diff --git "a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil" "b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil" new file mode 100644--- /dev/null +++ "b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil" @@ -0,0 +1,1253 @@ +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], [0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+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], [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, 1, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, true, 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 stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29))[name = tensor("stacked")]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor([1, 2, 345])]; + tensor input_1 = reshape(shape = var_36, x = stacked)[name = tensor("input_1")]; + tensor var_38 = const()[name = tensor("op_38"), val = tensor(0x1p+0)]; + tensor var_44 = const()[name = tensor("op_44"), val = tensor(true)]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor(0x1.4f8b58p-17)]; + tensor var_48 = const()[name = tensor("op_48"), val = tensor(0)]; + tensor var_50 = const()[name = tensor("op_50"), val = tensor(2)]; + tensor var_51 = const()[name = tensor("op_51"), val = tensor(-1)]; + tensor var_53 = const()[name = tensor("op_53"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_58 = const()[name = tensor("op_58"), val = tensor(0x1.5798eep-27)]; + tensor var_62 = const()[name = tensor("op_62"), 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_45, 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_45, 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_183 = const()[name = tensor("op_183"), val = tensor(0x1p-1)]; + tensor var_184 = mul(x = input_13, y = var_183)[name = tensor("op_184")]; + tensor input_15 = add(x = var_184, 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_45, 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_198 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_199 = const()[name = tensor("op_199"), val = tensor([1, 2, 4, 64])]; + tensor var_200 = reshape(shape = var_199, x = var_198)[name = tensor("op_200")]; + tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_204 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_205 = const()[name = tensor("op_205"), val = tensor(0x1p-3)]; + tensor var_206 = mul(x = var_204, y = var_205)[name = tensor("op_206")]; + tensor var_207 = const()[name = tensor("op_207"), val = tensor([1, 2, 4, 64])]; + tensor var_208 = reshape(shape = var_207, x = var_206)[name = tensor("op_208")]; + tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_212 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([1, 2, 4, 64])]; + tensor var_214 = reshape(shape = var_213, x = var_212)[name = tensor("op_214")]; + 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_208)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_200)[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_224 = const()[name = tensor("op_224"), val = tensor([2, 1])]; + tensor var_225 = reshape(shape = var_224, x = sqrt_s_t_1)[name = tensor("op_225")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_225)[name = tensor("M_1")]; + tensor var_227 = mul(x = qk_1, y = M_1)[name = tensor("op_227")]; + 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_214)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_227, y = v_1)[name = tensor("inner_1")]; + tensor var_229_transpose_x_0 = const()[name = tensor("op_229_transpose_x_0"), val = tensor(false)]; + tensor var_229_transpose_y_0 = const()[name = tensor("op_229_transpose_y_0"), val = tensor(false)]; + tensor var_229 = matmul(transpose_x = var_229_transpose_x_0, transpose_y = var_229_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_229")]; + tensor var_230 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_230")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1, 2, 1])]; + tensor var_232 = reshape(shape = var_231, x = var_230)[name = tensor("op_232")]; + tensor cross_1 = mul(x = var_229, y = var_232)[name = tensor("cross_1")]; + tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; + tensor var_235 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_235")]; + tensor var_237_transpose_x_1 = const()[name = tensor("op_237_transpose_x_1"), val = tensor(true)]; + tensor var_237_transpose_y_1 = const()[name = tensor("op_237_transpose_y_1"), val = tensor(false)]; + tensor var_237 = matmul(transpose_x = var_237_transpose_x_1, transpose_y = var_237_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_237")]; + tensor new_kv_unnorm_1 = add(x = var_235, y = var_237)[name = tensor("new_kv_unnorm_1")]; + tensor var_239 = const()[name = tensor("op_239"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_239)[name = tensor("new_scale_1")]; + tensor var_241 = sqrt(x = new_scale_1)[name = tensor("op_241")]; + tensor var_242 = real_div(x = new_kv_unnorm_1, y = var_241)[name = tensor("op_242")]; + tensor var_243_perm_0 = const()[name = tensor("op_243_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_243 = transpose(perm = var_243_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_53, x = var_243)[name = tensor("out_3")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_247, x = out_3)[name = tensor("out_5")]; + tensor var_249 = silu(x = input_19)[name = tensor("op_249")]; + tensor input_21 = mul(x = var_249, 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_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 0, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 1, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, false, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = x_3)[name = tensor("op_257")]; + tensor var_260_begin_0 = const()[name = tensor("op_260_begin_0"), val = tensor([0, 1, 0])]; + tensor var_260_end_0 = const()[name = tensor("op_260_end_0"), val = tensor([1, 16, 256])]; + tensor var_260_end_mask_0 = const()[name = tensor("op_260_end_mask_0"), val = tensor([true, true, true])]; + tensor var_260 = slice_by_index(begin = var_260_begin_0, end = var_260_end_0, end_mask = var_260_end_mask_0, x = window_1)[name = tensor("op_260")]; + tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; + tensor window_3 = concat(axis = var_62, interleave = window_3_interleave_0, values = (var_260, var_257))[name = tensor("window_3")]; + tensor var_265_begin_0 = const()[name = tensor("op_265_begin_0"), val = tensor([0, 1, 0])]; + tensor var_265_end_0 = const()[name = tensor("op_265_end_0"), val = tensor([1, 1, 256])]; + tensor var_265_end_mask_0 = const()[name = tensor("op_265_end_mask_0"), val = tensor([true, true, true])]; + tensor var_265 = slice_by_index(begin = var_265_begin_0, end = var_265_end_0, end_mask = var_265_end_mask_0, x = x_3)[name = tensor("op_265")]; + tensor var_268_begin_0 = const()[name = tensor("op_268_begin_0"), val = tensor([0, 1, 0])]; + tensor var_268_end_0 = const()[name = tensor("op_268_end_0"), val = tensor([1, 16, 256])]; + tensor var_268_end_mask_0 = const()[name = tensor("op_268_end_mask_0"), val = tensor([true, true, true])]; + tensor var_268 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = window_3)[name = tensor("op_268")]; + tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; + tensor window_5 = concat(axis = var_62, interleave = window_5_interleave_0, values = (var_268, var_265))[name = tensor("window_5")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_48, interleave = input_23_interleave_0, values = (window_3, window_5))[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_45, 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_293_split_sizes_0 = const()[name = tensor("op_293_split_sizes_0"), val = tensor([256, 256])]; + tensor var_293_axis_0 = const()[name = tensor("op_293_axis_0"), val = tensor(1)]; + tensor var_293_0, tensor var_293_1 = split(axis = var_293_axis_0, split_sizes = var_293_split_sizes_0, x = inputs_3)[name = tensor("op_293")]; + tensor var_295 = sigmoid(x = var_293_1)[name = tensor("op_295")]; + tensor inputs_5 = mul(x = var_293_0, y = var_295)[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([2, 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_45, 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_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, -1, 0])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([2, 16, 256])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_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_326 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = conv_out_1)[name = tensor("op_326")]; + tensor var_328_perm_0 = const()[name = tensor("op_328_perm_0"), val = tensor([1, 0, 2])]; + tensor var_328 = transpose(perm = var_328_perm_0, x = var_326)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_328)[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_45, 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_351 = const()[name = tensor("op_351"), val = tensor(0x1p-1)]; + tensor var_352 = mul(x = input_41, y = var_351)[name = tensor("op_352")]; + tensor input_43 = add(x = var_352, 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_45, 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_45, 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_381 = const()[name = tensor("op_381"), val = tensor(0x1p-1)]; + tensor var_382 = mul(x = input_53, y = var_381)[name = tensor("op_382")]; + tensor input_55 = add(x = var_382, 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_45, 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_396 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([1, 2, 4, 64])]; + tensor var_398 = reshape(shape = var_397, x = var_396)[name = tensor("op_398")]; + tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_402 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_403 = const()[name = tensor("op_403"), val = tensor(0x1p-3)]; + tensor var_404 = mul(x = var_402, y = var_403)[name = tensor("op_404")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor([1, 2, 4, 64])]; + tensor var_406 = reshape(shape = var_405, x = var_404)[name = tensor("op_406")]; + tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_410 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 2, 4, 64])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; + 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_406)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_398)[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_422 = const()[name = tensor("op_422"), val = tensor([2, 1])]; + tensor var_423 = reshape(shape = var_422, x = sqrt_s_t_3)[name = tensor("op_423")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_423)[name = tensor("M_3")]; + tensor var_425 = mul(x = qk_3, y = M_3)[name = tensor("op_425")]; + 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_412)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_425, y = v_3)[name = tensor("inner_3")]; + tensor var_427_transpose_x_0 = const()[name = tensor("op_427_transpose_x_0"), val = tensor(false)]; + tensor var_427_transpose_y_0 = const()[name = tensor("op_427_transpose_y_0"), val = tensor(false)]; + tensor var_427 = matmul(transpose_x = var_427_transpose_x_0, transpose_y = var_427_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_427")]; + tensor var_428 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_428")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1, 2, 1])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; + tensor cross_3 = mul(x = var_427, y = var_430)[name = tensor("cross_3")]; + tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; + tensor var_433 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_433")]; + tensor var_435_transpose_x_1 = const()[name = tensor("op_435_transpose_x_1"), val = tensor(true)]; + tensor var_435_transpose_y_1 = const()[name = tensor("op_435_transpose_y_1"), val = tensor(false)]; + tensor var_435 = matmul(transpose_x = var_435_transpose_x_1, transpose_y = var_435_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_435")]; + tensor new_kv_unnorm_3 = add(x = var_433, y = var_435)[name = tensor("new_kv_unnorm_3")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_437)[name = tensor("new_scale_3")]; + tensor var_439 = sqrt(x = new_scale_3)[name = tensor("op_439")]; + tensor var_440 = real_div(x = new_kv_unnorm_3, y = var_439)[name = tensor("op_440")]; + tensor var_441_perm_0 = const()[name = tensor("op_441_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_441 = transpose(perm = var_441_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_53, x = var_441)[name = tensor("out_9")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_445, x = out_9)[name = tensor("out_11")]; + tensor var_447 = silu(x = input_59)[name = tensor("op_447")]; + tensor input_61 = mul(x = var_447, 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_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; + tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; + tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 0, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 1, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, false, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = x_9)[name = tensor("op_455")]; + tensor var_458_begin_0 = const()[name = tensor("op_458_begin_0"), val = tensor([0, 1, 0])]; + tensor var_458_end_0 = const()[name = tensor("op_458_end_0"), val = tensor([1, 16, 256])]; + tensor var_458_end_mask_0 = const()[name = tensor("op_458_end_mask_0"), val = tensor([true, true, true])]; + tensor var_458 = slice_by_index(begin = var_458_begin_0, end = var_458_end_0, end_mask = var_458_end_mask_0, x = window_7)[name = tensor("op_458")]; + tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; + tensor window_9 = concat(axis = var_62, interleave = window_9_interleave_0, values = (var_458, var_455))[name = tensor("window_9")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 1, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = x_9)[name = tensor("op_463")]; + tensor var_466_begin_0 = const()[name = tensor("op_466_begin_0"), val = tensor([0, 1, 0])]; + tensor var_466_end_0 = const()[name = tensor("op_466_end_0"), val = tensor([1, 16, 256])]; + tensor var_466_end_mask_0 = const()[name = tensor("op_466_end_mask_0"), val = tensor([true, true, true])]; + tensor var_466 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, x = window_9)[name = tensor("op_466")]; + tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; + tensor window_11 = concat(axis = var_62, interleave = window_11_interleave_0, values = (var_466, var_463))[name = tensor("window_11")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_48, interleave = input_63_interleave_0, values = (window_9, window_11))[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_45, 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_491_split_sizes_0 = const()[name = tensor("op_491_split_sizes_0"), val = tensor([256, 256])]; + tensor var_491_axis_0 = const()[name = tensor("op_491_axis_0"), val = tensor(1)]; + tensor var_491_0, tensor var_491_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = inputs_13)[name = tensor("op_491")]; + tensor var_493 = sigmoid(x = var_491_1)[name = tensor("op_493")]; + tensor inputs_15 = mul(x = var_491_0, y = var_493)[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([2, 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_45, 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_524_begin_0 = const()[name = tensor("op_524_begin_0"), val = tensor([0, -1, 0])]; + tensor var_524_end_0 = const()[name = tensor("op_524_end_0"), val = tensor([2, 16, 256])]; + tensor var_524_end_mask_0 = const()[name = tensor("op_524_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_524 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = conv_out_3)[name = tensor("op_524")]; + tensor var_526_perm_0 = const()[name = tensor("op_526_perm_0"), val = tensor([1, 0, 2])]; + tensor var_526 = transpose(perm = var_526_perm_0, x = var_524)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_526)[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_45, 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_549 = const()[name = tensor("op_549"), val = tensor(0x1p-1)]; + tensor var_550 = mul(x = input_81, y = var_549)[name = tensor("op_550")]; + tensor input_83 = add(x = var_550, 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_45, 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_45, 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_579 = const()[name = tensor("op_579"), val = tensor(0x1p-1)]; + tensor var_580 = mul(x = input_93, y = var_579)[name = tensor("op_580")]; + tensor input_95 = add(x = var_580, 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_45, 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_594 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 2, 4, 64])]; + tensor var_596 = reshape(shape = var_595, x = var_594)[name = tensor("op_596")]; + tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_600 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-3)]; + tensor var_602 = mul(x = var_600, y = var_601)[name = tensor("op_602")]; + tensor var_603 = const()[name = tensor("op_603"), val = tensor([1, 2, 4, 64])]; + tensor var_604 = reshape(shape = var_603, x = var_602)[name = tensor("op_604")]; + tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_608 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 2, 4, 64])]; + tensor var_610 = reshape(shape = var_609, x = var_608)[name = tensor("op_610")]; + 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_604)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_596)[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_620 = const()[name = tensor("op_620"), val = tensor([2, 1])]; + tensor var_621 = reshape(shape = var_620, x = sqrt_s_t_5)[name = tensor("op_621")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_621)[name = tensor("M_5")]; + tensor var_623 = mul(x = qk_5, y = M_5)[name = tensor("op_623")]; + 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_610)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_623, y = v_5)[name = tensor("inner_5")]; + tensor var_625_transpose_x_0 = const()[name = tensor("op_625_transpose_x_0"), val = tensor(false)]; + tensor var_625_transpose_y_0 = const()[name = tensor("op_625_transpose_y_0"), val = tensor(false)]; + tensor var_625 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_625")]; + tensor var_626 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_626")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 2, 1])]; + tensor var_628 = reshape(shape = var_627, x = var_626)[name = tensor("op_628")]; + tensor cross_5 = mul(x = var_625, y = var_628)[name = tensor("cross_5")]; + tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; + tensor var_631 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_631")]; + tensor var_633_transpose_x_1 = const()[name = tensor("op_633_transpose_x_1"), val = tensor(true)]; + tensor var_633_transpose_y_1 = const()[name = tensor("op_633_transpose_y_1"), val = tensor(false)]; + tensor var_633 = matmul(transpose_x = var_633_transpose_x_1, transpose_y = var_633_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_633")]; + tensor new_kv_unnorm_5 = add(x = var_631, y = var_633)[name = tensor("new_kv_unnorm_5")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_635)[name = tensor("new_scale_5")]; + tensor var_637 = sqrt(x = new_scale_5)[name = tensor("op_637")]; + tensor var_638 = real_div(x = new_kv_unnorm_5, y = var_637)[name = tensor("op_638")]; + tensor var_639_perm_0 = const()[name = tensor("op_639_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_639 = transpose(perm = var_639_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_53, x = var_639)[name = tensor("out_15")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_643, x = out_15)[name = tensor("out_17")]; + tensor var_645 = silu(x = input_99)[name = tensor("op_645")]; + tensor input_101 = mul(x = var_645, 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_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; + tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; + tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; + tensor var_653_begin_0 = const()[name = tensor("op_653_begin_0"), val = tensor([0, 0, 0])]; + tensor var_653_end_0 = const()[name = tensor("op_653_end_0"), val = tensor([1, 1, 256])]; + tensor var_653_end_mask_0 = const()[name = tensor("op_653_end_mask_0"), val = tensor([true, false, true])]; + tensor var_653 = slice_by_index(begin = var_653_begin_0, end = var_653_end_0, end_mask = var_653_end_mask_0, x = x_15)[name = tensor("op_653")]; + tensor var_656_begin_0 = const()[name = tensor("op_656_begin_0"), val = tensor([0, 1, 0])]; + tensor var_656_end_0 = const()[name = tensor("op_656_end_0"), val = tensor([1, 16, 256])]; + tensor var_656_end_mask_0 = const()[name = tensor("op_656_end_mask_0"), val = tensor([true, true, true])]; + tensor var_656 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, x = window_13)[name = tensor("op_656")]; + tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; + tensor window_15 = concat(axis = var_62, interleave = window_15_interleave_0, values = (var_656, var_653))[name = tensor("window_15")]; + tensor var_661_begin_0 = const()[name = tensor("op_661_begin_0"), val = tensor([0, 1, 0])]; + tensor var_661_end_0 = const()[name = tensor("op_661_end_0"), val = tensor([1, 1, 256])]; + tensor var_661_end_mask_0 = const()[name = tensor("op_661_end_mask_0"), val = tensor([true, true, true])]; + tensor var_661 = slice_by_index(begin = var_661_begin_0, end = var_661_end_0, end_mask = var_661_end_mask_0, x = x_15)[name = tensor("op_661")]; + tensor var_664_begin_0 = const()[name = tensor("op_664_begin_0"), val = tensor([0, 1, 0])]; + tensor var_664_end_0 = const()[name = tensor("op_664_end_0"), val = tensor([1, 16, 256])]; + tensor var_664_end_mask_0 = const()[name = tensor("op_664_end_mask_0"), val = tensor([true, true, true])]; + tensor var_664 = slice_by_index(begin = var_664_begin_0, end = var_664_end_0, end_mask = var_664_end_mask_0, x = window_15)[name = tensor("op_664")]; + tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; + tensor window_17 = concat(axis = var_62, interleave = window_17_interleave_0, values = (var_664, var_661))[name = tensor("window_17")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_48, interleave = input_103_interleave_0, values = (window_15, window_17))[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_45, 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_689_split_sizes_0 = const()[name = tensor("op_689_split_sizes_0"), val = tensor([256, 256])]; + tensor var_689_axis_0 = const()[name = tensor("op_689_axis_0"), val = tensor(1)]; + tensor var_689_0, tensor var_689_1 = split(axis = var_689_axis_0, split_sizes = var_689_split_sizes_0, x = inputs_23)[name = tensor("op_689")]; + tensor var_691 = sigmoid(x = var_689_1)[name = tensor("op_691")]; + tensor inputs_25 = mul(x = var_689_0, y = var_691)[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([2, 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_45, 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_722_begin_0 = const()[name = tensor("op_722_begin_0"), val = tensor([0, -1, 0])]; + tensor var_722_end_0 = const()[name = tensor("op_722_end_0"), val = tensor([2, 16, 256])]; + tensor var_722_end_mask_0 = const()[name = tensor("op_722_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_722 = slice_by_index(begin = var_722_begin_0, end = var_722_end_0, end_mask = var_722_end_mask_0, x = conv_out_5)[name = tensor("op_722")]; + tensor var_724_perm_0 = const()[name = tensor("op_724_perm_0"), val = tensor([1, 0, 2])]; + tensor var_724 = transpose(perm = var_724_perm_0, x = var_722)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_724)[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_45, 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_747 = const()[name = tensor("op_747"), val = tensor(0x1p-1)]; + tensor var_748 = mul(x = input_121, y = var_747)[name = tensor("op_748")]; + tensor input_123 = add(x = var_748, 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_45, 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_45, 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_777 = const()[name = tensor("op_777"), val = tensor(0x1p-1)]; + tensor var_778 = mul(x = input_133, y = var_777)[name = tensor("op_778")]; + tensor input_135 = add(x = var_778, 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_45, 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_792 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 2, 4, 64])]; + tensor var_794 = reshape(shape = var_793, x = var_792)[name = tensor("op_794")]; + tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_798 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p-3)]; + tensor var_800 = mul(x = var_798, y = var_799)[name = tensor("op_800")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 2, 4, 64])]; + tensor var_802 = reshape(shape = var_801, x = var_800)[name = tensor("op_802")]; + tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_806 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 2, 4, 64])]; + tensor var_808 = reshape(shape = var_807, x = var_806)[name = tensor("op_808")]; + 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_802)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_794)[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_818 = const()[name = tensor("op_818"), val = tensor([2, 1])]; + tensor var_819 = reshape(shape = var_818, x = sqrt_s_t_7)[name = tensor("op_819")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_819)[name = tensor("M_7")]; + tensor var_821 = mul(x = qk_7, y = M_7)[name = tensor("op_821")]; + 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_808)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_821, y = v_7)[name = tensor("inner_7")]; + tensor var_823_transpose_x_0 = const()[name = tensor("op_823_transpose_x_0"), val = tensor(false)]; + tensor var_823_transpose_y_0 = const()[name = tensor("op_823_transpose_y_0"), val = tensor(false)]; + tensor var_823 = matmul(transpose_x = var_823_transpose_x_0, transpose_y = var_823_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_823")]; + tensor var_824 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_824")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1, 2, 1])]; + tensor var_826 = reshape(shape = var_825, x = var_824)[name = tensor("op_826")]; + tensor cross_7 = mul(x = var_823, y = var_826)[name = tensor("cross_7")]; + tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; + tensor var_829 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_829")]; + tensor var_831_transpose_x_1 = const()[name = tensor("op_831_transpose_x_1"), val = tensor(true)]; + tensor var_831_transpose_y_1 = const()[name = tensor("op_831_transpose_y_1"), val = tensor(false)]; + tensor var_831 = matmul(transpose_x = var_831_transpose_x_1, transpose_y = var_831_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_831")]; + tensor new_kv_unnorm_7 = add(x = var_829, y = var_831)[name = tensor("new_kv_unnorm_7")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_833)[name = tensor("new_scale_7")]; + tensor var_835 = sqrt(x = new_scale_7)[name = tensor("op_835")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_835)[name = tensor("nkv_1")]; + tensor var_837_perm_0 = const()[name = tensor("op_837_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_837 = transpose(perm = var_837_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_53, x = var_837)[name = tensor("out_21")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_841, x = out_21)[name = tensor("out_23")]; + tensor var_843 = silu(x = input_139)[name = tensor("op_843")]; + tensor input_141 = mul(x = var_843, 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_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; + tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; + tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; + tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; + tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 256])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([true, false, true])]; + tensor var_851 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = x_21)[name = tensor("op_851")]; + tensor var_854_begin_0 = const()[name = tensor("op_854_begin_0"), val = tensor([0, 1, 0])]; + tensor var_854_end_0 = const()[name = tensor("op_854_end_0"), val = tensor([1, 16, 256])]; + tensor var_854_end_mask_0 = const()[name = tensor("op_854_end_mask_0"), val = tensor([true, true, true])]; + tensor var_854 = slice_by_index(begin = var_854_begin_0, end = var_854_end_0, end_mask = var_854_end_mask_0, x = window_19)[name = tensor("op_854")]; + tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; + tensor window_21 = concat(axis = var_62, interleave = window_21_interleave_0, values = (var_854, var_851))[name = tensor("window_21")]; + tensor var_859_begin_0 = const()[name = tensor("op_859_begin_0"), val = tensor([0, 1, 0])]; + tensor var_859_end_0 = const()[name = tensor("op_859_end_0"), val = tensor([1, 1, 256])]; + tensor var_859_end_mask_0 = const()[name = tensor("op_859_end_mask_0"), val = tensor([true, true, true])]; + tensor var_859 = slice_by_index(begin = var_859_begin_0, end = var_859_end_0, end_mask = var_859_end_mask_0, x = x_21)[name = tensor("op_859")]; + tensor var_862_begin_0 = const()[name = tensor("op_862_begin_0"), val = tensor([0, 1, 0])]; + tensor var_862_end_0 = const()[name = tensor("op_862_end_0"), val = tensor([1, 16, 256])]; + tensor var_862_end_mask_0 = const()[name = tensor("op_862_end_mask_0"), val = tensor([true, true, true])]; + tensor var_862 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, x = window_21)[name = tensor("op_862")]; + tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; + tensor window = concat(axis = var_62, interleave = window_interleave_0, values = (var_862, var_859))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_48, interleave = input_143_interleave_0, values = (window_21, 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_45, 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_887_split_sizes_0 = const()[name = tensor("op_887_split_sizes_0"), val = tensor([256, 256])]; + tensor var_887_axis_0 = const()[name = tensor("op_887_axis_0"), val = tensor(1)]; + tensor var_887_0, tensor var_887_1 = split(axis = var_887_axis_0, split_sizes = var_887_split_sizes_0, x = inputs_33)[name = tensor("op_887")]; + tensor var_889 = sigmoid(x = var_887_1)[name = tensor("op_889")]; + tensor inputs_35 = mul(x = var_887_0, y = var_889)[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([2, 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_45, 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_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, -1, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([2, 16, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_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_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = conv_out_7)[name = tensor("op_920")]; + tensor var_922_perm_0 = const()[name = tensor("op_922_perm_0"), val = tensor([1, 0, 2])]; + tensor var_922 = transpose(perm = var_922_perm_0, x = var_920)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_922)[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_45, 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_945 = const()[name = tensor("op_945"), val = tensor(0x1p-1)]; + tensor var_946 = mul(x = input_161, y = var_945)[name = tensor("op_946")]; + tensor input_163 = add(x = var_946, 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_45, 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_50, 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_964_begin_0 = const()[name = tensor("op_964_begin_0"), val = tensor([0, 0, 2])]; + tensor var_964_end_0 = const()[name = tensor("op_964_end_0"), val = tensor([1, 256, 20])]; + tensor var_964_end_mask_0 = const()[name = tensor("op_964_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = cat)[name = tensor("op_964")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_967 = reduce_l2_norm(axes = var_966, keep_dims = var_44, x = input_165)[name = tensor("op_967")]; + tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; + tensor clip_0 = clip(alpha = var_58, beta = const_12, x = var_967)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_971_axis_0, values = (var_242, var_440, var_638, nkv_1))[name = tensor("op_971")]; + tensor var_973_axis_0 = const()[name = tensor("op_973_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_973_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_973")]; + tensor var_975_axis_0 = const()[name = tensor("op_975_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_975_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_975")]; + tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1043_axes_0 = const()[name = tensor("op_1043_axes_0"), val = tensor([2])]; + tensor var_1043 = expand_dims(axes = var_1043_axes_0, x = emb)[name = tensor("op_1043")]; + 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_1043)[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_51, 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_1051_perm_0 = const()[name = tensor("op_1051_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor([6, 2, 256])]; + tensor var_1051 = transpose(perm = var_1051_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1055, x = var_1051)[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_1063 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1064 = const()[name = tensor("op_1064"), val = tensor([6, 2, 4, 64])]; + tensor var_1065 = reshape(shape = var_1064, x = var_1063)[name = tensor("op_1065")]; + tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1069 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor(0x1p-3)]; + tensor var_1071 = mul(x = var_1069, y = var_1070)[name = tensor("op_1071")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([6, 2, 4, 64])]; + tensor var_1073 = reshape(shape = var_1072, x = var_1071)[name = tensor("op_1073")]; + tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1077 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([6, 2, 4, 64])]; + tensor var_1079 = reshape(shape = var_1078, x = var_1077)[name = tensor("op_1079")]; + 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_48, 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_38, 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_1073)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1065)[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_1091 = const()[name = tensor("op_1091"), val = tensor([1, 2])]; + tensor var_1092 = reshape(shape = var_1091, x = valid_mask)[name = tensor("op_1092")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1092)[name = tensor("causal_with_valid_1")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([2, 1])]; + tensor var_1095 = reshape(shape = var_1094, x = sqrt_s_t_9)[name = tensor("op_1095")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1095)[name = tensor("M_9")]; + tensor var_1097 = mul(x = qk_9, y = M_9)[name = tensor("op_1097")]; + 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_1079)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1097, y = v_9)[name = tensor("inner_9")]; + tensor var_1099_transpose_x_0 = const()[name = tensor("op_1099_transpose_x_0"), val = tensor(false)]; + tensor var_1099_transpose_y_0 = const()[name = tensor("op_1099_transpose_y_0"), val = tensor(false)]; + tensor var_1099 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1099")]; + tensor var_1100 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1100")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 1, 2, 1])]; + tensor var_1102 = reshape(shape = var_1101, x = var_1100)[name = tensor("op_1102")]; + tensor cross_9 = mul(x = var_1099, y = var_1102)[name = tensor("cross_9")]; + tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 1, 2, 1])]; + tensor var_1106 = reshape(shape = var_1105, x = valid_mask)[name = tensor("op_1106")]; + tensor v_masked_1 = mul(x = v_9, y = var_1106)[name = tensor("v_masked_1")]; + tensor var_1108 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1108")]; + tensor var_1110_transpose_x_1 = const()[name = tensor("op_1110_transpose_x_1"), val = tensor(true)]; + tensor var_1110_transpose_y_1 = const()[name = tensor("op_1110_transpose_y_1"), val = tensor(false)]; + tensor var_1110 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1110")]; + tensor new_kv_unnorm_9 = add(x = var_1108, y = var_1110)[name = tensor("new_kv_unnorm_9")]; + tensor var_1112_keep_dims_0 = const()[name = tensor("op_1112_keep_dims_0"), val = tensor(false)]; + tensor var_1112 = reduce_sum(keep_dims = var_1112_keep_dims_0, x = valid_mask)[name = tensor("op_1112")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1])]; + tensor var_1114 = reshape(shape = var_1113, x = var_1112)[name = tensor("op_1114")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1114)[name = tensor("new_scale_9")]; + tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; + tensor clip_2 = clip(alpha = var_38, 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_1118 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1118")]; + tensor var_1119_perm_0 = const()[name = tensor("op_1119_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_1119 = transpose(perm = var_1119_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_53, x = var_1119)[name = tensor("out_27")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([6, 2, 256])]; + tensor out_29 = reshape(shape = var_1123, x = out_27)[name = tensor("out_29")]; + tensor var_1125 = silu(x = input_171)[name = tensor("op_1125")]; + tensor input_173 = mul(x = var_1125, 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_45, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([1, 6, 2, 256])]; + tensor var_1136 = reshape(shape = var_1135, x = xt_1)[name = tensor("op_1136")]; + tensor var_1137_perm_0 = const()[name = tensor("op_1137_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1140 = const()[name = tensor("op_1140"), val = tensor([2, 6, 256])]; + tensor var_1137 = transpose(perm = var_1137_perm_0, x = var_1136)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1140, x = var_1137)[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_1163 = 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, 2, 3, 256])]; + tensor var_1165 = reshape(shape = concat_1, x = var_1163)[name = tensor("op_1165")]; + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([0])]; + tensor var_1166 = expand_dims(axes = var_1166_axes_0, x = var_1165)[name = tensor("op_1166")]; + tensor var_1167_perm_0 = const()[name = tensor("op_1167_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1168_axes_0 = const()[name = tensor("op_1168_axes_0"), val = tensor([-2])]; + tensor var_1167 = transpose(perm = var_1167_perm_0, x = var_1166)[name = tensor("transpose_21")]; + tensor var_1168 = squeeze(axes = var_1168_axes_0, x = var_1167)[name = tensor("op_1168")]; + 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, 2, 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_1168)[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, 2, 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_1168)[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, 2, 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_1168)[name = tensor("v_11")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([6, 8, 64])]; + tensor var_1177 = reshape(shape = var_1176, x = q_11)[name = tensor("op_1177")]; + tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([6, 8, 64])]; + tensor var_1184 = reshape(shape = var_1183, x = k_11)[name = tensor("op_1184")]; + tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([6, 8, 64])]; + tensor var_1191 = reshape(shape = var_1190, x = v_11)[name = tensor("op_1191")]; + tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([2, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1177)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1194, x = q_13)[name = tensor("q_15")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([2, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1184)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1196, x = k_13)[name = tensor("k_15")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([2, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1191)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1198, 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_1201 = const()[name = tensor("op_1201"), val = tensor([2, 0, 1, 3])]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([12, 256])]; + tensor var_1202 = transpose(perm = var_1201, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1206, x = var_1202)[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_1210 = const()[name = tensor("op_1210"), val = tensor([6, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1210, 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_45, 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_45, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 2, 6, 256])]; + tensor x_31 = reshape(shape = var_1230, x = xt_3)[name = tensor("x_31")]; + tensor var_1232_perm_0 = const()[name = tensor("op_1232_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([6, 2, 256])]; + tensor var_1232 = transpose(perm = var_1232_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1236, x = var_1232)[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_1244 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([6, 2, 4, 64])]; + tensor var_1246 = reshape(shape = var_1245, x = var_1244)[name = tensor("op_1246")]; + tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1250 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor(0x1p-3)]; + tensor var_1252 = mul(x = var_1250, y = var_1251)[name = tensor("op_1252")]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([6, 2, 4, 64])]; + tensor var_1254 = reshape(shape = var_1253, x = var_1252)[name = tensor("op_1254")]; + tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1258 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([6, 2, 4, 64])]; + tensor var_1260 = reshape(shape = var_1259, x = var_1258)[name = tensor("op_1260")]; + 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_38, 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_1254)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1246)[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_1275 = const()[name = tensor("op_1275"), val = tensor([2, 1])]; + tensor var_1276 = reshape(shape = var_1275, x = sqrt_s_t)[name = tensor("op_1276")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1276)[name = tensor("M")]; + tensor var_1278 = mul(x = qk, y = M)[name = tensor("op_1278")]; + 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_1260)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1278, y = v_17)[name = tensor("inner_11")]; + tensor var_1280_transpose_x_0 = const()[name = tensor("op_1280_transpose_x_0"), val = tensor(false)]; + tensor var_1280_transpose_y_0 = const()[name = tensor("op_1280_transpose_y_0"), val = tensor(false)]; + tensor var_1280 = matmul(transpose_x = var_1280_transpose_x_0, transpose_y = var_1280_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1280")]; + tensor var_1281 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1281")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([1, 1, 2, 1])]; + tensor var_1283 = reshape(shape = var_1282, x = var_1281)[name = tensor("op_1283")]; + tensor cross = mul(x = var_1280, y = var_1283)[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_1106)[name = tensor("v_masked")]; + tensor var_1289 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1289")]; + tensor var_1291_transpose_x_1 = const()[name = tensor("op_1291_transpose_x_1"), val = tensor(true)]; + tensor var_1291_transpose_y_1 = const()[name = tensor("op_1291_transpose_y_1"), val = tensor(false)]; + tensor var_1291 = matmul(transpose_x = var_1291_transpose_x_1, transpose_y = var_1291_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1291")]; + tensor new_kv_unnorm = add(x = var_1289, y = var_1291)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1114)[name = tensor("new_scale")]; + tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; + tensor clip_4 = clip(alpha = var_38, 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_1300_perm_0 = const()[name = tensor("op_1300_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_1300 = transpose(perm = var_1300_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_53, x = var_1300)[name = tensor("out_33")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([6, 2, 256])]; + tensor out = reshape(shape = var_1304, x = out_33)[name = tensor("out")]; + tensor var_1306 = silu(x = input_189)[name = tensor("op_1306")]; + tensor input_191 = mul(x = var_1306, 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_45, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 6, 2, 256])]; + tensor var_1317 = reshape(shape = var_1316, x = xt_5)[name = tensor("op_1317")]; + tensor var_1318_perm_0 = const()[name = tensor("op_1318_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([2, 6, 256])]; + tensor var_1318 = transpose(perm = var_1318_perm_0, x = var_1317)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1321, x = var_1318)[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_1344 = 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, 2, 3, 256])]; + tensor var_1346 = reshape(shape = concat_2, x = var_1344)[name = tensor("op_1346")]; + tensor var_1347_axes_0 = const()[name = tensor("op_1347_axes_0"), val = tensor([0])]; + tensor var_1347 = expand_dims(axes = var_1347_axes_0, x = var_1346)[name = tensor("op_1347")]; + tensor var_1348_perm_0 = const()[name = tensor("op_1348_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1349_axes_0 = const()[name = tensor("op_1349_axes_0"), val = tensor([-2])]; + tensor var_1348 = transpose(perm = var_1348_perm_0, x = var_1347)[name = tensor("transpose_8")]; + tensor var_1349 = squeeze(axes = var_1349_axes_0, x = var_1348)[name = tensor("op_1349")]; + 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, 2, 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_1349)[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, 2, 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_1349)[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, 2, 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_1349)[name = tensor("v_19")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([6, 8, 64])]; + tensor var_1358 = reshape(shape = var_1357, x = q_19)[name = tensor("op_1358")]; + tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([6, 8, 64])]; + tensor var_1365 = reshape(shape = var_1364, x = k_19)[name = tensor("op_1365")]; + tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([6, 8, 64])]; + tensor var_1372 = reshape(shape = var_1371, x = v_19)[name = tensor("op_1372")]; + tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([2, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1358)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1375, x = q_21)[name = tensor("q")]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor([2, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1365)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1377, x = k_21)[name = tensor("k")]; + tensor var_1379 = const()[name = tensor("op_1379"), val = tensor([2, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1372)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1379, 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_1382 = const()[name = tensor("op_1382"), val = tensor([2, 0, 1, 3])]; + tensor var_1387 = const()[name = tensor("op_1387"), val = tensor([12, 256])]; + tensor var_1383 = transpose(perm = var_1382, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1387, x = var_1383)[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_1391 = const()[name = tensor("op_1391"), val = tensor([6, 2, 256])]; + tensor attn_output = reshape(shape = var_1391, 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_45, 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_45, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 2, 6, 256])]; + tensor input = reshape(shape = var_1411, x = xt)[name = tensor("input")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([-1])]; + tensor var_1414 = reduce_l2_norm(axes = var_1413, keep_dims = var_44, x = input)[name = tensor("op_1414")]; + tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; + tensor clip_5 = clip(alpha = var_58, beta = const_42, x = var_1414)[name = tensor("clip_5")]; + tensor var_1416 = real_div(x = input, y = clip_5)[name = tensor("op_1416")]; + tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 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([2, 256, 6])]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1416)[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, 2, 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, 2, 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_1420")]; + tensor var_1422_axis_0 = const()[name = tensor("op_1422_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1422_axis_0, values = (var_1118, nkv))[name = tensor("op_1422")]; + tensor var_1424_axis_0 = const()[name = tensor("op_1424_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1424_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1424")]; + } -> (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