Contrastive Search peak memory reduction (#24120)

Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

src/transformers/generation/configuration_utils.py

@@ -189,6 +189,9 @@ class GenerationConfig(PushToHubMixin):
             The guidance scale for classifier free guidance (CFG). CFG is enabled by setting `guidance_scale > 1`.
             Higher guidance scale encourages the model to generate samples that are more closely linked to the input
             prompt, usually at the expense of poorer quality.
+        low_memory (`bool`, *optional*):
+            Switch to sequential topk for contrastive search to reduce peak memory. Used with contrastive search.
+
 
         > Parameters that define the output variables of `generate`
 
@@ -270,6 +273,7 @@ class GenerationConfig(PushToHubMixin):
         self.forced_decoder_ids = kwargs.pop("forced_decoder_ids", None)
         self.sequence_bias = kwargs.pop("sequence_bias", None)
         self.guidance_scale = kwargs.pop("guidance_scale", None)
+        self.low_memory = kwargs.pop("low_memory", None)
 
         # Parameters that define the output variables of `generate`
         self.num_return_sequences = kwargs.pop("num_return_sequences", 1)

src/transformers/generation/utils.py

@@ -1569,6 +1569,7 @@ class GenerationMixin:
                 return_dict_in_generate=generation_config.return_dict_in_generate,
                 synced_gpus=synced_gpus,
                 streamer=streamer,
+                sequential=generation_config.low_memory,
                 **model_kwargs,
             )
 
@@ -1832,6 +1833,7 @@ class GenerationMixin:
         return_dict_in_generate: Optional[bool] = None,
         synced_gpus: bool = False,
         streamer: Optional["BaseStreamer"] = None,
+        sequential: Optional[bool] = None,
         **model_kwargs,
     ) -> Union[ContrastiveSearchOutput, torch.LongTensor]:
         r"""
@@ -1882,6 +1884,8 @@ class GenerationMixin:
             streamer (`BaseStreamer`, *optional*):
                 Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                 through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
+            sequential (`bool`, *optional*):
+                Switches topk hidden state computation from parallel to sequential to reduce memory if True.
             model_kwargs:
                 Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
                 If model is an encoder-decoder model the kwargs should include `encoder_outputs`.
@@ -1921,6 +1925,7 @@ class GenerationMixin:
         stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
         pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
         eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
+        sequential = sequential if sequential is not None else self.generation_config.low_memory
         if isinstance(eos_token_id, int):
             eos_token_id = [eos_token_id]
         eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
@@ -1986,6 +1991,7 @@ class GenerationMixin:
                     last_hidden_states = outputs.decoder_hidden_states[-1]
                 else:
                     last_hidden_states = outputs.hidden_states[-1]
+
                 # next logit for contrastive search to select top-k candidate tokens
                 logit_for_next_step = outputs.logits[:, -1, :]
 
@@ -1995,11 +2001,11 @@ class GenerationMixin:
                     is_encoder_decoder=self.config.is_encoder_decoder,
                     standardize_cache_format=True,
                 )
-
-                # Expands model inputs top_k times, for batched forward passes (akin to beam search).
-                _, model_kwargs = self._expand_inputs_for_generation(
-                    expand_size=top_k, is_encoder_decoder=self.config.is_encoder_decoder, **model_kwargs
-                )
+                if not sequential:
+                    # Expands model inputs top_k times, for batched forward passes (akin to beam search).
+                    _, model_kwargs = self._expand_inputs_for_generation(
+                        expand_size=top_k, is_encoder_decoder=self.config.is_encoder_decoder, **model_kwargs
+                    )
 
                 past_key_values = model_kwargs.get("past_key_values")
                 if past_key_values is None:
@@ -2019,7 +2025,6 @@ class GenerationMixin:
             # contrastive_search main logic start:
             # contrastive search decoding consists of two steps: (1) candidate tokens recall; (2) candidate re-rank by
             # degeneration penalty
-
             logit_for_next_step = logits_processor(input_ids, logit_for_next_step)
             logit_for_next_step = logits_warper(input_ids, logit_for_next_step)
             next_probs = nn.functional.softmax(logit_for_next_step, dim=-1)
@@ -2049,25 +2054,74 @@ class GenerationMixin:
                 items = []
                 # item is either the key or the value matrix
                 for item in layer:
-                    items.append(item.repeat_interleave(top_k, dim=0))
+                    if sequential:
+                        items.append(item.repeat_interleave(1, dim=0))
+                    else:
+                        items.append(item.repeat_interleave(top_k, dim=0))
                 new_key_values.append(items)
             model_kwargs["past_key_values"] = new_key_values
 
-            # compute the candidate tokens by the language model and collects their hidden_states
-            next_model_inputs = self.prepare_inputs_for_generation(top_k_ids.view(-1, 1), **model_kwargs)
-            outputs = self(
-                **next_model_inputs, return_dict=True, output_hidden_states=True, output_attentions=output_attentions
-            )
-            next_past_key_values = self._extract_past_from_model_output(outputs, standardize_cache_format=True)
+            if sequential:
+                all_outputs = {key: [] for key in outputs}  # defined in first loop iteration
+                all_last_hstates, all_hstates, all_logits = [], [], []
+                for i in range(top_k):
+                    # compute the candidate tokens by the language model and collect their hidden_states
+                    next_model_inputs = self.prepare_inputs_for_generation(top_k_ids[:, i].view(-1, 1), **model_kwargs)
+
+                    outputs = self(
+                        **next_model_inputs,
+                        return_dict=True,
+                        output_hidden_states=True,
+                        output_attentions=output_attentions,
+                    )
+                    for key in all_outputs:
+                        all_outputs[key].append(outputs[key])
+
+                    if self.config.is_encoder_decoder:
+                        next_hidden = outputs.decoder_hidden_states[-1]
+                        full_hidden_states = outputs.decoder_hidden_states
+
+                    else:
+                        next_hidden = outputs.hidden_states[-1]
+                        full_hidden_states = outputs.hidden_states
+
+                    all_last_hstates.append(torch.squeeze(next_hidden, 0))
+                    all_hstates.append(full_hidden_states)
+                    all_logits.append(outputs.logits[:, -1, :])
+
+                # stack hidden states
+                next_hidden = torch.stack([all_last_hstates[i] for i in range(top_k)], dim=0)
+                final_full_hstates = [0 for i in range(len(full_hidden_states))]
+                for layer in range(len(full_hidden_states)):
+                    final_full_hstates[layer] = torch.stack(
+                        [torch.squeeze(all_hstates[i][layer], 0) for i in range(top_k)], dim=0
+                    )
+                full_hidden_states = tuple(final_full_hstates)
+
+                # stack logits
+                logits = torch.cat(all_logits, dim=0)
 
-            logits = outputs.logits[:, -1, :]
-            # name is different for encoder-decoder and decoder-only models
-            if self.config.is_encoder_decoder:
-                next_hidden = outputs.decoder_hidden_states[-1]
-                full_hidden_states = outputs.decoder_hidden_states
             else:
-                next_hidden = outputs.hidden_states[-1]
-                full_hidden_states = outputs.hidden_states
+                # compute the candidate tokens by the language model and collect their hidden_states
+                # assembles top_k_ids into batch of size k
+                next_model_inputs = self.prepare_inputs_for_generation(top_k_ids.view(-1, 1), **model_kwargs)
+
+                outputs = self(
+                    **next_model_inputs,
+                    return_dict=True,
+                    output_hidden_states=True,
+                    output_attentions=output_attentions,
+                )
+                # name is different for encoder-decoder and decoder-only models
+                if self.config.is_encoder_decoder:
+                    next_hidden = outputs.decoder_hidden_states[-1]
+                    full_hidden_states = outputs.decoder_hidden_states
+                else:
+                    next_hidden = outputs.hidden_states[-1]
+                    full_hidden_states = outputs.hidden_states
+
+                logits = outputs.logits[:, -1, :]
+
             context_hidden = last_hidden_states.repeat_interleave(top_k, dim=0)
 
             # compute the degeneration penalty and re-rank the candidates based on the degeneration penalty and the
@@ -2089,17 +2143,32 @@ class GenerationMixin:
                 layer = torch.stack(torch.split(layer, top_k))[range(batch_size), selected_idx, :]
                 next_decoder_hidden_states += (layer,)
 
-            # select the past_key_value
-            new_key_values = ()
-            for layer in next_past_key_values:
-                items = ()
-                # item is either the key or the value matrix
-                for item in layer:
-                    item = torch.stack(torch.split(item, top_k, dim=0))  # [B, K, num_head, seq_len, esz]
-                    item = item[range(batch_size), selected_idx, ...]  # [B, num_head, seq_len, esz]
-                    items += (item,)
-                new_key_values += (items,)
-            next_past_key_values = new_key_values
+            # generate past_key_values cache of only the selected token
+            if sequential:
+                next_model_input = self.prepare_inputs_for_generation(
+                    top_k_ids[:, selected_idx].view(-1, 1), **model_kwargs
+                )
+
+                selected_outputs = self(
+                    **next_model_input,
+                    return_dict=True,
+                    output_hidden_states=False,
+                    output_attentions=False,
+                )
+                next_past_key_values = selected_outputs["past_key_values"]
+
+            else:
+                next_past_key_values = self._extract_past_from_model_output(outputs, standardize_cache_format=True)
+                new_key_values = ()
+                for layer in next_past_key_values:
+                    items = ()
+                    # item is either the key or the value matrix
+                    for item in layer:
+                        item = torch.stack(torch.split(item, top_k, dim=0))  # [B, K, num_head, seq_len, esz]
+                        item = item[range(batch_size), selected_idx, ...]  # [B, num_head, seq_len, esz]
+                        items += (item,)
+                    new_key_values += (items,)
+                next_past_key_values = new_key_values
 
             logit_for_next_step = torch.stack(torch.split(logits, top_k))[range(batch_size), selected_idx, :]
 

tests/generation/test_utils.py

@@ -1457,6 +1457,49 @@ class GenerationTesterMixin:
             for output in (output_contrastive, output_generate):
                 self._check_outputs(output, input_ids, model.config, use_cache=True)
 
+    def test_contrastive_generate_low_memory(self):
+        # Check that choosing 'low_memory' does not change the model output
+        for model_class in self.all_generative_model_classes:
+            # won't fix: FSMT, Reformer, gptbigcode, and speech2text have a different cache variable type (and format).
+            if any(
+                model_name in model_class.__name__.lower()
+                for model_name in ["fsmt", "reformer", "gptbigcode", "speech2text"]
+            ):
+                return
+
+            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config(batch_size=1)
+
+            # NOTE: contrastive search only works with cache on at the moment.
+            if not hasattr(config, "use_cache"):
+                return
+
+            config.use_cache = True
+            config.is_decoder = True
+
+            # test output equality of low versus high memory
+            model = model_class(config).to(torch_device).eval()
+
+            low_output = model.generate(
+                input_ids,
+                top_k=4,
+                penalty_alpha=0.6,
+                low_memory=True,
+                max_length=max_length,
+                attention_mask=attention_mask,
+            )
+
+            high_output = model.generate(
+                input_ids,
+                top_k=4,
+                penalty_alpha=0.6,
+                low_memory=False,
+                max_length=max_length,
+                attention_mask=attention_mask,
+            )
+            self.assertListEqual(low_output.tolist(), high_output.tolist())
+
+        return
+
     @slow  # TODO(Joao): remove this. Some models (e.g. data2vec, xcom, roberta) have an error rate between 1 and 10%.
     def test_assisted_decoding_matches_greedy_search(self):
         # This test ensures that the assisted generation does not introduce output changes over greedy search.