Update past_key_values in GPT-2

src/transformers/generation_utils.py

@@ -504,7 +504,8 @@ def _update_model_kwargs_for_generation(
         return model_kwargs
 
     @staticmethod
-    def _reorder_cache(past: Tuple[torch.Tensor], beam_idx: torch.Tensor) -> Tuple[torch.Tensor]:
+    # def _reorder_cache(past: Tuple[torch.Tensor], beam_idx: torch.Tensor) -> Tuple[torch.Tensor]:
+    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
         """
         This function is used to re-order the :obj:`past_key_values` or :obj:`mems` cache if
         :meth:`~transformers.PretrainedModel.beam_search` or :meth:`~transformers.PretrainedModel.beam_sample` is
@@ -514,7 +515,11 @@ def _reorder_cache(past: Tuple[torch.Tensor], beam_idx: torch.Tensor) -> Tuple[t
         For custom re-ordering of :obj:`past_key_values` or :obj:`mems`, the function should be implemented in
         subclasses of :class:`~transformers.PreTrainedModel`.
         """
-        return tuple(layer_past.index_select(1, beam_idx.to(layer_past.device)) for layer_past in past)
+        # return tuple(layer_past.index_select(1, beam_idx.to(layer_past.device)) for layer_past in past)
+        return tuple(
+            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
+            for layer_past in past
+        )
 
     def _get_logits_warper(
         self, top_k: int = None, top_p: float = None, temperature: float = None, num_beams: int = None

src/transformers/models/gpt2/modeling_gpt2.py

@@ -17,7 +17,7 @@
 
 import os
 from dataclasses import dataclass
-from typing import List, Optional, Tuple
+from typing import Optional, Tuple
 
 import torch
 import torch.nn as nn
@@ -232,7 +232,7 @@ def forward(
             value = torch.cat((past_value, value), dim=-2)
 
         if use_cache is True:
-            present = torch.stack((key.transpose(-2, -1), value))  # transpose to have same shapes for stacking
+            present = (key.transpose(-2, -1), value)  # transpose to have same shapes
         else:
             present = None
 
@@ -369,9 +369,9 @@ class GPT2DoubleHeadsModelOutput(ModelOutput):
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         mc_logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_choices)`):
             Prediction scores of the multiple choice classification head (scores for each choice before SoftMax).
-        past_key_values (:obj:`List[torch.FloatTensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
-            List of :obj:`torch.FloatTensor` of length :obj:`config.n_layers`, with each tensor of shape :obj:`(2,
-            batch_size, num_heads, sequence_length, embed_size_per_head)`).
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            :obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers`, with each tuple having 2 tuples each of
+            which has a tensor of shape :obj:`(batch_size, num_heads, sequence_length, embed_size_per_head)`).
 
             Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
             :obj:`past_key_values` input) to speed up sequential decoding.
@@ -392,7 +392,7 @@ class GPT2DoubleHeadsModelOutput(ModelOutput):
     mc_loss: Optional[torch.FloatTensor] = None
     logits: torch.FloatTensor = None
     mc_logits: torch.FloatTensor = None
-    past_key_values: Optional[List[torch.FloatTensor]] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
     hidden_states: Optional[Tuple[torch.FloatTensor]] = None
     attentions: Optional[Tuple[torch.FloatTensor]] = None
 
@@ -418,7 +418,7 @@ class GPT2DoubleHeadsModelOutput(ModelOutput):
     Args:
         input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
             :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
-            ``past_key_values[0].shape[-2]`` (``sequence_length`` of input past key value states). Indices of input
+            ``past_key_values[0][0].shape[-2]`` (``sequence_length`` of input past key value states). Indices of input
             sequence tokens in the vocabulary.
 
             If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be
@@ -429,7 +429,7 @@ class GPT2DoubleHeadsModelOutput(ModelOutput):
             details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        past_key_values (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
+        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers`):
             Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
             :obj:`past_key_values` output below). Can be used to speed up sequential decoding. The ``input_ids`` which
             have their past given to this model should not be passed as ``input_ids`` as they have already been
@@ -639,7 +639,7 @@ def forward(
 
         if past_key_values is None:
             past_length = 0
-            past_key_values = [None] * len(self.h)
+            past_key_values = tuple([None] * len(self.h))
         else:
             past_length = past_key_values[0][0].size(-2)
         if position_ids is None:
@@ -707,7 +707,7 @@ def forward(
                 torch.cuda.set_device(hidden_states.device)
                 # Ensure layer_past is on same device as hidden_states (might not be correct)
                 if layer_past is not None:
-                    layer_past = layer_past.to(hidden_states.device)
+                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
                 # Ensure that attention_mask is always on the same device as hidden_states
                 if attention_mask is not None:
                     attention_mask = attention_mask.to(hidden_states.device)