"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_2898 = sub(x = var_2896, y = var_2897)[name = string("op_2898")]; |
| tensor<fp32, [1, 16, ?, 1]> var_2899 = mul(x = var_2894_0, y = var_2898)[name = string("op_2899")]; |
| fp32 var_2900 = const()[name = string("op_2900"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_39 = add(x = var_2899, y = var_2900)[name = string("gate_output_39")]; |
| tensor<int32, [3]> var_2905 = const()[name = string("op_2905"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_2906 = reshape(shape = var_2905, x = gate_output_39)[name = string("op_2906")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_39 = mul(x = var_2906, y = position_bias)[name = string("gated_position_bias_39")]; |
| tensor<int32, [3]> query_39_perm_0 = const()[name = string("query_39_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_39 = transpose(perm = query_39_perm_0, x = hidden_states_111)[name = string("transpose_177")]; |
| tensor<fp32, [?, 1, 1024]> q_77 = linear(bias = net_wavlm_encoder_layers_19_attention_q_proj_bias, weight = net_wavlm_encoder_layers_19_attention_q_proj_weight_quantized, x = query_39)[name = string("linear_135")]; |
| tensor<fp32, [?, 1, 1024]> k_77 = linear(bias = net_wavlm_encoder_layers_19_attention_k_proj_bias, weight = net_wavlm_encoder_layers_19_attention_k_proj_weight_quantized, x = query_39)[name = string("linear_136")]; |
| tensor<fp32, [?, 1, 1024]> v_77 = linear(bias = net_wavlm_encoder_layers_19_attention_v_proj_bias, weight = net_wavlm_encoder_layers_19_attention_v_proj_weight_quantized, x = query_39)[name = string("linear_137")]; |
| tensor<int32, [3]> concat_156x = const()[name = string("concat_156x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2933 = reshape(shape = concat_156x, x = q_77)[name = string("op_2933")]; |
| tensor<int32, [3]> concat_157x = const()[name = string("concat_157x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2940 = reshape(shape = concat_157x, x = k_77)[name = string("op_2940")]; |
| tensor<int32, [3]> concat_158x = const()[name = string("concat_158x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_2947 = reshape(shape = concat_158x, x = v_77)[name = string("op_2947")]; |
| tensor<int32, [3]> v_79_perm_0 = const()[name = string("v_79_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_2949 = const()[name = string("op_2949"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_39 = mul(x = var_2933, y = var_2949)[name = string("q_scaled_39")]; |
| bool input_441_bmm_transpose_x_0 = const()[name = string("input_441_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_441_bmm_transpose_y_0 = const()[name = string("input_441_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_134_perm_0 = const()[name = string("transpose_134_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_135_perm_0 = const()[name = string("transpose_135_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_135 = transpose(perm = transpose_135_perm_0, x = var_2940)[name = string("transpose_175")]; |
| tensor<fp32, [16, ?, 64]> transpose_134 = transpose(perm = transpose_134_perm_0, x = q_scaled_39)[name = string("transpose_176")]; |
| tensor<fp32, [16, ?, ?]> input_441_bmm = matmul(transpose_x = input_441_bmm_transpose_x_0, transpose_y = input_441_bmm_transpose_y_0, x = transpose_134, y = transpose_135)[name = string("input_441_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_441 = add(x = gated_position_bias_39, y = input_441_bmm)[name = string("input_441")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_77 = softmax(axis = var_9, x = input_441)[name = string("attn_output_weights_77")]; |
| bool attn_output_153_transpose_x_0 = const()[name = string("attn_output_153_transpose_x_0"), val = bool(false)]; |
| bool attn_output_153_transpose_y_0 = const()[name = string("attn_output_153_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_79 = transpose(perm = v_79_perm_0, x = var_2947)[name = string("transpose_174")]; |
| tensor<fp32, [16, ?, 64]> attn_output_153 = matmul(transpose_x = attn_output_153_transpose_x_0, transpose_y = attn_output_153_transpose_y_0, x = attn_output_weights_77, y = v_79)[name = string("attn_output_153")]; |
| tensor<int32, [3]> var_2955_perm_0 = const()[name = string("op_2955_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_159x = const()[name = string("concat_159x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_2955 = transpose(perm = var_2955_perm_0, x = attn_output_153)[name = string("transpose_173")]; |
| tensor<fp32, [?, 1024]> attn_output_155 = reshape(shape = concat_159x, x = var_2955)[name = string("attn_output_155")]; |
| tensor<fp32, [?, 1024]> attn_output_157 = linear(bias = net_wavlm_encoder_layers_19_attention_out_proj_bias, weight = net_wavlm_encoder_layers_19_attention_out_proj_weight_quantized, x = attn_output_155)[name = string("linear_138")]; |
| tensor<int32, [3]> concat_160x = const()[name = string("concat_160x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_159 = reshape(shape = concat_160x, x = attn_output_157)[name = string("attn_output_159")]; |
| tensor<int32, [3]> input_443_perm_0 = const()[name = string("input_443_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_443 = transpose(perm = input_443_perm_0, x = attn_output_159)[name = string("transpose_172")]; |
| tensor<fp32, [1, ?, 1024]> input_445 = add(x = input_437, y = input_443)[name = string("input_445")]; |
| tensor<int32, [1]> input_447_axes_0 = const()[name = string("input_447_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_447 = layer_norm(axes = input_447_axes_0, beta = net_wavlm_encoder_layers_19_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_19_final_layer_norm_weight, x = input_445)[name = string("input_447")]; |
| tensor<fp32, [1, ?, 4096]> input_449 = linear(bias = net_wavlm_encoder_layers_19_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_19_feed_forward_intermediate_dense_weight_quantized, x = input_447)[name = string("linear_139")]; |
| string input_451_mode_0 = const()[name = string("input_451_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_451 = gelu(mode = input_451_mode_0, x = input_449)[name = string("input_451")]; |
| tensor<fp32, [1, ?, 1024]> input_455 = linear(bias = net_wavlm_encoder_layers_19_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_19_feed_forward_output_dense_weight_quantized, x = input_451)[name = string("linear_140")]; |
| tensor<fp32, [1, ?, 1024]> input_457 = add(x = input_445, y = input_455)[name = string("input_457")]; |
| tensor<int32, [1]> hidden_states_115_axes_0 = const()[name = string("hidden_states_115_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_115 = layer_norm(axes = hidden_states_115_axes_0, beta = net_wavlm_encoder_layers_20_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_20_layer_norm_weight, x = input_457)[name = string("hidden_states_115")]; |
| tensor<int32, [4]> concat_161x = const()[name = string("concat_161x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_41 = reshape(shape = concat_161x, x = hidden_states_115)[name = string("gated_hidden_states_41")]; |
| tensor<int32, [4]> var_3016 = const()[name = string("op_3016"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_459 = transpose(perm = var_3016, x = gated_hidden_states_41)[name = string("transpose_171")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_81 = linear(bias = net_wavlm_encoder_layers_20_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_20_attention_gru_rel_pos_linear_weight, x = input_459)[name = string("linear_141")]; |
| tensor<int32, [5]> concat_162x = const()[name = string("concat_162x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_3025 = reshape(shape = concat_162x, x = relative_position_proj_81)[name = string("op_3025")]; |
| tensor<int32, [1]> relative_position_proj_83_axes_0 = const()[name = string("relative_position_proj_83_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_83_keep_dims_0 = const()[name = string("relative_position_proj_83_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_83 = reduce_sum(axes = relative_position_proj_83_axes_0, keep_dims = relative_position_proj_83_keep_dims_0, x = var_3025)[name = string("relative_position_proj_83")]; |
| tensor<fp32, [1, 16, ?, 2]> var_3028 = sigmoid(x = relative_position_proj_83)[name = string("op_3028")]; |
| tensor<int32, [2]> var_3029_split_sizes_0 = const()[name = string("op_3029_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_3029_axis_0 = const()[name = string("op_3029_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3029_0, tensor<fp32, [1, 16, ?, 1]> var_3029_1 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = var_3028)[name = string("op_3029")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3031 = mul(x = var_3029_1, y = net_wavlm_encoder_layers_20_attention_gru_rel_pos_const)[name = string("op_3031")]; |
| fp32 var_3032 = const()[name = string("op_3032"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3033 = sub(x = var_3031, y = var_3032)[name = string("op_3033")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3034 = mul(x = var_3029_0, y = var_3033)[name = string("op_3034")]; |
| fp32 var_3035 = const()[name = string("op_3035"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_41 = add(x = var_3034, y = var_3035)[name = string("gate_output_41")]; |
| tensor<int32, [3]> var_3040 = const()[name = string("op_3040"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_3041 = reshape(shape = var_3040, x = gate_output_41)[name = string("op_3041")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_41 = mul(x = var_3041, y = position_bias)[name = string("gated_position_bias_41")]; |
| tensor<int32, [3]> query_41_perm_0 = const()[name = string("query_41_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_41 = transpose(perm = query_41_perm_0, x = hidden_states_115)[name = string("transpose_170")]; |
| tensor<fp32, [?, 1, 1024]> q_81 = linear(bias = net_wavlm_encoder_layers_20_attention_q_proj_bias, weight = net_wavlm_encoder_layers_20_attention_q_proj_weight_quantized, x = query_41)[name = string("linear_142")]; |
| tensor<fp32, [?, 1, 1024]> k_81 = linear(bias = net_wavlm_encoder_layers_20_attention_k_proj_bias, weight = net_wavlm_encoder_layers_20_attention_k_proj_weight_quantized, x = query_41)[name = string("linear_143")]; |
| tensor<fp32, [?, 1, 1024]> v_81 = linear(bias = net_wavlm_encoder_layers_20_attention_v_proj_bias, weight = net_wavlm_encoder_layers_20_attention_v_proj_weight_quantized, x = query_41)[name = string("linear_144")]; |
| tensor<int32, [3]> concat_164x = const()[name = string("concat_164x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3068 = reshape(shape = concat_164x, x = q_81)[name = string("op_3068")]; |
| tensor<int32, [3]> concat_165x = const()[name = string("concat_165x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3075 = reshape(shape = concat_165x, x = k_81)[name = string("op_3075")]; |
| tensor<int32, [3]> concat_166x = const()[name = string("concat_166x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3082 = reshape(shape = concat_166x, x = v_81)[name = string("op_3082")]; |
| tensor<int32, [3]> v_83_perm_0 = const()[name = string("v_83_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_3084 = const()[name = string("op_3084"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_41 = mul(x = var_3068, y = var_3084)[name = string("q_scaled_41")]; |
| bool input_461_bmm_transpose_x_0 = const()[name = string("input_461_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_461_bmm_transpose_y_0 = const()[name = string("input_461_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_136_perm_0 = const()[name = string("transpose_136_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_137_perm_0 = const()[name = string("transpose_137_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_137 = transpose(perm = transpose_137_perm_0, x = var_3075)[name = string("transpose_168")]; |
| tensor<fp32, [16, ?, 64]> transpose_136 = transpose(perm = transpose_136_perm_0, x = q_scaled_41)[name = string("transpose_169")]; |
| tensor<fp32, [16, ?, ?]> input_461_bmm = matmul(transpose_x = input_461_bmm_transpose_x_0, transpose_y = input_461_bmm_transpose_y_0, x = transpose_136, y = transpose_137)[name = string("input_461_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_461 = add(x = gated_position_bias_41, y = input_461_bmm)[name = string("input_461")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_81 = softmax(axis = var_9, x = input_461)[name = string("attn_output_weights_81")]; |
| bool attn_output_161_transpose_x_0 = const()[name = string("attn_output_161_transpose_x_0"), val = bool(false)]; |
| bool attn_output_161_transpose_y_0 = const()[name = string("attn_output_161_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_83 = transpose(perm = v_83_perm_0, x = var_3082)[name = string("transpose_167")]; |
| tensor<fp32, [16, ?, 64]> attn_output_161 = matmul(transpose_x = attn_output_161_transpose_x_0, transpose_y = attn_output_161_transpose_y_0, x = attn_output_weights_81, y = v_83)[name = string("attn_output_161")]; |
| tensor<int32, [3]> var_3090_perm_0 = const()[name = string("op_3090_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_167x = const()[name = string("concat_167x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_3090 = transpose(perm = var_3090_perm_0, x = attn_output_161)[name = string("transpose_166")]; |
| tensor<fp32, [?, 1024]> attn_output_163 = reshape(shape = concat_167x, x = var_3090)[name = string("attn_output_163")]; |
| tensor<fp32, [?, 1024]> attn_output_165 = linear(bias = net_wavlm_encoder_layers_20_attention_out_proj_bias, weight = net_wavlm_encoder_layers_20_attention_out_proj_weight_quantized, x = attn_output_163)[name = string("linear_145")]; |
| tensor<int32, [3]> concat_168x = const()[name = string("concat_168x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_167 = reshape(shape = concat_168x, x = attn_output_165)[name = string("attn_output_167")]; |
| tensor<int32, [3]> input_463_perm_0 = const()[name = string("input_463_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_463 = transpose(perm = input_463_perm_0, x = attn_output_167)[name = string("transpose_165")]; |
| tensor<fp32, [1, ?, 1024]> input_465 = add(x = input_457, y = input_463)[name = string("input_465")]; |
| tensor<int32, [1]> input_467_axes_0 = const()[name = string("input_467_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_467 = layer_norm(axes = input_467_axes_0, beta = net_wavlm_encoder_layers_20_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_20_final_layer_norm_weight, x = input_465)[name = string("input_467")]; |
| tensor<fp32, [1, ?, 4096]> input_469 = linear(bias = net_wavlm_encoder_layers_20_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_20_feed_forward_intermediate_dense_weight_quantized, x = input_467)[name = string("linear_146")]; |
| string input_471_mode_0 = const()[name = string("input_471_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_471 = gelu(mode = input_471_mode_0, x = input_469)[name = string("input_471")]; |
| tensor<fp32, [1, ?, 1024]> input_475 = linear(bias = net_wavlm_encoder_layers_20_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_20_feed_forward_output_dense_weight_quantized, x = input_471)[name = string("linear_147")]; |
| tensor<fp32, [1, ?, 1024]> input_477 = add(x = input_465, y = input_475)[name = string("input_477")]; |
| tensor<int32, [1]> hidden_states_119_axes_0 = const()[name = string("hidden_states_119_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_119 = layer_norm(axes = hidden_states_119_axes_0, beta = net_wavlm_encoder_layers_21_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_21_layer_norm_weight, x = input_477)[name = string("hidden_states_119")]; |
| tensor<int32, [4]> concat_169x = const()[name = string("concat_169x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_43 = reshape(shape = concat_169x, x = hidden_states_119)[name = string("gated_hidden_states_43")]; |
| tensor<int32, [4]> var_3151 = const()[name = string("op_3151"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_479 = transpose(perm = var_3151, x = gated_hidden_states_43)[name = string("transpose_164")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_85 = linear(bias = net_wavlm_encoder_layers_21_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_21_attention_gru_rel_pos_linear_weight, x = input_479)[name = string("linear_148")]; |
| tensor<int32, [5]> concat_170x = const()[name = string("concat_170x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_3160 = reshape(shape = concat_170x, x = relative_position_proj_85)[name = string("op_3160")]; |
| tensor<int32, [1]> relative_position_proj_87_axes_0 = const()[name = string("relative_position_proj_87_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_87_keep_dims_0 = const()[name = string("relative_position_proj_87_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_87 = reduce_sum(axes = relative_position_proj_87_axes_0, keep_dims = relative_position_proj_87_keep_dims_0, x = var_3160)[name = string("relative_position_proj_87")]; |
| tensor<fp32, [1, 16, ?, 2]> var_3163 = sigmoid(x = relative_position_proj_87)[name = string("op_3163")]; |
| tensor<int32, [2]> var_3164_split_sizes_0 = const()[name = string("op_3164_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_3164_axis_0 = const()[name = string("op_3164_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3164_0, tensor<fp32, [1, 16, ?, 1]> var_3164_1 = split(axis = var_3164_axis_0, split_sizes = var_3164_split_sizes_0, x = var_3163)[name = string("op_3164")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3166 = mul(x = var_3164_1, y = net_wavlm_encoder_layers_21_attention_gru_rel_pos_const)[name = string("op_3166")]; |
| fp32 var_3167 = const()[name = string("op_3167"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3168 = sub(x = var_3166, y = var_3167)[name = string("op_3168")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3169 = mul(x = var_3164_0, y = var_3168)[name = string("op_3169")]; |
| fp32 var_3170 = const()[name = string("op_3170"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_43 = add(x = var_3169, y = var_3170)[name = string("gate_output_43")]; |
| tensor<int32, [3]> var_3175 = const()[name = string("op_3175"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_3176 = reshape(shape = var_3175, x = gate_output_43)[name = string("op_3176")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_43 = mul(x = var_3176, y = position_bias)[name = string("gated_position_bias_43")]; |
| tensor<int32, [3]> query_43_perm_0 = const()[name = string("query_43_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_43 = transpose(perm = query_43_perm_0, x = hidden_states_119)[name = string("transpose_163")]; |
| tensor<fp32, [?, 1, 1024]> q_85 = linear(bias = net_wavlm_encoder_layers_21_attention_q_proj_bias, weight = net_wavlm_encoder_layers_21_attention_q_proj_weight_quantized, x = query_43)[name = string("linear_149")]; |
| tensor<fp32, [?, 1, 1024]> k_85 = linear(bias = net_wavlm_encoder_layers_21_attention_k_proj_bias, weight = net_wavlm_encoder_layers_21_attention_k_proj_weight_quantized, x = query_43)[name = string("linear_150")]; |
| tensor<fp32, [?, 1, 1024]> v_85 = linear(bias = net_wavlm_encoder_layers_21_attention_v_proj_bias, weight = net_wavlm_encoder_layers_21_attention_v_proj_weight_quantized, x = query_43)[name = string("linear_151")]; |
| tensor<int32, [3]> concat_172x = const()[name = string("concat_172x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3203 = reshape(shape = concat_172x, x = q_85)[name = string("op_3203")]; |
| tensor<int32, [3]> concat_173x = const()[name = string("concat_173x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3210 = reshape(shape = concat_173x, x = k_85)[name = string("op_3210")]; |
| tensor<int32, [3]> concat_174x = const()[name = string("concat_174x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3217 = reshape(shape = concat_174x, x = v_85)[name = string("op_3217")]; |
| tensor<int32, [3]> v_87_perm_0 = const()[name = string("v_87_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_3219 = const()[name = string("op_3219"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_43 = mul(x = var_3203, y = var_3219)[name = string("q_scaled_43")]; |
| bool input_481_bmm_transpose_x_0 = const()[name = string("input_481_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_481_bmm_transpose_y_0 = const()[name = string("input_481_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_138_perm_0 = const()[name = string("transpose_138_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_139_perm_0 = const()[name = string("transpose_139_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_139 = transpose(perm = transpose_139_perm_0, x = var_3210)[name = string("transpose_161")]; |
| tensor<fp32, [16, ?, 64]> transpose_138 = transpose(perm = transpose_138_perm_0, x = q_scaled_43)[name = string("transpose_162")]; |
| tensor<fp32, [16, ?, ?]> input_481_bmm = matmul(transpose_x = input_481_bmm_transpose_x_0, transpose_y = input_481_bmm_transpose_y_0, x = transpose_138, y = transpose_139)[name = string("input_481_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_481 = add(x = gated_position_bias_43, y = input_481_bmm)[name = string("input_481")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_85 = softmax(axis = var_9, x = input_481)[name = string("attn_output_weights_85")]; |
| bool attn_output_169_transpose_x_0 = const()[name = string("attn_output_169_transpose_x_0"), val = bool(false)]; |
| bool attn_output_169_transpose_y_0 = const()[name = string("attn_output_169_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_87 = transpose(perm = v_87_perm_0, x = var_3217)[name = string("transpose_160")]; |
| tensor<fp32, [16, ?, 64]> attn_output_169 = matmul(transpose_x = attn_output_169_transpose_x_0, transpose_y = attn_output_169_transpose_y_0, x = attn_output_weights_85, y = v_87)[name = string("attn_output_169")]; |
| tensor<int32, [3]> var_3225_perm_0 = const()[name = string("op_3225_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_175x = const()[name = string("concat_175x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_3225 = transpose(perm = var_3225_perm_0, x = attn_output_169)[name = string("transpose_159")]; |
| tensor<fp32, [?, 1024]> attn_output_171 = reshape(shape = concat_175x, x = var_3225)[name = string("attn_output_171")]; |
| tensor<fp32, [?, 1024]> attn_output_173 = linear(bias = net_wavlm_encoder_layers_21_attention_out_proj_bias, weight = net_wavlm_encoder_layers_21_attention_out_proj_weight_quantized, x = attn_output_171)[name = string("linear_152")]; |
| tensor<int32, [3]> concat_176x = const()[name = string("concat_176x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_175 = reshape(shape = concat_176x, x = attn_output_173)[name = string("attn_output_175")]; |
| tensor<int32, [3]> input_483_perm_0 = const()[name = string("input_483_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_483 = transpose(perm = input_483_perm_0, x = attn_output_175)[name = string("transpose_158")]; |
| tensor<fp32, [1, ?, 1024]> input_485 = add(x = input_477, y = input_483)[name = string("input_485")]; |
| tensor<int32, [1]> input_487_axes_0 = const()[name = string("input_487_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_487 = layer_norm(axes = input_487_axes_0, beta = net_wavlm_encoder_layers_21_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_21_final_layer_norm_weight, x = input_485)[name = string("input_487")]; |
| tensor<fp32, [1, ?, 4096]> input_489 = linear(bias = net_wavlm_encoder_layers_21_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_21_feed_forward_intermediate_dense_weight_quantized, x = input_487)[name = string("linear_153")]; |
| string input_491_mode_0 = const()[name = string("input_491_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_491 = gelu(mode = input_491_mode_0, x = input_489)[name = string("input_491")]; |
| tensor<fp32, [1, ?, 1024]> input_495 = linear(bias = net_wavlm_encoder_layers_21_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_21_feed_forward_output_dense_weight_quantized, x = input_491)[name = string("linear_154")]; |
| tensor<fp32, [1, ?, 1024]> input_497 = add(x = input_485, y = input_495)[name = string("input_497")]; |
| tensor<int32, [1]> hidden_states_123_axes_0 = const()[name = string("hidden_states_123_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_123 = layer_norm(axes = hidden_states_123_axes_0, beta = net_wavlm_encoder_layers_22_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_22_layer_norm_weight, x = input_497)[name = string("hidden_states_123")]; |
| tensor<int32, [4]> concat_177x = const()[name = string("concat_177x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states_45 = reshape(shape = concat_177x, x = hidden_states_123)[name = string("gated_hidden_states_45")]; |
| tensor<int32, [4]> var_3286 = const()[name = string("op_3286"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_499 = transpose(perm = var_3286, x = gated_hidden_states_45)[name = string("transpose_157")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_89 = linear(bias = net_wavlm_encoder_layers_22_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_22_attention_gru_rel_pos_linear_weight, x = input_499)[name = string("linear_155")]; |
| tensor<int32, [5]> concat_178x = const()[name = string("concat_178x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_3295 = reshape(shape = concat_178x, x = relative_position_proj_89)[name = string("op_3295")]; |
| tensor<int32, [1]> relative_position_proj_91_axes_0 = const()[name = string("relative_position_proj_91_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_91_keep_dims_0 = const()[name = string("relative_position_proj_91_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj_91 = reduce_sum(axes = relative_position_proj_91_axes_0, keep_dims = relative_position_proj_91_keep_dims_0, x = var_3295)[name = string("relative_position_proj_91")]; |
| tensor<fp32, [1, 16, ?, 2]> var_3298 = sigmoid(x = relative_position_proj_91)[name = string("op_3298")]; |
| tensor<int32, [2]> var_3299_split_sizes_0 = const()[name = string("op_3299_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_3299_axis_0 = const()[name = string("op_3299_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3299_0, tensor<fp32, [1, 16, ?, 1]> var_3299_1 = split(axis = var_3299_axis_0, split_sizes = var_3299_split_sizes_0, x = var_3298)[name = string("op_3299")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3301 = mul(x = var_3299_1, y = net_wavlm_encoder_layers_22_attention_gru_rel_pos_const)[name = string("op_3301")]; |
| fp32 var_3302 = const()[name = string("op_3302"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3303 = sub(x = var_3301, y = var_3302)[name = string("op_3303")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3304 = mul(x = var_3299_0, y = var_3303)[name = string("op_3304")]; |
| fp32 var_3305 = const()[name = string("op_3305"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output_45 = add(x = var_3304, y = var_3305)[name = string("gate_output_45")]; |
| tensor<int32, [3]> var_3310 = const()[name = string("op_3310"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_3311 = reshape(shape = var_3310, x = gate_output_45)[name = string("op_3311")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias_45 = mul(x = var_3311, y = position_bias)[name = string("gated_position_bias_45")]; |
| tensor<int32, [3]> query_45_perm_0 = const()[name = string("query_45_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query_45 = transpose(perm = query_45_perm_0, x = hidden_states_123)[name = string("transpose_156")]; |
| tensor<fp32, [?, 1, 1024]> q_89 = linear(bias = net_wavlm_encoder_layers_22_attention_q_proj_bias, weight = net_wavlm_encoder_layers_22_attention_q_proj_weight_quantized, x = query_45)[name = string("linear_156")]; |
| tensor<fp32, [?, 1, 1024]> k_89 = linear(bias = net_wavlm_encoder_layers_22_attention_k_proj_bias, weight = net_wavlm_encoder_layers_22_attention_k_proj_weight_quantized, x = query_45)[name = string("linear_157")]; |
| tensor<fp32, [?, 1, 1024]> v_89 = linear(bias = net_wavlm_encoder_layers_22_attention_v_proj_bias, weight = net_wavlm_encoder_layers_22_attention_v_proj_weight_quantized, x = query_45)[name = string("linear_158")]; |
| tensor<int32, [3]> concat_180x = const()[name = string("concat_180x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3338 = reshape(shape = concat_180x, x = q_89)[name = string("op_3338")]; |
| tensor<int32, [3]> concat_181x = const()[name = string("concat_181x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3345 = reshape(shape = concat_181x, x = k_89)[name = string("op_3345")]; |
| tensor<int32, [3]> concat_182x = const()[name = string("concat_182x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3352 = reshape(shape = concat_182x, x = v_89)[name = string("op_3352")]; |
| tensor<int32, [3]> v_91_perm_0 = const()[name = string("v_91_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_3354 = const()[name = string("op_3354"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled_45 = mul(x = var_3338, y = var_3354)[name = string("q_scaled_45")]; |
| bool input_501_bmm_transpose_x_0 = const()[name = string("input_501_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_501_bmm_transpose_y_0 = const()[name = string("input_501_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_140_perm_0 = const()[name = string("transpose_140_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_141_perm_0 = const()[name = string("transpose_141_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_141 = transpose(perm = transpose_141_perm_0, x = var_3345)[name = string("transpose_154")]; |
| tensor<fp32, [16, ?, 64]> transpose_140 = transpose(perm = transpose_140_perm_0, x = q_scaled_45)[name = string("transpose_155")]; |
| tensor<fp32, [16, ?, ?]> input_501_bmm = matmul(transpose_x = input_501_bmm_transpose_x_0, transpose_y = input_501_bmm_transpose_y_0, x = transpose_140, y = transpose_141)[name = string("input_501_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_501 = add(x = gated_position_bias_45, y = input_501_bmm)[name = string("input_501")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_89 = softmax(axis = var_9, x = input_501)[name = string("attn_output_weights_89")]; |
| bool attn_output_177_transpose_x_0 = const()[name = string("attn_output_177_transpose_x_0"), val = bool(false)]; |
| bool attn_output_177_transpose_y_0 = const()[name = string("attn_output_177_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v_91 = transpose(perm = v_91_perm_0, x = var_3352)[name = string("transpose_153")]; |
| tensor<fp32, [16, ?, 64]> attn_output_177 = matmul(transpose_x = attn_output_177_transpose_x_0, transpose_y = attn_output_177_transpose_y_0, x = attn_output_weights_89, y = v_91)[name = string("attn_output_177")]; |
| tensor<int32, [3]> var_3360_perm_0 = const()[name = string("op_3360_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_183x = const()[name = string("concat_183x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_3360 = transpose(perm = var_3360_perm_0, x = attn_output_177)[name = string("transpose_152")]; |
| tensor<fp32, [?, 1024]> attn_output_179 = reshape(shape = concat_183x, x = var_3360)[name = string("attn_output_179")]; |
| tensor<fp32, [?, 1024]> attn_output_181 = linear(bias = net_wavlm_encoder_layers_22_attention_out_proj_bias, weight = net_wavlm_encoder_layers_22_attention_out_proj_weight_quantized, x = attn_output_179)[name = string("linear_159")]; |
| tensor<int32, [3]> concat_184x = const()[name = string("concat_184x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output_183 = reshape(shape = concat_184x, x = attn_output_181)[name = string("attn_output_183")]; |
| tensor<int32, [3]> input_503_perm_0 = const()[name = string("input_503_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_503 = transpose(perm = input_503_perm_0, x = attn_output_183)[name = string("transpose_151")]; |
| tensor<fp32, [1, ?, 1024]> input_505 = add(x = input_497, y = input_503)[name = string("input_505")]; |
| tensor<int32, [1]> input_507_axes_0 = const()[name = string("input_507_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_507 = layer_norm(axes = input_507_axes_0, beta = net_wavlm_encoder_layers_22_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_22_final_layer_norm_weight, x = input_505)[name = string("input_507")]; |
| tensor<fp32, [1, ?, 4096]> input_509 = linear(bias = net_wavlm_encoder_layers_22_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_22_feed_forward_intermediate_dense_weight_quantized, x = input_507)[name = string("linear_160")]; |
| string input_511_mode_0 = const()[name = string("input_511_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_511 = gelu(mode = input_511_mode_0, x = input_509)[name = string("input_511")]; |
| tensor<fp32, [1, ?, 1024]> input_515 = linear(bias = net_wavlm_encoder_layers_22_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_22_feed_forward_output_dense_weight_quantized, x = input_511)[name = string("linear_161")]; |
| tensor<fp32, [1, ?, 1024]> input_517 = add(x = input_505, y = input_515)[name = string("input_517")]; |
| tensor<int32, [1]> hidden_states_127_axes_0 = const()[name = string("hidden_states_127_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> hidden_states_127 = layer_norm(axes = hidden_states_127_axes_0, beta = net_wavlm_encoder_layers_23_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_23_layer_norm_weight, x = input_517)[name = string("hidden_states_127")]; |
| tensor<int32, [4]> concat_185x = const()[name = string("concat_185x"), val = tensor<int32, [4]>([1, -1, 16, 64])]; |
| tensor<fp32, [1, ?, 16, 64]> gated_hidden_states = reshape(shape = concat_185x, x = hidden_states_127)[name = string("gated_hidden_states")]; |
| tensor<int32, [4]> var_3421 = const()[name = string("op_3421"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp32, [1, 16, ?, 64]> input_519 = transpose(perm = var_3421, x = gated_hidden_states)[name = string("transpose_150")]; |
| tensor<fp32, [1, 16, ?, 8]> relative_position_proj_93 = linear(bias = net_wavlm_encoder_layers_23_attention_gru_rel_pos_linear_bias, weight = net_wavlm_encoder_layers_23_attention_gru_rel_pos_linear_weight, x = input_519)[name = string("linear_162")]; |
| tensor<int32, [5]> concat_186x = const()[name = string("concat_186x"), val = tensor<int32, [5]>([1, 16, -1, 2, 4])]; |
| tensor<fp32, [1, 16, ?, 2, 4]> var_3430 = reshape(shape = concat_186x, x = relative_position_proj_93)[name = string("op_3430")]; |
| tensor<int32, [1]> relative_position_proj_axes_0 = const()[name = string("relative_position_proj_axes_0"), val = tensor<int32, [1]>([-1])]; |
| bool relative_position_proj_keep_dims_0 = const()[name = string("relative_position_proj_keep_dims_0"), val = bool(false)]; |
| tensor<fp32, [1, 16, ?, 2]> relative_position_proj = reduce_sum(axes = relative_position_proj_axes_0, keep_dims = relative_position_proj_keep_dims_0, x = var_3430)[name = string("relative_position_proj")]; |
| tensor<fp32, [1, 16, ?, 2]> var_3433 = sigmoid(x = relative_position_proj)[name = string("op_3433")]; |
| tensor<int32, [2]> var_3434_split_sizes_0 = const()[name = string("op_3434_split_sizes_0"), val = tensor<int32, [2]>([1, 1])]; |
| int32 var_3434_axis_0 = const()[name = string("op_3434_axis_0"), val = int32(-1)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3434_0, tensor<fp32, [1, 16, ?, 1]> var_3434_1 = split(axis = var_3434_axis_0, split_sizes = var_3434_split_sizes_0, x = var_3433)[name = string("op_3434")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3436 = mul(x = var_3434_1, y = net_wavlm_encoder_layers_23_attention_gru_rel_pos_const)[name = string("op_3436")]; |
| fp32 var_3437 = const()[name = string("op_3437"), val = fp32(0x1p+0)]; |
| tensor<fp32, [1, 16, ?, 1]> var_3438 = sub(x = var_3436, y = var_3437)[name = string("op_3438")]; |
| tensor<fp32, [1, 16, ?, 1]> var_3439 = mul(x = var_3434_0, y = var_3438)[name = string("op_3439")]; |
| fp32 var_3440 = const()[name = string("op_3440"), val = fp32(0x1p+1)]; |
| tensor<fp32, [1, 16, ?, 1]> gate_output = add(x = var_3439, y = var_3440)[name = string("gate_output")]; |
| tensor<int32, [3]> var_3445 = const()[name = string("op_3445"), val = tensor<int32, [3]>([16, -1, 1])]; |
| tensor<fp32, [16, ?, 1]> var_3446 = reshape(shape = var_3445, x = gate_output)[name = string("op_3446")]; |
| tensor<fp32, [16, ?, ?]> gated_position_bias = mul(x = var_3446, y = position_bias)[name = string("gated_position_bias")]; |
| tensor<int32, [3]> query_perm_0 = const()[name = string("query_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [?, 1, 1024]> query = transpose(perm = query_perm_0, x = hidden_states_127)[name = string("transpose_149")]; |
| tensor<fp32, [?, 1, 1024]> q_93 = linear(bias = net_wavlm_encoder_layers_23_attention_q_proj_bias, weight = net_wavlm_encoder_layers_23_attention_q_proj_weight_quantized, x = query)[name = string("linear_163")]; |
| tensor<fp32, [?, 1, 1024]> k_93 = linear(bias = net_wavlm_encoder_layers_23_attention_k_proj_bias, weight = net_wavlm_encoder_layers_23_attention_k_proj_weight_quantized, x = query)[name = string("linear_164")]; |
| tensor<fp32, [?, 1, 1024]> v_93 = linear(bias = net_wavlm_encoder_layers_23_attention_v_proj_bias, weight = net_wavlm_encoder_layers_23_attention_v_proj_weight_quantized, x = query)[name = string("linear_165")]; |
| tensor<int32, [3]> concat_188x = const()[name = string("concat_188x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3473 = reshape(shape = concat_188x, x = q_93)[name = string("op_3473")]; |
| tensor<int32, [3]> concat_189x = const()[name = string("concat_189x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3480 = reshape(shape = concat_189x, x = k_93)[name = string("op_3480")]; |
| tensor<int32, [3]> concat_190x = const()[name = string("concat_190x"), val = tensor<int32, [3]>([-1, 16, 64])]; |
| tensor<fp32, [?, 16, 64]> var_3487 = reshape(shape = concat_190x, x = v_93)[name = string("op_3487")]; |
| tensor<int32, [3]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| fp32 var_3489 = const()[name = string("op_3489"), val = fp32(0x1p-3)]; |
| tensor<fp32, [?, 16, 64]> q_scaled = mul(x = var_3473, y = var_3489)[name = string("q_scaled")]; |
| bool input_521_bmm_transpose_x_0 = const()[name = string("input_521_bmm_transpose_x_0"), val = bool(false)]; |
| bool input_521_bmm_transpose_y_0 = const()[name = string("input_521_bmm_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [3]> transpose_142_perm_0 = const()[name = string("transpose_142_perm_0"), val = tensor<int32, [3]>([1, -3, -1])]; |
| tensor<int32, [3]> transpose_143_perm_0 = const()[name = string("transpose_143_perm_0"), val = tensor<int32, [3]>([1, -1, -3])]; |
| tensor<fp32, [16, 64, ?]> transpose_143 = transpose(perm = transpose_143_perm_0, x = var_3480)[name = string("transpose_147")]; |
| tensor<fp32, [16, ?, 64]> transpose_142 = transpose(perm = transpose_142_perm_0, x = q_scaled)[name = string("transpose_148")]; |
| tensor<fp32, [16, ?, ?]> input_521_bmm = matmul(transpose_x = input_521_bmm_transpose_x_0, transpose_y = input_521_bmm_transpose_y_0, x = transpose_142, y = transpose_143)[name = string("input_521_bmm")]; |
| tensor<fp32, [16, ?, ?]> input_521 = add(x = gated_position_bias, y = input_521_bmm)[name = string("input_521")]; |
| tensor<fp32, [16, ?, ?]> attn_output_weights_93 = softmax(axis = var_9, x = input_521)[name = string("attn_output_weights_93")]; |
| bool attn_output_185_transpose_x_0 = const()[name = string("attn_output_185_transpose_x_0"), val = bool(false)]; |
| bool attn_output_185_transpose_y_0 = const()[name = string("attn_output_185_transpose_y_0"), val = bool(false)]; |
| tensor<fp32, [16, ?, 64]> v = transpose(perm = v_perm_0, x = var_3487)[name = string("transpose_146")]; |
| tensor<fp32, [16, ?, 64]> attn_output_185 = matmul(transpose_x = attn_output_185_transpose_x_0, transpose_y = attn_output_185_transpose_y_0, x = attn_output_weights_93, y = v)[name = string("attn_output_185")]; |
| tensor<int32, [3]> var_3495_perm_0 = const()[name = string("op_3495_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<int32, [2]> concat_191x = const()[name = string("concat_191x"), val = tensor<int32, [2]>([-1, 1024])]; |
| tensor<fp32, [?, 16, 64]> var_3495 = transpose(perm = var_3495_perm_0, x = attn_output_185)[name = string("transpose_145")]; |
| tensor<fp32, [?, 1024]> attn_output_187 = reshape(shape = concat_191x, x = var_3495)[name = string("attn_output_187")]; |
| tensor<fp32, [?, 1024]> attn_output_189 = linear(bias = net_wavlm_encoder_layers_23_attention_out_proj_bias, weight = net_wavlm_encoder_layers_23_attention_out_proj_weight_quantized, x = attn_output_187)[name = string("linear_166")]; |
| tensor<int32, [3]> concat_192x = const()[name = string("concat_192x"), val = tensor<int32, [3]>([-1, 1, 1024])]; |
| tensor<fp32, [?, 1, 1024]> attn_output = reshape(shape = concat_192x, x = attn_output_189)[name = string("attn_output")]; |
| tensor<int32, [3]> input_523_perm_0 = const()[name = string("input_523_perm_0"), val = tensor<int32, [3]>([1, 0, 2])]; |
| tensor<fp32, [1, ?, 1024]> input_523 = transpose(perm = input_523_perm_0, x = attn_output)[name = string("transpose_144")]; |
| tensor<fp32, [1, ?, 1024]> input_525 = add(x = input_517, y = input_523)[name = string("input_525")]; |
| tensor<int32, [1]> input_527_axes_0 = const()[name = string("input_527_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_527 = layer_norm(axes = input_527_axes_0, beta = net_wavlm_encoder_layers_23_final_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layers_23_final_layer_norm_weight, x = input_525)[name = string("input_527")]; |
| tensor<fp32, [1, ?, 4096]> input_529 = linear(bias = net_wavlm_encoder_layers_23_feed_forward_intermediate_dense_bias, weight = net_wavlm_encoder_layers_23_feed_forward_intermediate_dense_weight_quantized, x = input_527)[name = string("linear_167")]; |
| string input_531_mode_0 = const()[name = string("input_531_mode_0"), val = string("EXACT")]; |
| tensor<fp32, [1, ?, 4096]> input_531 = gelu(mode = input_531_mode_0, x = input_529)[name = string("input_531")]; |
| tensor<fp32, [1, ?, 1024]> input_535 = linear(bias = net_wavlm_encoder_layers_23_feed_forward_output_dense_bias, weight = net_wavlm_encoder_layers_23_feed_forward_output_dense_weight_quantized, x = input_531)[name = string("linear_168")]; |
| tensor<fp32, [1, ?, 1024]> input_537 = add(x = input_525, y = input_535)[name = string("input_537")]; |
| tensor<int32, [1]> input_539_axes_0 = const()[name = string("input_539_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp32, [1, ?, 1024]> input_539 = layer_norm(axes = input_539_axes_0, beta = net_wavlm_encoder_layer_norm_bias, epsilon = var_10, gamma = net_wavlm_encoder_layer_norm_weight, x = input_537)[name = string("input_539")]; |
| tensor<fp32, [1, ?, 46]> var_3530 = linear(bias = net_lm_head_bias, weight = net_lm_head_weight_quantized, x = input_539)[name = string("linear_169")]; |
| int32 var_3531 = const()[name = string("op_3531"), val = int32(-1)]; |
| tensor<fp32, [1, ?, 46]> var_3533_softmax = softmax(axis = var_3531, x = var_3530)[name = string("op_3533_softmax")]; |
| fp32 var_3533_epsilon_0 = const()[name = string("op_3533_epsilon_0"), val = fp32(0x1p-149)]; |
| tensor<fp32, [1, ?, 46]> logprobs = log(epsilon = var_3533_epsilon_0, x = var_3533_softmax)[name = string("op_3533")]; |
| } -> (logprobs); |
| } |
