torch beamsearch.txt

graph(%self : __torch__.pytorch_translate.ensemble_export.BeamSearchAndDecode,
      %src_tokens.1 : Tensor,
      %src_lengths.1 : Tensor,
      %prev_token.1 : Tensor,
      %prev_scores.1 : Tensor,
      %attn_weights.1 : Tensor,
      %prev_hypos_indices.1 : Tensor,
      %num_steps.1 : int):
  %8 : None = prim::Constant()
  %9 : __torch__.pytorch_translate.ensemble_export.BeamSearch = prim::GetAttr[name="beam_search"](%self)
  %17 : Long(1, strides=[1], requires_grad=0, device=cpu) = prim::Constant[value={-1}]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
  %18 : Device = prim::Constant[value="cpu"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
  %19 : int = prim::Constant[value=4]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
  %20 : bool = prim::Constant[value=0]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
  %21 : None = prim::Constant()
  %22 : Float(103, strides=[1], requires_grad=0, device=cpu) = prim::Constant[value=<Tensor>]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:323:0
  %23 : int = prim::Constant[value=-1]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:332:0
  %24 : Long(requires_grad=0, device=cpu) = prim::Constant[value={1}]() # /usr/local/lib/python3.7/dist-packages/torch/tensor.py:559:0
  %25 : int = prim::Constant[value=1]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1062:45
  %26 : int = prim::Constant[value=0]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1058:79
  %27 : bool = prim::Constant[value=1]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1084:8
  %28 : __torch__.pytorch_translate.ensemble_export.EncoderEnsemble = prim::GetAttr[name="encoder_ens"](%9)
  %29 : __torch__.pytorch_translate.rnn.RNNModel = prim::GetAttr[name="model_0"](%28)
  %30 : __torch__.pytorch_translate.rnn.RNNDecoder = prim::GetAttr[name="decoder"](%29)
  %31 : Tensor = prim::GetAttr[name="initial_attn_context"](%30)
  %src_tokens : Long(1, 5, strides=[1, 1], requires_grad=0, device=cpu) = aten::t(%src_tokens.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:167:0
  %33 : Future[(Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Int(1, strides=[1], requires_grad=0, device=cpu), Long(1, 5, strides=[1, 1], requires_grad=0, device=cpu), Float(5, 1, 10, strides=[10, 50, 1], requires_grad=1, device=cpu))] = prim::fork_0(%28, %src_tokens, %src_lengths.1)
  %34 : (Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Int(1, strides=[1], requires_grad=0, device=cpu), Long(1, 5, strides=[1, 1], requires_grad=0, device=cpu), Float(5, 1, 10, strides=[10, 50, 1], requires_grad=1, device=cpu)) = aten::wait(%33)
  %35 : Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), %final_hiddens : Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), %final_cells : Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), %38 : Int(1, strides=[1], requires_grad=0, device=cpu), %39 : Long(1, 5, strides=[1, 1], requires_grad=0, device=cpu), %40 : Float(5, 1, 10, strides=[10, 50, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%34)
  %41 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::select(%final_hiddens, %26, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/rnn.py:1313:0
  %42 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::select(%final_hiddens, %26, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/rnn.py:1313:0
  %43 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::select(%final_cells, %26, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/rnn.py:1314:0
  %44 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::select(%final_cells, %26, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/rnn.py:1314:0
  %enc_states.1 : (Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(10, strides=[1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%35, %41, %43, %42, %44, %31)
  %46 : int[] = prim::ListConstruct(%25)
  %47 : Tensor = aten::repeat(%prev_token.1, %46) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1058:21
  %all_tokens.1 : Tensor = aten::unsqueeze(%47, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1058:21
  %49 : int[] = prim::ListConstruct(%25)
  %50 : Tensor = aten::repeat(%prev_scores.1, %49) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1059:21
  %all_scores.1 : Tensor = aten::unsqueeze(%50, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1059:21
  %52 : Tensor = aten::unsqueeze(%attn_weights.1, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1061:12
  %53 : int[] = prim::ListConstruct(%25, %25)
  %54 : Tensor = aten::repeat(%52, %53) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1061:12
  %all_weights.1 : Tensor = aten::unsqueeze(%54, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1061:12
  %all_prev_indices.1 : Tensor = aten::unsqueeze(%prev_hypos_indices.1, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1065:27
  %57 : __torch__.pytorch_translate.ensemble_export.DecoderBatchedStepEnsemble = prim::GetAttr[name="decoder_ens_tile"](%9)
  %58 : Tensor, %59 : Tensor, %60 : Tensor, %61 : Tensor, %62 : Tensor, %63 : Tensor = prim::TupleUnpack(%enc_states.1)
  %x.2 : Long(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::unsqueeze(%prev_token.1, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:301:0
  %65 : Long(2, strides=[1], requires_grad=0, device=cpu) = aten::_shape_as_tensor(%x.2) # /usr/local/lib/python3.7/dist-packages/torch/onnx/operators.py:16:0
  %batch_size.1 : Long(requires_grad=0, device=cpu) = aten::select(%65, %26, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:408:0
  %67 : int[] = prim::ListConstruct(%25)
  %68 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::view(%batch_size.1, %67) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
  %69 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::to(%17, %18, %19, %20, %20, %21) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
  %70 : Tensor[] = prim::ListConstruct(%68, %69)
  %shape.2 : Long(2, strides=[1], requires_grad=0, device=cpu) = aten::cat(%70, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
  %72 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::_reshape_from_tensor(%63, %shape.2) # /usr/local/lib/python3.7/dist-packages/torch/onnx/operators.py:20:0
  %73 : Future[(Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu))] = prim::fork_1(%57, %x.2, %72, %59, %60, %61, %62, %58)
  %74 : (Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = aten::wait(%73)
  %75 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), %76 : Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), %77 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), %78 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), %state.5 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%74)
  %state.6 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %state.7 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%77)
  %state.8 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %state.9 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%78)
  %84 : Tensor[] = prim::ListConstruct(%75)
  %85 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::cat(%84, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:314:0
  %86 : int[] = prim::ListConstruct(%25)
  %average_log_probs.1 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::mean(%85, %86, %27, %21) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:314:0
  %88 : Float(1, 103, strides=[103, 1], requires_grad=0, device=cpu) = aten::unsqueeze(%22, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:323:0
  %word_rewards.1 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=0, device=cpu) = aten::unsqueeze(%88, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:323:0
  %average_log_probs_with_rewards.1 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::add(%average_log_probs.1, %word_rewards.1, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:325:0
  %91 : Tensor[] = prim::ListConstruct(%76)
  %92 : Float(1, 1, 5, strides=[5, 5, 1], requires_grad=1, device=cpu) = aten::cat(%91, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:328:0
  %93 : int[] = prim::ListConstruct(%25)
  %average_attn_weights.1 : Float(1, 1, 5, strides=[5, 5, 1], requires_grad=1, device=cpu) = aten::mean(%92, %93, %27, %21) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:328:0
  %95 : Float(1, 103, strides=[103, 1], requires_grad=1, device=cpu) = aten::squeeze(%average_log_probs_with_rewards.1, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:332:0
  %best_scores_k_by_k.1 : Float(1, 1, strides=[1, 1], requires_grad=1, device=cpu), %best_tokens_k_by_k.1 : Long(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::topk(%95, %25, %23, %27, %27) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:332:0
  %98 : int[] = prim::ListConstruct(%23, %25)
  %99 : Float(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::view(%prev_scores.1, %98) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:335:0
  %100 : int[] = prim::ListConstruct(%23, %25)
  %prev_scores_k_by_k.1 : Float(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::expand(%99, %100, %20) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:335:0
  %total_scores_k_by_k.1 : Float(1, 1, strides=[1, 1], requires_grad=1, device=cpu) = aten::add(%best_scores_k_by_k.1, %prev_scores_k_by_k.1, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:336:0
  %103 : int[] = prim::ListConstruct(%23)
  %total_scores_flat.1 : Float(1, strides=[1], requires_grad=1, device=cpu) = aten::view(%total_scores_k_by_k.1, %103) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:339:0
  %105 : int[] = prim::ListConstruct(%23)
  %best_tokens_flat.1 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::view(%best_tokens_k_by_k.1, %105) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:340:0
  %107 : Float(1, strides=[1], requires_grad=1, device=cpu), %best_indices.2 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::topk(%total_scores_flat.1, %25, %23, %27, %27) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:342:0
  %109 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::index_select(%best_tokens_flat.1, %26, %best_indices.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:344:0
  %110 : int[] = prim::ListConstruct(%23)
  %111 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::view(%109, %110) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:344:0
  %prev_hypos.1 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::floor_divide(%best_indices.2, %24) # /usr/local/lib/python3.7/dist-packages/torch/tensor.py:559:0
  %attention_weights.1 : Float(1, 1, 5, strides=[5, 5, 1], requires_grad=1, device=cpu) = aten::index_select(%average_attn_weights.1, %26, %prev_hypos.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:349:0
  %114 : Float(1, 5, strides=[5, 1], requires_grad=1, device=cpu) = aten::squeeze(%attention_weights.1, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:364:0
  %115 : int[] = prim::ListConstruct(%25, %25, %25)
  %116 : Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu) = aten::repeat(%58, %115) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:376:0
  %117 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.6, %26, %prev_hypos.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
  %118 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.8, %26, %prev_hypos.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
  %119 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.7, %26, %prev_hypos.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
  %120 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.9, %26, %prev_hypos.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
  %121 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.5, %26, %prev_hypos.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
  %122 : (Long(1, strides=[1], requires_grad=0, device=cpu), Float(1, strides=[1], requires_grad=1, device=cpu), Long(1, strides=[1], requires_grad=0, device=cpu), Float(1, 5, strides=[5, 1], requires_grad=1, device=cpu), Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%111, %107, %prev_hypos.1, %114, %116, %117, %118, %119, %120, %121)
  %prev_token.4 : Tensor, %prev_scores.4 : Tensor, %prev_hypos_indices.3 : Tensor, %attn_weights.3 : Tensor, %127 : Tensor, %128 : Tensor, %129 : Tensor, %130 : Tensor, %131 : Tensor, %132 : Tensor = prim::TupleUnpack(%122)
  %states.1 : (Tensor, Tensor, Tensor, Tensor, Tensor, Tensor) = prim::TupleConstruct(%127, %128, %129, %130, %131, %132)
  %134 : Tensor = aten::unsqueeze(%prev_token.4, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1077:44
  %135 : Tensor[] = prim::ListConstruct(%all_tokens.1, %134)
  %all_tokens.3 : Tensor = aten::cat(%135, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1077:21
  %137 : Tensor = aten::unsqueeze(%prev_scores.4, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1078:44
  %138 : Tensor[] = prim::ListConstruct(%all_scores.1, %137)
  %all_scores.3 : Tensor = aten::cat(%138, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1078:21
  %140 : Tensor = aten::unsqueeze(%attn_weights.3, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1079:46
  %141 : Tensor[] = prim::ListConstruct(%all_weights.1, %140)
  %all_weights.3 : Tensor = aten::cat(%141, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1079:22
  %143 : Tensor = aten::unsqueeze(%prev_hypos_indices.3, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1081:31
  %144 : Tensor[] = prim::ListConstruct(%all_prev_indices.1, %143)
  %all_prev_indices.3 : Tensor = aten::cat(%144, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1080:27
  %146 : int = aten::sub(%num_steps.1, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1084:23
  %all_tokens : Tensor, %all_scores : Tensor, %all_weights : Tensor, %all_prev_indices : Tensor, %prev_scores : Tensor, %prev_token : Tensor, %states : (Tensor, Tensor, Tensor, Tensor, Tensor, Tensor) = prim::Loop(%146, %27, %all_tokens.3, %all_scores.3, %all_weights.3, %all_prev_indices.3, %prev_scores.4, %prev_token.4, %states.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1084:8
    block0(%i.1 : int, %all_tokens.8 : Tensor, %all_scores.8 : Tensor, %all_weights.8 : Tensor, %all_prev_indices.8 : Tensor, %prev_scores.10 : Tensor, %prev_token.10 : Tensor, %states.5 : (Tensor, Tensor, Tensor, Tensor, Tensor, Tensor)):
      %162 : __torch__.pytorch_translate.ensemble_export.___torch_mangle_32.DecoderBatchedStepEnsemble = prim::GetAttr[name="decoder_ens"](%9)
      %163 : Tensor, %164 : Tensor, %165 : Tensor, %166 : Tensor, %167 : Tensor, %168 : Tensor = prim::TupleUnpack(%states.5)
      %x.1 : Long(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::unsqueeze(%prev_token.10, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:301:0
      %170 : Long(2, strides=[1], requires_grad=0, device=cpu) = aten::_shape_as_tensor(%x.1) # /usr/local/lib/python3.7/dist-packages/torch/onnx/operators.py:16:0
      %batch_size : Long(requires_grad=0, device=cpu) = aten::select(%170, %26, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:408:0
      %172 : int[] = prim::ListConstruct(%25)
      %173 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::view(%batch_size, %172) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
      %174 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::to(%17, %18, %19, %20, %20, %21) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
      %175 : Tensor[] = prim::ListConstruct(%173, %174)
      %shape.1 : Long(2, strides=[1], requires_grad=0, device=cpu) = aten::cat(%175, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:438:0
      %177 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::_reshape_from_tensor(%168, %shape.1) # /usr/local/lib/python3.7/dist-packages/torch/onnx/operators.py:20:0
      %178 : Future[(Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu))] = prim::fork_2(%162, %x.1, %177, %164, %165, %166, %167, %163)
      %179 : (Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = aten::wait(%178)
      %180 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), %181 : Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), %182 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), %183 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), %state : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%179)
      %state.10 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %state.3 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%182)
      %state.2 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %state.4 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%183)
      %189 : Tensor[] = prim::ListConstruct(%180)
      %190 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::cat(%189, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:314:0
      %191 : int[] = prim::ListConstruct(%25)
      %average_log_probs : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::mean(%190, %191, %27, %21) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:314:0
      %193 : Float(1, 103, strides=[103, 1], requires_grad=0, device=cpu) = aten::unsqueeze(%22, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:323:0
      %word_rewards : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=0, device=cpu) = aten::unsqueeze(%193, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:323:0
      %average_log_probs_with_rewards : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::add(%average_log_probs, %word_rewards, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:325:0
      %196 : Tensor[] = prim::ListConstruct(%181)
      %197 : Float(1, 1, 5, strides=[5, 5, 1], requires_grad=1, device=cpu) = aten::cat(%196, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:328:0
      %198 : int[] = prim::ListConstruct(%25)
      %average_attn_weights : Float(1, 1, 5, strides=[5, 5, 1], requires_grad=1, device=cpu) = aten::mean(%197, %198, %27, %21) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:328:0
      %200 : Float(1, 103, strides=[103, 1], requires_grad=1, device=cpu) = aten::squeeze(%average_log_probs_with_rewards, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:332:0
      %best_scores_k_by_k : Float(1, 1, strides=[1, 1], requires_grad=1, device=cpu), %best_tokens_k_by_k : Long(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::topk(%200, %25, %23, %27, %27) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:332:0
      %203 : int[] = prim::ListConstruct(%23, %25)
      %204 : Float(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::view(%prev_scores.10, %203) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:335:0
      %205 : int[] = prim::ListConstruct(%23, %25)
      %prev_scores_k_by_k : Float(1, 1, strides=[1, 1], requires_grad=0, device=cpu) = aten::expand(%204, %205, %20) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:335:0
      %total_scores_k_by_k : Float(1, 1, strides=[1, 1], requires_grad=1, device=cpu) = aten::add(%best_scores_k_by_k, %prev_scores_k_by_k, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:336:0
      %208 : int[] = prim::ListConstruct(%23)
      %total_scores_flat : Float(1, strides=[1], requires_grad=1, device=cpu) = aten::view(%total_scores_k_by_k, %208) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:339:0
      %210 : int[] = prim::ListConstruct(%23)
      %best_tokens_flat : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::view(%best_tokens_k_by_k, %210) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:340:0
      %212 : Float(1, strides=[1], requires_grad=1, device=cpu), %best_indices : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::topk(%total_scores_flat, %25, %23, %27, %27) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:342:0
      %214 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::index_select(%best_tokens_flat, %26, %best_indices) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:344:0
      %215 : int[] = prim::ListConstruct(%23)
      %216 : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::view(%214, %215) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:344:0
      %prev_hypos : Long(1, strides=[1], requires_grad=0, device=cpu) = aten::floor_divide(%best_indices, %24) # /usr/local/lib/python3.7/dist-packages/torch/tensor.py:559:0
      %attention_weights : Float(1, 1, 5, strides=[5, 5, 1], requires_grad=1, device=cpu) = aten::index_select(%average_attn_weights, %26, %prev_hypos) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:349:0
      %219 : Float(1, 5, strides=[5, 1], requires_grad=1, device=cpu) = aten::squeeze(%attention_weights, %25) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:364:0
      %220 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.10, %26, %prev_hypos) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
      %221 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.2, %26, %prev_hypos) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
      %222 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.3, %26, %prev_hypos) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
      %223 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state.4, %26, %prev_hypos) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
      %224 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = aten::index_select(%state, %26, %prev_hypos) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:389:0
      %225 : (Long(1, strides=[1], requires_grad=0, device=cpu), Float(1, strides=[1], requires_grad=1, device=cpu), Long(1, strides=[1], requires_grad=0, device=cpu), Float(1, 5, strides=[5, 1], requires_grad=1, device=cpu), Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%216, %212, %prev_hypos, %219, %163, %220, %221, %222, %223, %224)
      %prev_token.7 : Tensor, %prev_scores.7 : Tensor, %prev_hypos_indices.5 : Tensor, %attn_weights.5 : Tensor, %230 : Tensor, %231 : Tensor, %232 : Tensor, %233 : Tensor, %234 : Tensor, %235 : Tensor = prim::TupleUnpack(%225)
      %states.3 : (Tensor, Tensor, Tensor, Tensor, Tensor, Tensor) = prim::TupleConstruct(%230, %231, %232, %233, %234, %235)
      %237 : Tensor = aten::unsqueeze(%prev_token.7, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1095:48
      %238 : Tensor[] = prim::ListConstruct(%all_tokens.8, %237)
      %all_tokens.5 : Tensor = aten::cat(%238, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1095:25
      %240 : Tensor = aten::unsqueeze(%prev_scores.7, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1096:48
      %241 : Tensor[] = prim::ListConstruct(%all_scores.8, %240)
      %all_scores.5 : Tensor = aten::cat(%241, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1096:25
      %243 : Tensor = aten::unsqueeze(%attn_weights.5, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1097:50
      %244 : Tensor[] = prim::ListConstruct(%all_weights.8, %243)
      %all_weights.5 : Tensor = aten::cat(%244, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1097:26
      %246 : Tensor = aten::unsqueeze(%prev_hypos_indices.5, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1099:35
      %247 : Tensor[] = prim::ListConstruct(%all_prev_indices.8, %246)
      %all_prev_indices.5 : Tensor = aten::cat(%247, %26) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/ensemble_export.py:1098:31
      -> (%27, %all_tokens.5, %all_scores.5, %all_weights.5, %all_prev_indices.5, %prev_scores.7, %prev_token.7, %states.3)
  %beam_search_out.1 : (Tensor, Tensor, Tensor, Tensor) = prim::TupleConstruct(%all_tokens, %all_scores, %all_weights, %all_prev_indices)
  %all_tokens.2 : Tensor, %all_scores.2 : Tensor, %all_weights.2 : Tensor, %all_prev_indices.2 : Tensor = prim::TupleUnpack(%beam_search_out.1)
  %15 : __torch__.pytorch_translate.beam_decode.BeamDecode = prim::GetAttr[name="beam_decode"](%self)
  %250 : float = prim::Constant[value=inf]()
  %251 : str = prim::Constant[value="Dimension of beam_prev_indices : {} and num_steps : {} are not consistent"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1025:12
  %252 : str = prim::Constant[value="Dimension of token_weights : {} and num_steps : {} are not consistent"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1021:12
  %253 : str = prim::Constant[value="Dimension of beam_scores : {} and num_steps : {} are not consistent"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1017:12
  %254 : str = prim::Constant[value="Dimension of beam_tokens : {} and num_steps : {} are not consistent"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1013:12
  %255 : str = prim::Constant[value="AssertionError: Dimension of beam_prev_indices : {} and beam size : {} "]()
  %256 : str = prim::Constant[value="Dimension of token_weights : {} and beam size : {} are not consistent"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1004:12
  %257 : str = prim::Constant[value="Dimension of beam_scores : {} and beam size : {} are not consistent"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1000:12
  %258 : str = prim::Constant[value="Dimension of beam_tokens : {} and beam size : {} are not consistent"]() # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:996:11
  %259 : int = prim::Constant[value=9223372036854775807]() # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:855:12
  %260 : int = prim::Constant[value=2]() # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:845:34
  %261 : int = prim::Constant[value=1]() # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:844:33
  %262 : int = prim::Constant[value=0]() # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:838:43
  %263 : bool = prim::Constant[value=1]() # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:836:8
  %264 : None = prim::Constant()
  %265 : bool = prim::Constant[value=0]()
  %266 : str = prim::Constant[value="AssertionError: "]()
  %prev_beam_index.1 : int = prim::Constant[value=-1]() # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:854:30
  %268 : int = aten::size(%all_tokens.2, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:995:12
  %269 : int = prim::GetAttr[name="beam_size"](%15)
  %270 : bool = aten::eq(%268, %269) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:995:12
   = prim::If(%270) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:994:8
    block0():
      -> ()
    block1():
      %271 : int[] = aten::size(%all_tokens.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:997:12
      %272 : int = prim::GetAttr[name="beam_size"](%15)
      %273 : str = aten::format(%258, %271, %272) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:996:11
      %274 : str = aten::add(%266, %273) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:994:8
       = prim::RaiseException(%274) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:994:8
      -> ()
  %275 : int = aten::size(%all_scores.2, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:999:15
  %276 : int = prim::GetAttr[name="beam_size"](%15)
  %277 : bool = aten::eq(%275, %276) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:999:15
   = prim::If(%277) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:999:8
    block0():
      -> ()
    block1():
      %278 : int[] = aten::size(%all_scores.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1001:40
      %279 : int = prim::GetAttr[name="beam_size"](%15)
      %280 : str = aten::format(%257, %278, %279) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1000:12
      %281 : str = aten::add(%266, %280) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:999:8
       = prim::RaiseException(%281) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:999:8
      -> ()
  %282 : int = aten::size(%all_weights.2, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1003:15
  %283 : int = prim::GetAttr[name="beam_size"](%15)
  %284 : bool = aten::eq(%282, %283) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1003:15
   = prim::If(%284) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1003:8
    block0():
      -> ()
    block1():
      %285 : int[] = aten::size(%all_weights.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1005:40
      %286 : int = prim::GetAttr[name="beam_size"](%15)
      %287 : str = aten::format(%256, %285, %286) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1004:12
      %288 : str = aten::add(%266, %287) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1003:8
       = prim::RaiseException(%288) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1003:8
      -> ()
  %289 : int = aten::size(%all_prev_indices.2, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1008:12
  %290 : int = prim::GetAttr[name="beam_size"](%15)
  %291 : bool = aten::eq(%289, %290) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1008:12
   = prim::If(%291) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1007:8
    block0():
      -> ()
    block1():
       = prim::RaiseException(%255) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1007:8
      -> ()
  %292 : int = aten::size(%all_tokens.2, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1012:15
  %293 : int = aten::add(%num_steps.1, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1012:38
  %294 : bool = aten::le(%292, %293) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1012:15
   = prim::If(%294) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1012:8
    block0():
      -> ()
    block1():
      %295 : int[] = aten::size(%all_tokens.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1014:40
      %296 : str = aten::format(%254, %295, %num_steps.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1013:12
      %297 : str = aten::add(%266, %296) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1012:8
       = prim::RaiseException(%297) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1012:8
      -> ()
  %298 : int = aten::size(%all_scores.2, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1016:15
  %299 : int = aten::add(%num_steps.1, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1016:38
  %300 : bool = aten::le(%298, %299) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1016:15
   = prim::If(%300) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1016:8
    block0():
      -> ()
    block1():
      %301 : int[] = aten::size(%all_scores.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1018:40
      %302 : str = aten::format(%253, %301, %num_steps.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1017:12
      %303 : str = aten::add(%266, %302) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1016:8
       = prim::RaiseException(%303) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1016:8
      -> ()
  %304 : int = aten::size(%all_weights.2, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1020:15
  %305 : int = aten::add(%num_steps.1, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1020:40
  %306 : bool = aten::le(%304, %305) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1020:15
   = prim::If(%306) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1020:8
    block0():
      -> ()
    block1():
      %307 : int[] = aten::size(%all_weights.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1022:40
      %308 : str = aten::format(%252, %307, %num_steps.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1021:12
      %309 : str = aten::add(%266, %308) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1020:8
       = prim::RaiseException(%309) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1020:8
      -> ()
  %310 : int = aten::size(%all_prev_indices.2, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1024:15
  %311 : int = aten::add(%num_steps.1, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1024:44
  %312 : bool = aten::le(%310, %311) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1024:15
   = prim::If(%312) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1024:8
    block0():
      -> ()
    block1():
      %313 : int[] = aten::size(%all_prev_indices.2) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1026:40
      %314 : str = aten::format(%251, %313, %num_steps.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1025:12
      %315 : str = aten::add(%266, %314) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1024:8
       = prim::RaiseException(%315) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:1024:8
      -> ()
  %end_states.1 : Tensor[] = prim::ListConstruct()
  %317 : int = prim::GetAttr[name="beam_size"](%15)
  %318 : int[] = prim::ListConstruct(%317)
  %319 : Tensor = aten::zeros(%318, %264, %264, %264, %264) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:938:32
  %prev_hypo_is_finished.1 : Tensor = aten::to(%319, %262, %265, %265, %264) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:938:32
  %321 : bool = aten::le(%261, %num_steps.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:941:19
  %end_states : Tensor[], %min_index.4 : int, %min_score.1 : float, %position : int, %prev_hypo_is_finished : Tensor = prim::Loop(%259, %321, %end_states.1, %prev_beam_index.1, %250, %261, %prev_hypo_is_finished.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:941:8
    block0(%327 : int, %end_states.21 : Tensor[], %min_index.15 : int, %min_score.15 : float, %position.18 : int, %prev_hypo_is_finished.4 : Tensor):
      %333 : int = prim::GetAttr[name="beam_size"](%15)
      %334 : int[] = prim::ListConstruct(%333)
      %335 : Tensor = aten::zeros(%334, %264, %264, %264, %264) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:942:31
      %hypo_is_finished.1 : Tensor = aten::to(%335, %262, %265, %265, %264) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:942:31
      %337 : int = prim::GetAttr[name="beam_size"](%15)
      %end_states.20 : Tensor[], %min_index.14 : int, %min_score.14 : float = prim::Loop(%337, %263, %end_states.21, %min_index.15, %min_score.15) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:944:12
        block0(%hyp_index.1 : int, %end_states.19 : Tensor[], %min_index.13 : int, %min_score.13 : float):
          %345 : Tensor = aten::select(%all_prev_indices.2, %262, %position.18) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:945:27
          %prev_pos.1 : Tensor = aten::select(%345, %262, %hyp_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:945:27
          %347 : Tensor?[] = prim::ListConstruct(%prev_pos.1)
          %348 : Tensor = aten::index(%prev_hypo_is_finished.4, %347) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:946:46
          %349 : Tensor = aten::select(%hypo_is_finished.1, %262, %hyp_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:946:16
          %350 : Tensor = aten::copy_(%349, %348, %265) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:946:16
          %351 : Tensor = aten::select(%hypo_is_finished.1, %262, %hyp_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:950:24
          %352 : Tensor = aten::eq(%351, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:950:24
          %353 : bool = aten::Bool(%352) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:950:19
          %end_states.18 : Tensor[], %min_index.12 : int, %min_score.16 : float = prim::If(%353) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:950:16
            block0():
              %357 : Tensor = aten::select(%all_tokens.2, %262, %position.18) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:954:24
              %358 : Tensor = aten::select(%357, %262, %hyp_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:954:24
              %359 : int = prim::GetAttr[name="eos_token_id"](%15)
              %360 : Tensor = aten::eq(%358, %359) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:954:24
              %361 : bool = aten::Bool(%360) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:953:23
              %362 : bool = prim::If(%361) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:953:23
                block0():
                  -> (%263)
                block1():
                  %363 : bool = aten::eq(%position.18, %num_steps.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:955:30
                  -> (%363)
              %end_states.17 : Tensor[], %min_index.11 : int, %min_score.11 : float = prim::If(%362) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:953:20
                block0():
                  %367 : int = prim::GetAttr[name="stop_at_eos"](%15)
                  %368 : bool = aten::Bool(%367) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:957:27
                   = prim::If(%368) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:957:24
                    block0():
                      %369 : Tensor = aten::select(%hypo_is_finished.1, %262, %hyp_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:958:28
                      %370 : int = prim::dtype(%369)
                      %371 : Device = prim::device(%369)
                      %372 : Tensor = aten::tensor(%261, %370, %371, %265)
                      %373 : Tensor = aten::copy_(%369, %372, %265) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:958:28
                      -> ()
                    block1():
                      -> ()
                  %374 : Tensor = aten::select(%all_scores.2, %262, %position.18) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:960:43
                  %375 : Tensor = aten::select(%374, %262, %hyp_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:960:43
                  %hypo_score.1 : float = aten::Float(%375) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:960:37
                  %377 : float = prim::GetAttr[name="length_penalty"](%15)
                  %378 : bool = aten::ne(%377, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:961:32
                  %hypo_score : float = prim::If(%378) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:961:24
                    block0():
                      %380 : float = aten::Float(%position.18) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:962:54
                      %381 : float = prim::GetAttr[name="length_penalty"](%15)
                      %382 : float = aten::pow(%380, %381) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:962:54
                      %hypo_score.3 : float = aten::div(%hypo_score.1, %382) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:962:41
                      -> (%hypo_score.3)
                    block1():
                      -> (%hypo_score.1)
                  %384 : float = aten::Float(%position.18) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:970:45
                  %385 : float = aten::Float(%hyp_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:970:62
                  %386 : float[] = prim::ListConstruct(%hypo_score, %384, %385)
                  %387 : Tensor = aten::tensor(%386, %264, %264, %265) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:969:28
                  %388 : int = aten::len(%end_states.19) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:905:11
                  %389 : int = prim::GetAttr[name="nbest"](%15)
                  %390 : bool = aten::lt(%388, %389) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:905:11
                  %min_score : float, %min_index : int = prim::If(%390) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:905:8
                    block0():
                      %393 : Tensor[] = aten::append(%end_states.19, %387) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:906:12
                      %394 : Tensor = aten::select(%387, %262, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:908:21
                      %395 : float = aten::Float(%394) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:908:15
                      %396 : bool = aten::le(%395, %min_score.13) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:908:15
                      %min_score.25 : float, %min_index.22 : int = prim::If(%396) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:908:12
                        block0():
                          %399 : Tensor = aten::select(%387, %262, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:909:34
                          %min_score.3 : float = aten::Float(%399) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:909:28
                          %401 : int = aten::len(%end_states.19) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:910:28
                          %min_index.2 : int = aten::sub(%401, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:910:28
                          -> (%min_score.3, %min_index.2)
                        block1():
                          -> (%min_score.13, %min_index.13)
                      -> (%min_score.25, %min_index.22)
                    block1():
                      %403 : Tensor = aten::select(%387, %262, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:911:18
                      %404 : Tensor = aten::gt(%403, %min_score.13) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:911:18
                      %405 : bool = aten::Bool(%404) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:911:13
                      %min_score.29 : float, %min_index.26 : int = prim::If(%405) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:911:8
                        block0():
                          %408 : Tensor[] = aten::_set_item(%end_states.19, %min_index.13, %387) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:913:12
                          %409 : int = aten::len(%end_states.19) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:917:29
                          %min_score.28 : float, %min_index.25 : int = prim::Loop(%409, %263, %250, %prev_beam_index.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:917:12
                            block0(%idx.1 : int, %min_score.27 : float, %min_index.24 : int):
                              %s.1 : Tensor = aten::__getitem__(%end_states.19, %idx.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:918:20
                              %416 : Tensor = aten::select(%s.1, %262, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:919:30
                              %417 : float = aten::Float(%416) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:919:24
                              %418 : bool = aten::le(%417, %min_score.27) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:919:24
                              %min_score.26 : float, %min_index.23 : int = prim::If(%418) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:919:16
                                block0():
                                  %421 : Tensor = aten::select(%s.1, %262, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:921:38
                                  %min_score.12 : float = aten::Float(%421) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:921:32
                                  -> (%min_score.12, %idx.1)
                                block1():
                                  -> (%min_score.27, %min_index.24)
                              -> (%263, %min_score.26, %min_index.23)
                          -> (%min_score.28, %min_index.25)
                        block1():
                          -> (%min_score.13, %min_index.13)
                      -> (%min_score.29, %min_index.26)
                  %423 : (Tensor[], float, int) = prim::TupleConstruct(%end_states.19, %min_score, %min_index)
                  %end_states.3 : Tensor[], %min_score.4 : float, %min_index.3 : int = prim::TupleUnpack(%423)
                  -> (%end_states.3, %min_index.3, %min_score.4)
                block1():
                  -> (%end_states.19, %min_index.13, %min_score.13)
              -> (%end_states.17, %min_index.11, %min_score.11)
            block1():
              -> (%end_states.19, %min_index.13, %min_score.13)
          -> (%263, %end_states.18, %min_index.12, %min_score.16)
      %position.16 : int = aten::add(%position.18, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:976:23
      %428 : bool = aten::le(%position.16, %num_steps.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:941:19
      -> (%428, %end_states.20, %min_index.14, %min_score.14, %position.16, %hypo_is_finished.1)
  %end_states.12 : Tensor = aten::stack(%end_states, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:978:21
  %430 : Tensor = aten::slice(%end_states.12, %262, %262, %259, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:980:38
  %431 : Tensor = aten::select(%430, %261, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:980:38
  %432 : Tensor, %sorted_end_state_indices.1 : Tensor = aten::sort(%431, %262, %263) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:980:38
  %434 : Tensor = aten::slice(%end_states.12, %261, %262, %259, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:981:21
  %435 : Tensor?[] = prim::ListConstruct(%sorted_end_state_indices.1)
  %end_states.2 : Tensor = aten::index(%434, %435) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:981:21
  %outputs.2 : (Tensor, float, float[], Tensor, Tensor)[] = prim::ListConstruct()
  %438 : int = aten::len(%end_states.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:836:31
   = prim::Loop(%438, %263) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:836:8
    block0(%state_idx.1 : int):
      %state.1 : Tensor = aten::select(%end_states.2, %262, %state_idx.1) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:837:20
      %441 : Tensor = aten::select(%state.1, %262, %262) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:838:37
      %hypothesis_score.1 : float = aten::Float(%441) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:838:31
      %443 : Tensor = aten::select(%state.1, %262, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:889:31
      %present_position.1 : int = aten::Int(%443) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:889:27
      %445 : Tensor = aten::select(%state.1, %262, %260) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:890:25
      %beam_index.1 : int = aten::Int(%445) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:890:21
      %beam_indices.2 : int[] = prim::ListConstruct()
      %448 : bool = aten::ge(%present_position.1, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:892:14
      %beam_index : int, %present_position : int = prim::Loop(%259, %448, %beam_index.1, %present_position.1) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:892:8
        block0(%451 : int, %beam_index.8 : int, %present_position.9 : int):
           = aten::insert(%beam_indices.2, %262, %beam_index.8) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:893:12
          %454 : Tensor = aten::select(%all_prev_indices.2, %262, %present_position.9) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:894:29
          %455 : Tensor = aten::select(%454, %262, %beam_index.8) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:894:29
          %beam_index.6 : int = aten::Int(%455) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:894:25
          %present_position.7 : int = aten::sub(%present_position.9, %261) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:895:31
          %458 : bool = aten::ge(%present_position.7, %262) # /usr/local/lib/python3.7/dist-packages/pytorch_translate/beam_decode.py:892:14
          -> (%458, %beam_index.6, %present_position.7)
      %beam_output.1 : Tensor[] = prim::ListConstruct()
      %token_level_scores.1 : float[] = prim::ListConstruct()
      %461 : Tensor = aten::select(%state.1, %262, %261) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:844:27
      %position.2 : int = aten::Int(%461) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:844:23
      %463 : Tensor = aten::select(%state.1, %262, %260) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:845:28
      %hyp_index.2 : int = aten::Int(%463) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:845:24
      %465 : int[] = prim::ListConstruct(%position.2, %hyp_index.2)
      %best_indices.1 : Tensor = aten::tensor(%465, %264, %264, %265) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:849:27
      %back_alignment_weights.1 : Tensor[] = prim::ListConstruct()
      %468 : int = aten::add(%position.2, %261) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:852:19
      %469 : int = aten::len(%beam_indices.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:852:35
      %470 : bool = aten::eq(%468, %469) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:852:19
       = prim::If(%470) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:852:12
        block0():
          -> ()
        block1():
           = prim::RaiseException(%266) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:852:12
          -> ()
      %471 : int = aten::len(%beam_indices.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:855:24
      %472 : bool = aten::lt(%261, %471) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:855:18
      %pos : int, %prev_beam_index : int = prim::Loop(%259, %472, %261, %prev_beam_index.1) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:855:12
        block0(%475 : int, %pos.22 : int, %prev_beam_index.6 : int):
          %beam_index.2 : int = aten::__getitem__(%beam_indices.2, %pos.22) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:856:29
          %479 : Tensor = aten::select(%all_tokens.2, %262, %pos.22) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:857:35
          %480 : Tensor = aten::select(%479, %262, %beam_index.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:857:35
          %481 : Tensor[] = aten::append(%beam_output.1, %480) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:857:16
          %482 : bool = aten::eq(%pos.22, %261) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:858:19
           = prim::If(%482) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:858:16
            block0():
              %483 : Tensor = aten::select(%all_scores.2, %262, %pos.22) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:860:52
              %484 : Tensor = aten::select(%483, %262, %beam_index.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:860:52
              %485 : float = aten::Float(%484) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:860:46
              %486 : float[] = aten::append(%token_level_scores.1, %485) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:860:20
              -> ()
            block1():
              %487 : Tensor = aten::select(%all_scores.2, %262, %pos.22) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:863:30
              %488 : Tensor = aten::select(%487, %262, %beam_index.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:863:30
              %489 : float = aten::Float(%488) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:863:24
              %490 : int = aten::sub(%pos.22, %261) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:864:44
              %491 : Tensor = aten::select(%all_scores.2, %262, %490) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:864:32
              %492 : Tensor = aten::select(%491, %262, %prev_beam_index.6) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:864:32
              %493 : float = aten::Float(%492) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:864:26
              %494 : float = aten::sub(%489, %493) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:863:24
              %495 : float[] = aten::append(%token_level_scores.1, %494) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:862:20
              -> ()
          %496 : Tensor = aten::select(%all_weights.2, %262, %pos.22) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:866:46
          %497 : Tensor = aten::select(%496, %262, %beam_index.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:866:46
          %498 : Tensor = aten::detach(%497) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:866:46
          %499 : Tensor[] = aten::append(%back_alignment_weights.1, %498) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:866:16
          %pos.20 : int = aten::add(%pos.22, %261) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:868:16
          %501 : int = aten::len(%beam_indices.2) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:855:24
          %502 : bool = aten::lt(%pos.20, %501) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:855:18
          -> (%502, %pos.20, %beam_index.2)
      %503 : Tensor = aten::stack(%beam_output.1, %262) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:871:20
      %504 : Tensor = aten::stack(%back_alignment_weights.1, %261) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:874:20
      %505 : (Tensor, float, float[], Tensor, Tensor) = prim::TupleConstruct(%503, %hypothesis_score.1, %token_level_scores.1, %504, %best_indices.1)
      %506 : (Tensor, float, float[], Tensor, Tensor)[] = aten::append(%outputs.2, %505) # /usr/local/lib/python3.7/dist-packages/torch/autograd/grad_mode.py:869:12
      -> (%263)
  return (%outputs.2)
with prim::fork_0 = graph(%self.1 : __torch__.pytorch_translate.ensemble_export.EncoderEnsemble,
      %src_tokens : Long(1, 5, strides=[1, 1], requires_grad=0, device=cpu),
      %lengths.1 : Int(1, strides=[1], requires_grad=0, device=cpu)):
  %3 : __torch__.pytorch_translate.rnn.RNNModel = prim::GetAttr[name="model_0"](%self.1)
  %4 : __torch__.pytorch_translate.rnn.LSTMSequenceEncoder = prim::GetAttr[name="encoder"](%3)
  %5 : (Tensor, Tensor, Tensor, Tensor, Tensor) = prim::CallMethod[name="forward"](%4, %src_tokens, %lengths.1)
  %6 : Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), %7 : Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), %8 : Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), %9 : Int(1, strides=[1], requires_grad=0, device=cpu), %10 : Float(5, 1, 10, strides=[10, 50, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%5)
  %11 : (Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Float(2, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu), Int(1, strides=[1], requires_grad=0, device=cpu), Long(1, 5, strides=[1, 1], requires_grad=0, device=cpu), Float(5, 1, 10, strides=[10, 50, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%6, %7, %8, %9, %src_tokens, %10)
  return (%11)
with prim::fork_1 = graph(%self.1 : __torch__.pytorch_translate.ensemble_export.DecoderBatchedStepEnsemble,
      %x.1 : Long(1, 1, strides=[1, 1], requires_grad=0, device=cpu),
      %2 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx.1 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx.2 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx.3 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %encoder_output : Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu)):
  %8 : __torch__.pytorch_translate.rnn.___torch_mangle_17.RNNModel = prim::GetAttr[name="model_0"](%self.1)
  %9 : __torch__.pytorch_translate.rnn.___torch_mangle_16.RNNDecoder = prim::GetAttr[name="decoder"](%8)
  %10 : (Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor) = prim::CallMethod[name="forward"](%9, %x.1, %2, %hx.1, %hx.2, %hx.3, %hx, %encoder_output)
  %11 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), %12 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %13 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %14 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %15 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %16 : Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), %17 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%10)
  %18 : int = prim::Constant[value=2]() # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1672:0
  %19 : None = prim::Constant()
  %20 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::log_softmax(%11, %18, %19) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1672:0
  %21 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%12, %13)
  %22 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%14, %15)
  %23 : (Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%20, %16, %21, %22, %17)
  return (%23)
with prim::fork_2 = graph(%self.1 : __torch__.pytorch_translate.ensemble_export.___torch_mangle_32.DecoderBatchedStepEnsemble,
      %x.1 : Long(1, 1, strides=[1, 1], requires_grad=0, device=cpu),
      %2 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx.1 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx.2 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx.3 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %hx : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu),
      %encoder_output : Float(5, 1, 10, strides=[10, 10, 1], requires_grad=1, device=cpu)):
  %8 : __torch__.pytorch_translate.rnn.___torch_mangle_31.RNNModel = prim::GetAttr[name="model_0"](%self.1)
  %9 : __torch__.pytorch_translate.rnn.___torch_mangle_30.RNNDecoder = prim::GetAttr[name="decoder"](%8)
  %10 : (Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor) = prim::CallMethod[name="forward"](%9, %x.1, %2, %hx.1, %hx.2, %hx.3, %hx, %encoder_output)
  %11 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), %12 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %13 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %14 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %15 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), %16 : Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), %17 : Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu) = prim::TupleUnpack(%10)
  %18 : int = prim::Constant[value=2]() # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1672:0
  %19 : None = prim::Constant()
  %20 : Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu) = aten::log_softmax(%11, %18, %19) # /usr/local/lib/python3.7/dist-packages/torch/nn/functional.py:1672:0
  %21 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%12, %13)
  %22 : (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%14, %15)
  %23 : (Float(1, 1, 103, strides=[103, 103, 1], requires_grad=1, device=cpu), Float(1, 1, 5, strides=[1, 1, 1], requires_grad=1, device=cpu), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), (Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)), Float(1, 10, strides=[10, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%20, %16, %21, %22, %17)
  return (%23)