diff --git "a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil" "b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil" @@ -0,0 +1,1297 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 1, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, true, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39))[name = tensor("stacked")]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor([1, 3, 345])]; + tensor input_1 = reshape(shape = var_46, x = stacked)[name = tensor("input_1")]; + tensor var_48 = const()[name = tensor("op_48"), val = tensor(0x1p+0)]; + tensor var_54 = const()[name = tensor("op_54"), val = tensor(true)]; + tensor var_55 = const()[name = tensor("op_55"), val = tensor(0x1.4f8b58p-17)]; + tensor var_58 = const()[name = tensor("op_58"), val = tensor(0)]; + tensor var_60 = const()[name = tensor("op_60"), val = tensor(2)]; + tensor var_61 = const()[name = tensor("op_61"), val = tensor(-1)]; + tensor var_63 = const()[name = tensor("op_63"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_68 = const()[name = tensor("op_68"), val = tensor(0x1.5798eep-27)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_55, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor(0x1p-1)]; + tensor var_194 = mul(x = input_13, y = var_193)[name = tensor("op_194")]; + tensor input_15 = add(x = var_194, y = input_5)[name = tensor("input_15")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; + tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; + tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_1_squeeze_mask_0 = const()[name = tensor("prev_kv_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_1 = slice_by_index(begin = prev_kv_1_begin_0, end = prev_kv_1_end_0, end_mask = prev_kv_1_end_mask_0, squeeze_mask = prev_kv_1_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_1")]; + tensor prev_scale_1_begin_0 = const()[name = tensor("prev_scale_1_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_1_end_0 = const()[name = tensor("prev_scale_1_end_0"), val = tensor([1, 1])]; + tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; + tensor var_208 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_209 = const()[name = tensor("op_209"), val = tensor([1, 3, 4, 64])]; + tensor var_210 = reshape(shape = var_209, x = var_208)[name = tensor("op_210")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_214 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_215 = const()[name = tensor("op_215"), val = tensor(0x1p-3)]; + tensor var_216 = mul(x = var_214, y = var_215)[name = tensor("op_216")]; + tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 3, 4, 64])]; + tensor var_218 = reshape(shape = var_217, x = var_216)[name = tensor("op_218")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_222 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 3, 4, 64])]; + tensor var_224 = reshape(shape = var_223, x = var_222)[name = tensor("op_224")]; + tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; + tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; + tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; + tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_218)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_210)[name = tensor("transpose_58")]; + tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; + tensor var_234 = const()[name = tensor("op_234"), val = tensor([3, 1])]; + tensor var_235 = reshape(shape = var_234, x = sqrt_s_t_1)[name = tensor("op_235")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_235)[name = tensor("M_1")]; + tensor var_237 = mul(x = qk_1, y = M_1)[name = tensor("op_237")]; + tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; + tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_224)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_237, y = v_1)[name = tensor("inner_1")]; + tensor var_239_transpose_x_0 = const()[name = tensor("op_239_transpose_x_0"), val = tensor(false)]; + tensor var_239_transpose_y_0 = const()[name = tensor("op_239_transpose_y_0"), val = tensor(false)]; + tensor var_239 = matmul(transpose_x = var_239_transpose_x_0, transpose_y = var_239_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_239")]; + tensor var_240 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_240")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1, 3, 1])]; + tensor var_242 = reshape(shape = var_241, x = var_240)[name = tensor("op_242")]; + tensor cross_1 = mul(x = var_239, y = var_242)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_245 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_245")]; + tensor var_247_transpose_x_1 = const()[name = tensor("op_247_transpose_x_1"), val = tensor(true)]; + tensor var_247_transpose_y_1 = const()[name = tensor("op_247_transpose_y_1"), val = tensor(false)]; + tensor var_247 = matmul(transpose_x = var_247_transpose_x_1, transpose_y = var_247_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_247")]; + tensor new_kv_unnorm_1 = add(x = var_245, y = var_247)[name = tensor("new_kv_unnorm_1")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_249)[name = tensor("new_scale_1")]; + tensor var_251 = sqrt(x = new_scale_1)[name = tensor("op_251")]; + tensor var_252 = real_div(x = new_kv_unnorm_1, y = var_251)[name = tensor("op_252")]; + tensor var_253_perm_0 = const()[name = tensor("op_253_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; + tensor var_253 = transpose(perm = var_253_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_63, x = var_253)[name = tensor("out_3")]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_257, x = out_3)[name = tensor("out_5")]; + tensor var_259 = silu(x = input_19)[name = tensor("op_259")]; + tensor input_21 = mul(x = var_259, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; + tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; + tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; + tensor var_267_begin_0 = const()[name = tensor("op_267_begin_0"), val = tensor([0, 0, 0])]; + tensor var_267_end_0 = const()[name = tensor("op_267_end_0"), val = tensor([1, 1, 256])]; + tensor var_267_end_mask_0 = const()[name = tensor("op_267_end_mask_0"), val = tensor([true, false, true])]; + tensor var_267 = slice_by_index(begin = var_267_begin_0, end = var_267_end_0, end_mask = var_267_end_mask_0, x = x_3)[name = tensor("op_267")]; + tensor var_270_begin_0 = const()[name = tensor("op_270_begin_0"), val = tensor([0, 1, 0])]; + tensor var_270_end_0 = const()[name = tensor("op_270_end_0"), val = tensor([1, 16, 256])]; + tensor var_270_end_mask_0 = const()[name = tensor("op_270_end_mask_0"), val = tensor([true, true, true])]; + tensor var_270 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = window_1)[name = tensor("op_270")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_72, interleave = window_3_interleave_0, values = (var_270, var_267))[name = tensor("window_3")]; + tensor var_275_begin_0 = const()[name = tensor("op_275_begin_0"), val = tensor([0, 1, 0])]; + tensor var_275_end_0 = const()[name = tensor("op_275_end_0"), val = tensor([1, 2, 256])]; + tensor var_275_end_mask_0 = const()[name = tensor("op_275_end_mask_0"), val = tensor([true, false, true])]; + tensor var_275 = slice_by_index(begin = var_275_begin_0, end = var_275_end_0, end_mask = var_275_end_mask_0, x = x_3)[name = tensor("op_275")]; + tensor var_278_begin_0 = const()[name = tensor("op_278_begin_0"), val = tensor([0, 1, 0])]; + tensor var_278_end_0 = const()[name = tensor("op_278_end_0"), val = tensor([1, 16, 256])]; + tensor var_278_end_mask_0 = const()[name = tensor("op_278_end_mask_0"), val = tensor([true, true, true])]; + tensor var_278 = slice_by_index(begin = var_278_begin_0, end = var_278_end_0, end_mask = var_278_end_mask_0, x = window_3)[name = tensor("op_278")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_72, interleave = window_5_interleave_0, values = (var_278, var_275))[name = tensor("window_5")]; + tensor var_283_begin_0 = const()[name = tensor("op_283_begin_0"), val = tensor([0, 2, 0])]; + tensor var_283_end_0 = const()[name = tensor("op_283_end_0"), val = tensor([1, 1, 256])]; + tensor var_283_end_mask_0 = const()[name = tensor("op_283_end_mask_0"), val = tensor([true, true, true])]; + tensor var_283 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = x_3)[name = tensor("op_283")]; + tensor var_286_begin_0 = const()[name = tensor("op_286_begin_0"), val = tensor([0, 1, 0])]; + tensor var_286_end_0 = const()[name = tensor("op_286_end_0"), val = tensor([1, 16, 256])]; + tensor var_286_end_mask_0 = const()[name = tensor("op_286_end_mask_0"), val = tensor([true, true, true])]; + tensor var_286 = slice_by_index(begin = var_286_begin_0, end = var_286_end_0, end_mask = var_286_end_mask_0, x = window_5)[name = tensor("op_286")]; + tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; + tensor window_7 = concat(axis = var_72, interleave = window_7_interleave_0, values = (var_286, var_283))[name = tensor("window_7")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_58, interleave = input_23_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_23")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; + tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; + tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; + tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; + tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_311_split_sizes_0 = const()[name = tensor("op_311_split_sizes_0"), val = tensor([256, 256])]; + tensor var_311_axis_0 = const()[name = tensor("op_311_axis_0"), val = tensor(1)]; + tensor var_311_0, tensor var_311_1 = split(axis = var_311_axis_0, split_sizes = var_311_split_sizes_0, x = inputs_3)[name = tensor("op_311")]; + tensor var_313 = sigmoid(x = var_311_1)[name = tensor("op_313")]; + tensor inputs_5 = mul(x = var_311_0, y = var_313)[name = tensor("inputs_5")]; + tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; + tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; + tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([3, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; + tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, -1, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([3, 16, 256])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = conv_out_1)[name = tensor("op_344")]; + tensor var_346_perm_0 = const()[name = tensor("op_346_perm_0"), val = tensor([1, 0, 2])]; + tensor var_346 = transpose(perm = var_346_perm_0, x = var_344)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_346)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor(0x1p-1)]; + tensor var_370 = mul(x = input_41, y = var_369)[name = tensor("op_370")]; + tensor input_43 = add(x = var_370, y = input_33)[name = tensor("input_43")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor(0x1p-1)]; + tensor var_400 = mul(x = input_53, y = var_399)[name = tensor("op_400")]; + tensor input_55 = add(x = var_400, y = input_45)[name = tensor("input_55")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; + tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; + tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_3_squeeze_mask_0 = const()[name = tensor("prev_kv_3_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_3 = slice_by_index(begin = prev_kv_3_begin_0, end = prev_kv_3_end_0, end_mask = prev_kv_3_end_mask_0, squeeze_mask = prev_kv_3_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_3")]; + tensor prev_scale_3_begin_0 = const()[name = tensor("prev_scale_3_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_3_end_0 = const()[name = tensor("prev_scale_3_end_0"), val = tensor([2, 1])]; + tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; + tensor var_414 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 3, 4, 64])]; + tensor var_416 = reshape(shape = var_415, x = var_414)[name = tensor("op_416")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_420 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_421 = const()[name = tensor("op_421"), val = tensor(0x1p-3)]; + tensor var_422 = mul(x = var_420, y = var_421)[name = tensor("op_422")]; + tensor var_423 = const()[name = tensor("op_423"), val = tensor([1, 3, 4, 64])]; + tensor var_424 = reshape(shape = var_423, x = var_422)[name = tensor("op_424")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_428 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 3, 4, 64])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; + tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; + tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; + tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; + tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_424)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_416)[name = tensor("transpose_51")]; + tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor([3, 1])]; + tensor var_441 = reshape(shape = var_440, x = sqrt_s_t_3)[name = tensor("op_441")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_441)[name = tensor("M_3")]; + tensor var_443 = mul(x = qk_3, y = M_3)[name = tensor("op_443")]; + tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; + tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_430)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_443, y = v_3)[name = tensor("inner_3")]; + tensor var_445_transpose_x_0 = const()[name = tensor("op_445_transpose_x_0"), val = tensor(false)]; + tensor var_445_transpose_y_0 = const()[name = tensor("op_445_transpose_y_0"), val = tensor(false)]; + tensor var_445 = matmul(transpose_x = var_445_transpose_x_0, transpose_y = var_445_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_445")]; + tensor var_446 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_446")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 3, 1])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; + tensor cross_3 = mul(x = var_445, y = var_448)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_451 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_451")]; + tensor var_453_transpose_x_1 = const()[name = tensor("op_453_transpose_x_1"), val = tensor(true)]; + tensor var_453_transpose_y_1 = const()[name = tensor("op_453_transpose_y_1"), val = tensor(false)]; + tensor var_453 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_453")]; + tensor new_kv_unnorm_3 = add(x = var_451, y = var_453)[name = tensor("new_kv_unnorm_3")]; + tensor var_455 = const()[name = tensor("op_455"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_455)[name = tensor("new_scale_3")]; + tensor var_457 = sqrt(x = new_scale_3)[name = tensor("op_457")]; + tensor var_458 = real_div(x = new_kv_unnorm_3, y = var_457)[name = tensor("op_458")]; + tensor var_459_perm_0 = const()[name = tensor("op_459_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; + tensor var_459 = transpose(perm = var_459_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_63, x = var_459)[name = tensor("out_9")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_463, x = out_9)[name = tensor("out_11")]; + tensor var_465 = silu(x = input_59)[name = tensor("op_465")]; + tensor input_61 = mul(x = var_465, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; + tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; + tensor var_473_begin_0 = const()[name = tensor("op_473_begin_0"), val = tensor([0, 0, 0])]; + tensor var_473_end_0 = const()[name = tensor("op_473_end_0"), val = tensor([1, 1, 256])]; + tensor var_473_end_mask_0 = const()[name = tensor("op_473_end_mask_0"), val = tensor([true, false, true])]; + tensor var_473 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = x_9)[name = tensor("op_473")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 1, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 16, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = window_9)[name = tensor("op_476")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_72, interleave = window_11_interleave_0, values = (var_476, var_473))[name = tensor("window_11")]; + tensor var_481_begin_0 = const()[name = tensor("op_481_begin_0"), val = tensor([0, 1, 0])]; + tensor var_481_end_0 = const()[name = tensor("op_481_end_0"), val = tensor([1, 2, 256])]; + tensor var_481_end_mask_0 = const()[name = tensor("op_481_end_mask_0"), val = tensor([true, false, true])]; + tensor var_481 = slice_by_index(begin = var_481_begin_0, end = var_481_end_0, end_mask = var_481_end_mask_0, x = x_9)[name = tensor("op_481")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([0, 1, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([1, 16, 256])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([true, true, true])]; + tensor var_484 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, x = window_11)[name = tensor("op_484")]; + tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; + tensor window_13 = concat(axis = var_72, interleave = window_13_interleave_0, values = (var_484, var_481))[name = tensor("window_13")]; + tensor var_489_begin_0 = const()[name = tensor("op_489_begin_0"), val = tensor([0, 2, 0])]; + tensor var_489_end_0 = const()[name = tensor("op_489_end_0"), val = tensor([1, 1, 256])]; + tensor var_489_end_mask_0 = const()[name = tensor("op_489_end_mask_0"), val = tensor([true, true, true])]; + tensor var_489 = slice_by_index(begin = var_489_begin_0, end = var_489_end_0, end_mask = var_489_end_mask_0, x = x_9)[name = tensor("op_489")]; + tensor var_492_begin_0 = const()[name = tensor("op_492_begin_0"), val = tensor([0, 1, 0])]; + tensor var_492_end_0 = const()[name = tensor("op_492_end_0"), val = tensor([1, 16, 256])]; + tensor var_492_end_mask_0 = const()[name = tensor("op_492_end_mask_0"), val = tensor([true, true, true])]; + tensor var_492 = slice_by_index(begin = var_492_begin_0, end = var_492_end_0, end_mask = var_492_end_mask_0, x = window_13)[name = tensor("op_492")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_72, interleave = window_15_interleave_0, values = (var_492, var_489))[name = tensor("window_15")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_58, interleave = input_63_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_63")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; + tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; + tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; + tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; + tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_517_split_sizes_0 = const()[name = tensor("op_517_split_sizes_0"), val = tensor([256, 256])]; + tensor var_517_axis_0 = const()[name = tensor("op_517_axis_0"), val = tensor(1)]; + tensor var_517_0, tensor var_517_1 = split(axis = var_517_axis_0, split_sizes = var_517_split_sizes_0, x = inputs_13)[name = tensor("op_517")]; + tensor var_519 = sigmoid(x = var_517_1)[name = tensor("op_519")]; + tensor inputs_15 = mul(x = var_517_0, y = var_519)[name = tensor("inputs_15")]; + tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; + tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; + tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([3, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; + tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; + tensor var_550_begin_0 = const()[name = tensor("op_550_begin_0"), val = tensor([0, -1, 0])]; + tensor var_550_end_0 = const()[name = tensor("op_550_end_0"), val = tensor([3, 16, 256])]; + tensor var_550_end_mask_0 = const()[name = tensor("op_550_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_550 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = conv_out_3)[name = tensor("op_550")]; + tensor var_552_perm_0 = const()[name = tensor("op_552_perm_0"), val = tensor([1, 0, 2])]; + tensor var_552 = transpose(perm = var_552_perm_0, x = var_550)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_552)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_81, y = var_575)[name = tensor("op_576")]; + tensor input_83 = add(x = var_576, y = input_73)[name = tensor("input_83")]; + tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor(0x1p-1)]; + tensor var_606 = mul(x = input_93, y = var_605)[name = tensor("op_606")]; + tensor input_95 = add(x = var_606, y = input_85)[name = tensor("input_95")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; + tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; + tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; + tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_5_squeeze_mask_0 = const()[name = tensor("prev_kv_5_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_5 = slice_by_index(begin = prev_kv_5_begin_0, end = prev_kv_5_end_0, end_mask = prev_kv_5_end_mask_0, squeeze_mask = prev_kv_5_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_5")]; + tensor prev_scale_5_begin_0 = const()[name = tensor("prev_scale_5_begin_0"), val = tensor([2, 0])]; + tensor prev_scale_5_end_0 = const()[name = tensor("prev_scale_5_end_0"), val = tensor([3, 1])]; + tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; + tensor var_620 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 3, 4, 64])]; + tensor var_622 = reshape(shape = var_621, x = var_620)[name = tensor("op_622")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_626 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor(0x1p-3)]; + tensor var_628 = mul(x = var_626, y = var_627)[name = tensor("op_628")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 3, 4, 64])]; + tensor var_630 = reshape(shape = var_629, x = var_628)[name = tensor("op_630")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_634 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([1, 3, 4, 64])]; + tensor var_636 = reshape(shape = var_635, x = var_634)[name = tensor("op_636")]; + tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; + tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; + tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; + tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_630)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_622)[name = tensor("transpose_44")]; + tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; + tensor var_646 = const()[name = tensor("op_646"), val = tensor([3, 1])]; + tensor var_647 = reshape(shape = var_646, x = sqrt_s_t_5)[name = tensor("op_647")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_647)[name = tensor("M_5")]; + tensor var_649 = mul(x = qk_5, y = M_5)[name = tensor("op_649")]; + tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; + tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_636)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_649, y = v_5)[name = tensor("inner_5")]; + tensor var_651_transpose_x_0 = const()[name = tensor("op_651_transpose_x_0"), val = tensor(false)]; + tensor var_651_transpose_y_0 = const()[name = tensor("op_651_transpose_y_0"), val = tensor(false)]; + tensor var_651 = matmul(transpose_x = var_651_transpose_x_0, transpose_y = var_651_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_651")]; + tensor var_652 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_652")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1, 3, 1])]; + tensor var_654 = reshape(shape = var_653, x = var_652)[name = tensor("op_654")]; + tensor cross_5 = mul(x = var_651, y = var_654)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_657 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_657")]; + tensor var_659_transpose_x_1 = const()[name = tensor("op_659_transpose_x_1"), val = tensor(true)]; + tensor var_659_transpose_y_1 = const()[name = tensor("op_659_transpose_y_1"), val = tensor(false)]; + tensor var_659 = matmul(transpose_x = var_659_transpose_x_1, transpose_y = var_659_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_659")]; + tensor new_kv_unnorm_5 = add(x = var_657, y = var_659)[name = tensor("new_kv_unnorm_5")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_661)[name = tensor("new_scale_5")]; + tensor var_663 = sqrt(x = new_scale_5)[name = tensor("op_663")]; + tensor var_664 = real_div(x = new_kv_unnorm_5, y = var_663)[name = tensor("op_664")]; + tensor var_665_perm_0 = const()[name = tensor("op_665_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; + tensor var_665 = transpose(perm = var_665_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_63, x = var_665)[name = tensor("out_15")]; + tensor var_669 = const()[name = tensor("op_669"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_669, x = out_15)[name = tensor("out_17")]; + tensor var_671 = silu(x = input_99)[name = tensor("op_671")]; + tensor input_101 = mul(x = var_671, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; + tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; + tensor var_679_begin_0 = const()[name = tensor("op_679_begin_0"), val = tensor([0, 0, 0])]; + tensor var_679_end_0 = const()[name = tensor("op_679_end_0"), val = tensor([1, 1, 256])]; + tensor var_679_end_mask_0 = const()[name = tensor("op_679_end_mask_0"), val = tensor([true, false, true])]; + tensor var_679 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, x = x_15)[name = tensor("op_679")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 1, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 16, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, true, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = window_17)[name = tensor("op_682")]; + tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; + tensor window_19 = concat(axis = var_72, interleave = window_19_interleave_0, values = (var_682, var_679))[name = tensor("window_19")]; + tensor var_687_begin_0 = const()[name = tensor("op_687_begin_0"), val = tensor([0, 1, 0])]; + tensor var_687_end_0 = const()[name = tensor("op_687_end_0"), val = tensor([1, 2, 256])]; + tensor var_687_end_mask_0 = const()[name = tensor("op_687_end_mask_0"), val = tensor([true, false, true])]; + tensor var_687 = slice_by_index(begin = var_687_begin_0, end = var_687_end_0, end_mask = var_687_end_mask_0, x = x_15)[name = tensor("op_687")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 1, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 16, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, true, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = window_19)[name = tensor("op_690")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_72, interleave = window_21_interleave_0, values = (var_690, var_687))[name = tensor("window_21")]; + tensor var_695_begin_0 = const()[name = tensor("op_695_begin_0"), val = tensor([0, 2, 0])]; + tensor var_695_end_0 = const()[name = tensor("op_695_end_0"), val = tensor([1, 1, 256])]; + tensor var_695_end_mask_0 = const()[name = tensor("op_695_end_mask_0"), val = tensor([true, true, true])]; + tensor var_695 = slice_by_index(begin = var_695_begin_0, end = var_695_end_0, end_mask = var_695_end_mask_0, x = x_15)[name = tensor("op_695")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 1, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 16, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = window_21)[name = tensor("op_698")]; + tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; + tensor window_23 = concat(axis = var_72, interleave = window_23_interleave_0, values = (var_698, var_695))[name = tensor("window_23")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_58, interleave = input_103_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_103")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; + tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; + tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; + tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; + tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_723_split_sizes_0 = const()[name = tensor("op_723_split_sizes_0"), val = tensor([256, 256])]; + tensor var_723_axis_0 = const()[name = tensor("op_723_axis_0"), val = tensor(1)]; + tensor var_723_0, tensor var_723_1 = split(axis = var_723_axis_0, split_sizes = var_723_split_sizes_0, x = inputs_23)[name = tensor("op_723")]; + tensor var_725 = sigmoid(x = var_723_1)[name = tensor("op_725")]; + tensor inputs_25 = mul(x = var_723_0, y = var_725)[name = tensor("inputs_25")]; + tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; + tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; + tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([3, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; + tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, -1, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([3, 16, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = conv_out_5)[name = tensor("op_756")]; + tensor var_758_perm_0 = const()[name = tensor("op_758_perm_0"), val = tensor([1, 0, 2])]; + tensor var_758 = transpose(perm = var_758_perm_0, x = var_756)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_758)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor(0x1p-1)]; + tensor var_782 = mul(x = input_121, y = var_781)[name = tensor("op_782")]; + tensor input_123 = add(x = var_782, y = input_113)[name = tensor("input_123")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-1)]; + tensor var_812 = mul(x = input_133, y = var_811)[name = tensor("op_812")]; + tensor input_135 = add(x = var_812, y = input_125)[name = tensor("input_135")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; + tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; + tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; + tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_7_squeeze_mask_0 = const()[name = tensor("prev_kv_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_7 = slice_by_index(begin = prev_kv_7_begin_0, end = prev_kv_7_end_0, end_mask = prev_kv_7_end_mask_0, squeeze_mask = prev_kv_7_squeeze_mask_0, x = enc_kv)[name = tensor("prev_kv_7")]; + tensor prev_scale_7_begin_0 = const()[name = tensor("prev_scale_7_begin_0"), val = tensor([3, 0])]; + tensor prev_scale_7_end_0 = const()[name = tensor("prev_scale_7_end_0"), val = tensor([4, 1])]; + tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; + tensor var_826 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor([1, 3, 4, 64])]; + tensor var_828 = reshape(shape = var_827, x = var_826)[name = tensor("op_828")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_832 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p-3)]; + tensor var_834 = mul(x = var_832, y = var_833)[name = tensor("op_834")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 3, 4, 64])]; + tensor var_836 = reshape(shape = var_835, x = var_834)[name = tensor("op_836")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_840 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 3, 4, 64])]; + tensor var_842 = reshape(shape = var_841, x = var_840)[name = tensor("op_842")]; + tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; + tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; + tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; + tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_836)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_828)[name = tensor("transpose_37")]; + tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; + tensor var_852 = const()[name = tensor("op_852"), val = tensor([3, 1])]; + tensor var_853 = reshape(shape = var_852, x = sqrt_s_t_7)[name = tensor("op_853")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_853)[name = tensor("M_7")]; + tensor var_855 = mul(x = qk_7, y = M_7)[name = tensor("op_855")]; + tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; + tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_842)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_855, y = v_7)[name = tensor("inner_7")]; + tensor var_857_transpose_x_0 = const()[name = tensor("op_857_transpose_x_0"), val = tensor(false)]; + tensor var_857_transpose_y_0 = const()[name = tensor("op_857_transpose_y_0"), val = tensor(false)]; + tensor var_857 = matmul(transpose_x = var_857_transpose_x_0, transpose_y = var_857_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_857")]; + tensor var_858 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_858")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor([1, 1, 3, 1])]; + tensor var_860 = reshape(shape = var_859, x = var_858)[name = tensor("op_860")]; + tensor cross_7 = mul(x = var_857, y = var_860)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_863 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_863")]; + tensor var_865_transpose_x_1 = const()[name = tensor("op_865_transpose_x_1"), val = tensor(true)]; + tensor var_865_transpose_y_1 = const()[name = tensor("op_865_transpose_y_1"), val = tensor(false)]; + tensor var_865 = matmul(transpose_x = var_865_transpose_x_1, transpose_y = var_865_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_865")]; + tensor new_kv_unnorm_7 = add(x = var_863, y = var_865)[name = tensor("new_kv_unnorm_7")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_867)[name = tensor("new_scale_7")]; + tensor var_869 = sqrt(x = new_scale_7)[name = tensor("op_869")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_869)[name = tensor("nkv_1")]; + tensor var_871_perm_0 = const()[name = tensor("op_871_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; + tensor var_871 = transpose(perm = var_871_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_63, x = var_871)[name = tensor("out_21")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_875, x = out_21)[name = tensor("out_23")]; + tensor var_877 = silu(x = input_139)[name = tensor("op_877")]; + tensor input_141 = mul(x = var_877, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; + tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; + tensor var_885_begin_0 = const()[name = tensor("op_885_begin_0"), val = tensor([0, 0, 0])]; + tensor var_885_end_0 = const()[name = tensor("op_885_end_0"), val = tensor([1, 1, 256])]; + tensor var_885_end_mask_0 = const()[name = tensor("op_885_end_mask_0"), val = tensor([true, false, true])]; + tensor var_885 = slice_by_index(begin = var_885_begin_0, end = var_885_end_0, end_mask = var_885_end_mask_0, x = x_21)[name = tensor("op_885")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 1, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 16, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, true, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = window_25)[name = tensor("op_888")]; + tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; + tensor window_27 = concat(axis = var_72, interleave = window_27_interleave_0, values = (var_888, var_885))[name = tensor("window_27")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 1, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 2, 256])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([true, false, true])]; + tensor var_893 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = x_21)[name = tensor("op_893")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 16, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = window_27)[name = tensor("op_896")]; + tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; + tensor window_29 = concat(axis = var_72, interleave = window_29_interleave_0, values = (var_896, var_893))[name = tensor("window_29")]; + tensor var_901_begin_0 = const()[name = tensor("op_901_begin_0"), val = tensor([0, 2, 0])]; + tensor var_901_end_0 = const()[name = tensor("op_901_end_0"), val = tensor([1, 1, 256])]; + tensor var_901_end_mask_0 = const()[name = tensor("op_901_end_mask_0"), val = tensor([true, true, true])]; + tensor var_901 = slice_by_index(begin = var_901_begin_0, end = var_901_end_0, end_mask = var_901_end_mask_0, x = x_21)[name = tensor("op_901")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 1, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 16, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = window_29)[name = tensor("op_904")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_72, interleave = window_interleave_0, values = (var_904, var_901))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_58, interleave = input_143_interleave_0, values = (window_27, window_29, window))[name = tensor("input_143")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; + tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; + tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; + tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; + tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; + tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_929_split_sizes_0 = const()[name = tensor("op_929_split_sizes_0"), val = tensor([256, 256])]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(1)]; + tensor var_929_0, tensor var_929_1 = split(axis = var_929_axis_0, split_sizes = var_929_split_sizes_0, x = inputs_33)[name = tensor("op_929")]; + tensor var_931 = sigmoid(x = var_929_1)[name = tensor("op_931")]; + tensor inputs_35 = mul(x = var_929_0, y = var_931)[name = tensor("inputs_35")]; + tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; + tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; + tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; + tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; + tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([3, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; + tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, -1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([3, 16, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = conv_out_7)[name = tensor("op_962")]; + tensor var_964_perm_0 = const()[name = tensor("op_964_perm_0"), val = tensor([1, 0, 2])]; + tensor var_964 = transpose(perm = var_964_perm_0, x = var_962)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_964)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1p-1)]; + tensor var_988 = mul(x = input_161, y = var_987)[name = tensor("op_988")]; + tensor input_163 = add(x = var_988, y = input_153)[name = tensor("input_163")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; + tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; + tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; + tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; + tensor cat = concat(axis = var_60, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; + tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; + tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; + tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; + tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1006_begin_0 = const()[name = tensor("op_1006_begin_0"), val = tensor([0, 0, 3])]; + tensor var_1006_end_0 = const()[name = tensor("op_1006_end_0"), val = tensor([1, 256, 21])]; + tensor var_1006_end_mask_0 = const()[name = tensor("op_1006_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, x = cat)[name = tensor("op_1006")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1009 = reduce_l2_norm(axes = var_1008, keep_dims = var_54, x = input_165)[name = tensor("op_1009")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_68, beta = const_12, x = var_1009)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1013_axis_0 = const()[name = tensor("op_1013_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1013_axis_0, values = (var_252, var_458, var_664, nkv_1))[name = tensor("op_1013")]; + tensor var_1015_axis_0 = const()[name = tensor("op_1015_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1015_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1015")]; + tensor var_1017_axis_0 = const()[name = tensor("op_1017_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1017_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_1017")]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1085_axes_0 = const()[name = tensor("op_1085_axes_0"), val = tensor([2])]; + tensor var_1085 = expand_dims(axes = var_1085_axes_0, x = emb)[name = tensor("op_1085")]; + tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1085)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_61, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1093_perm_0 = const()[name = tensor("op_1093_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([6, 3, 256])]; + tensor var_1093 = transpose(perm = var_1093_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1097, x = var_1093)[name = tensor("x_29")]; + tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; + tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_9_squeeze_mask_0 = const()[name = tensor("prev_kv_9_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv_9 = slice_by_index(begin = prev_kv_9_begin_0, end = prev_kv_9_end_0, end_mask = prev_kv_9_end_mask_0, squeeze_mask = prev_kv_9_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv_9")]; + tensor prev_scale_9_begin_0 = const()[name = tensor("prev_scale_9_begin_0"), val = tensor([0, 0])]; + tensor prev_scale_9_end_0 = const()[name = tensor("prev_scale_9_end_0"), val = tensor([1, 1])]; + tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; + tensor var_1105 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([6, 3, 4, 64])]; + tensor var_1107 = reshape(shape = var_1106, x = var_1105)[name = tensor("op_1107")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1111 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1112 = const()[name = tensor("op_1112"), val = tensor(0x1p-3)]; + tensor var_1113 = mul(x = var_1111, y = var_1112)[name = tensor("op_1113")]; + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([6, 3, 4, 64])]; + tensor var_1115 = reshape(shape = var_1114, x = var_1113)[name = tensor("op_1115")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1119 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([6, 3, 4, 64])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; + tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; + tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; + tensor cumsum_mask_1 = cumsum(axis = var_58, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; + tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; + tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; + tensor clip_1 = clip(alpha = var_48, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; + tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; + tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1115)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1107)[name = tensor("transpose_27")]; + tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 3])]; + tensor var_1134 = reshape(shape = var_1133, x = valid_mask)[name = tensor("op_1134")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1134)[name = tensor("causal_with_valid_1")]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([3, 1])]; + tensor var_1137 = reshape(shape = var_1136, x = sqrt_s_t_9)[name = tensor("op_1137")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1137)[name = tensor("M_9")]; + tensor var_1139 = mul(x = qk_9, y = M_9)[name = tensor("op_1139")]; + tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; + tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1121)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1139, y = v_9)[name = tensor("inner_9")]; + tensor var_1141_transpose_x_0 = const()[name = tensor("op_1141_transpose_x_0"), val = tensor(false)]; + tensor var_1141_transpose_y_0 = const()[name = tensor("op_1141_transpose_y_0"), val = tensor(false)]; + tensor var_1141 = matmul(transpose_x = var_1141_transpose_x_0, transpose_y = var_1141_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1141")]; + tensor var_1142 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1142")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1, 3, 1])]; + tensor var_1144 = reshape(shape = var_1143, x = var_1142)[name = tensor("op_1144")]; + tensor cross_9 = mul(x = var_1141, y = var_1144)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1147 = const()[name = tensor("op_1147"), val = tensor([1, 1, 3, 1])]; + tensor var_1148 = reshape(shape = var_1147, x = valid_mask)[name = tensor("op_1148")]; + tensor v_masked_1 = mul(x = v_9, y = var_1148)[name = tensor("v_masked_1")]; + tensor var_1150 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1150")]; + tensor var_1152_transpose_x_1 = const()[name = tensor("op_1152_transpose_x_1"), val = tensor(true)]; + tensor var_1152_transpose_y_1 = const()[name = tensor("op_1152_transpose_y_1"), val = tensor(false)]; + tensor var_1152 = matmul(transpose_x = var_1152_transpose_x_1, transpose_y = var_1152_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1152")]; + tensor new_kv_unnorm_9 = add(x = var_1150, y = var_1152)[name = tensor("new_kv_unnorm_9")]; + tensor var_1154_keep_dims_0 = const()[name = tensor("op_1154_keep_dims_0"), val = tensor(false)]; + tensor var_1154 = reduce_sum(keep_dims = var_1154_keep_dims_0, x = valid_mask)[name = tensor("op_1154")]; + tensor var_1155 = const()[name = tensor("op_1155"), val = tensor([1])]; + tensor var_1156 = reshape(shape = var_1155, x = var_1154)[name = tensor("op_1156")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1156)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_48, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; + tensor var_1160 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1160")]; + tensor var_1161_perm_0 = const()[name = tensor("op_1161_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; + tensor var_1161 = transpose(perm = var_1161_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_63, x = var_1161)[name = tensor("out_27")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([6, 3, 256])]; + tensor out_29 = reshape(shape = var_1165, x = out_27)[name = tensor("out_29")]; + tensor var_1167 = silu(x = input_171)[name = tensor("op_1167")]; + tensor input_173 = mul(x = var_1167, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; + tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_55, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 6, 3, 256])]; + tensor var_1178 = reshape(shape = var_1177, x = xt_1)[name = tensor("op_1178")]; + tensor var_1179_perm_0 = const()[name = tensor("op_1179_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([3, 6, 256])]; + tensor var_1179 = transpose(perm = var_1179_perm_0, x = var_1178)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1182, x = var_1179)[name = tensor("query_1")]; + tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; + tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; + tensor var_1205 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 3, 3, 256])]; + tensor var_1207 = reshape(shape = concat_1, x = var_1205)[name = tensor("op_1207")]; + tensor var_1208_axes_0 = const()[name = tensor("op_1208_axes_0"), val = tensor([0])]; + tensor var_1208 = expand_dims(axes = var_1208_axes_0, x = var_1207)[name = tensor("op_1208")]; + tensor var_1209_perm_0 = const()[name = tensor("op_1209_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1210_axes_0 = const()[name = tensor("op_1210_axes_0"), val = tensor([-2])]; + tensor var_1209 = transpose(perm = var_1209_perm_0, x = var_1208)[name = tensor("transpose_21")]; + tensor var_1210 = squeeze(axes = var_1210_axes_0, x = var_1209)[name = tensor("op_1210")]; + tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 3, 256])]; + tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1210)[name = tensor("q_11")]; + tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 3, 256])]; + tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1210)[name = tensor("k_11")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 3, 256])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1210)[name = tensor("v_11")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([6, 12, 64])]; + tensor var_1219 = reshape(shape = var_1218, x = q_11)[name = tensor("op_1219")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([6, 12, 64])]; + tensor var_1226 = reshape(shape = var_1225, x = k_11)[name = tensor("op_1226")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([6, 12, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = v_11)[name = tensor("op_1233")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([3, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1219)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1236, x = q_13)[name = tensor("q_15")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([3, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1226)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1238, x = k_13)[name = tensor("k_15")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([3, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1233)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1240, x = v_13)[name = tensor("v_15")]; + tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; + tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; + tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; + tensor matmul_0_transpose_x_0 = const()[name = tensor("matmul_0_transpose_x_0"), val = tensor(false)]; + tensor matmul_0 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = mul_1, y = k_15)[name = tensor("matmul_0")]; + tensor softmax_0_axis_0 = const()[name = tensor("softmax_0_axis_0"), val = tensor(-1)]; + tensor softmax_0 = softmax(axis = softmax_0_axis_0, x = matmul_0)[name = tensor("softmax_0")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; + tensor var_1243 = const()[name = tensor("op_1243"), val = tensor([2, 0, 1, 3])]; + tensor var_1248 = const()[name = tensor("op_1248"), val = tensor([18, 256])]; + tensor var_1244 = transpose(perm = var_1243, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1248, x = var_1244)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([6, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1252, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_55, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; + tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_55, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 3, 6, 256])]; + tensor x_31 = reshape(shape = var_1272, x = xt_3)[name = tensor("x_31")]; + tensor var_1274_perm_0 = const()[name = tensor("op_1274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([6, 3, 256])]; + tensor var_1274 = transpose(perm = var_1274_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1278, x = var_1274)[name = tensor("x")]; + tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; + tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor prev_kv_squeeze_mask_0 = const()[name = tensor("prev_kv_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor prev_kv = slice_by_index(begin = prev_kv_begin_0, end = prev_kv_end_0, end_mask = prev_kv_end_mask_0, squeeze_mask = prev_kv_squeeze_mask_0, x = dec_kv)[name = tensor("prev_kv")]; + tensor prev_scale_begin_0 = const()[name = tensor("prev_scale_begin_0"), val = tensor([1, 0])]; + tensor prev_scale_end_0 = const()[name = tensor("prev_scale_end_0"), val = tensor([2, 1])]; + tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; + tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; + tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; + tensor var_1286 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([6, 3, 4, 64])]; + tensor var_1288 = reshape(shape = var_1287, x = var_1286)[name = tensor("op_1288")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1292 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1293 = const()[name = tensor("op_1293"), val = tensor(0x1p-3)]; + tensor var_1294 = mul(x = var_1292, y = var_1293)[name = tensor("op_1294")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([6, 3, 4, 64])]; + tensor var_1296 = reshape(shape = var_1295, x = var_1294)[name = tensor("op_1296")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1300 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([6, 3, 4, 64])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; + tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; + tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; + tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; + tensor clip_3 = clip(alpha = var_48, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; + tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; + tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1296)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1288)[name = tensor("transpose_14")]; + tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([3, 1])]; + tensor var_1318 = reshape(shape = var_1317, x = sqrt_s_t)[name = tensor("op_1318")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1318)[name = tensor("M")]; + tensor var_1320 = mul(x = qk, y = M)[name = tensor("op_1320")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1302)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1320, y = v_17)[name = tensor("inner_11")]; + tensor var_1322_transpose_x_0 = const()[name = tensor("op_1322_transpose_x_0"), val = tensor(false)]; + tensor var_1322_transpose_y_0 = const()[name = tensor("op_1322_transpose_y_0"), val = tensor(false)]; + tensor var_1322 = matmul(transpose_x = var_1322_transpose_x_0, transpose_y = var_1322_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1322")]; + tensor var_1323 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1323")]; + tensor var_1324 = const()[name = tensor("op_1324"), val = tensor([1, 1, 3, 1])]; + tensor var_1325 = reshape(shape = var_1324, x = var_1323)[name = tensor("op_1325")]; + tensor cross = mul(x = var_1322, y = var_1325)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1148)[name = tensor("v_masked")]; + tensor var_1331 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1331")]; + tensor var_1333_transpose_x_1 = const()[name = tensor("op_1333_transpose_x_1"), val = tensor(true)]; + tensor var_1333_transpose_y_1 = const()[name = tensor("op_1333_transpose_y_1"), val = tensor(false)]; + tensor var_1333 = matmul(transpose_x = var_1333_transpose_x_1, transpose_y = var_1333_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1333")]; + tensor new_kv_unnorm = add(x = var_1331, y = var_1333)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1156)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_48, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; + tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; + tensor var_1342_perm_0 = const()[name = tensor("op_1342_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; + tensor var_1342 = transpose(perm = var_1342_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_63, x = var_1342)[name = tensor("out_33")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([6, 3, 256])]; + tensor out = reshape(shape = var_1346, x = out_33)[name = tensor("out")]; + tensor var_1348 = silu(x = input_189)[name = tensor("op_1348")]; + tensor input_191 = mul(x = var_1348, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; + tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_55, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 6, 3, 256])]; + tensor var_1359 = reshape(shape = var_1358, x = xt_5)[name = tensor("op_1359")]; + tensor var_1360_perm_0 = const()[name = tensor("op_1360_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([3, 6, 256])]; + tensor var_1360 = transpose(perm = var_1360_perm_0, x = var_1359)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1363, x = var_1360)[name = tensor("query_5")]; + tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; + tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; + tensor var_1386 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 3, 3, 256])]; + tensor var_1388 = reshape(shape = concat_2, x = var_1386)[name = tensor("op_1388")]; + tensor var_1389_axes_0 = const()[name = tensor("op_1389_axes_0"), val = tensor([0])]; + tensor var_1389 = expand_dims(axes = var_1389_axes_0, x = var_1388)[name = tensor("op_1389")]; + tensor var_1390_perm_0 = const()[name = tensor("op_1390_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1391_axes_0 = const()[name = tensor("op_1391_axes_0"), val = tensor([-2])]; + tensor var_1390 = transpose(perm = var_1390_perm_0, x = var_1389)[name = tensor("transpose_8")]; + tensor var_1391 = squeeze(axes = var_1391_axes_0, x = var_1390)[name = tensor("op_1391")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 3, 256])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1391)[name = tensor("q_19")]; + tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 3, 256])]; + tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1391)[name = tensor("k_19")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 3, 256])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1391)[name = tensor("v_19")]; + tensor var_1399 = const()[name = tensor("op_1399"), val = tensor([6, 12, 64])]; + tensor var_1400 = reshape(shape = var_1399, x = q_19)[name = tensor("op_1400")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([6, 12, 64])]; + tensor var_1407 = reshape(shape = var_1406, x = k_19)[name = tensor("op_1407")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([6, 12, 64])]; + tensor var_1414 = reshape(shape = var_1413, x = v_19)[name = tensor("op_1414")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1417 = const()[name = tensor("op_1417"), val = tensor([3, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1400)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1417, x = q_21)[name = tensor("q")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([3, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1407)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1419, x = k_21)[name = tensor("k")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([3, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1414)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1421, x = v_21)[name = tensor("v")]; + tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; + tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; + tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; + tensor matmul_1_transpose_x_0 = const()[name = tensor("matmul_1_transpose_x_0"), val = tensor(false)]; + tensor matmul_1 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = mul_3, y = k)[name = tensor("matmul_1")]; + tensor softmax_1_axis_0 = const()[name = tensor("softmax_1_axis_0"), val = tensor(-1)]; + tensor softmax_1 = softmax(axis = softmax_1_axis_0, x = matmul_1)[name = tensor("softmax_1")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; + tensor var_1424 = const()[name = tensor("op_1424"), val = tensor([2, 0, 1, 3])]; + tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([18, 256])]; + tensor var_1425 = transpose(perm = var_1424, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1429, x = var_1425)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([6, 3, 256])]; + tensor attn_output = reshape(shape = var_1433, x = attn_output_13)[name = tensor("attn_output")]; + tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; + tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_55, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; + tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_55, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1453 = const()[name = tensor("op_1453"), val = tensor([1, 3, 6, 256])]; + tensor input = reshape(shape = var_1453, x = xt)[name = tensor("input")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([-1])]; + tensor var_1456 = reduce_l2_norm(axes = var_1455, keep_dims = var_54, x = input)[name = tensor("op_1456")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_68, beta = const_42, x = var_1456)[name = tensor("clip_5")]; + tensor var_1458 = real_div(x = input, y = clip_5)[name = tensor("op_1458")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; + tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; + tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; + tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1458)[name = tensor("transpose_2")]; + tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; + tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; + tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; + tensor matmul_2 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = reshape_0, y = reshape_1)[name = tensor("matmul_2")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([1, 3, 6])]; + tensor reshape_2 = reshape(shape = concat_10, x = matmul_2)[name = tensor("reshape_2")]; + tensor output_begin_0 = const()[name = tensor("output_begin_0"), val = tensor([0, 0, 1])]; + tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 5])]; + tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; + tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1462")]; + tensor var_1464_axis_0 = const()[name = tensor("op_1464_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1464_axis_0, values = (var_1160, nkv))[name = tensor("op_1464")]; + tensor var_1466_axis_0 = const()[name = tensor("op_1466_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1466_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1466")]; + } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); +} \ No newline at end of file