Fix on "cache position" for assisted generation (#30068)

* clean commit history I hope

* get kv seq length correctly

* PR suggestions

* Update src/transformers/testing_utils.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* add comment

* give gpt bigcode it's own overriden method

* remove code

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

src/transformers/generation/utils.py

@@ -641,6 +641,7 @@ class GenerationMixin:
         model_kwargs: Dict[str, Any],
         is_encoder_decoder: bool = False,
         standardize_cache_format: bool = False,
+        num_new_tokens: int = 1,
     ) -> Dict[str, Any]:
         # update past_key_values
         model_kwargs["past_key_values"] = self._extract_past_from_model_output(
@@ -671,7 +672,7 @@ class GenerationMixin:
                 )
 
         if "cache_position" in model_kwargs and model_kwargs["cache_position"] is not None:
-            model_kwargs["cache_position"] = model_kwargs["cache_position"][-1:] + 1
+            model_kwargs["cache_position"] = model_kwargs["cache_position"][-1:] + num_new_tokens
 
         return model_kwargs
 
@@ -1294,6 +1295,21 @@ class GenerationMixin:
 
         return generation_config, model_kwargs
 
+    def _get_initial_cache_position(self, input_ids, model_kwargs):
+        """Calculates `cache_position` for the pre-fill stage based on `input_ids` and optionally past length"""
+        past_length = 0
+        if "past_key_values" in model_kwargs:
+            if isinstance(model_kwargs["past_key_values"], Cache):
+                past_length = model_kwargs["past_key_values"].get_seq_length()
+            else:
+                past_length = model_kwargs["past_key_values"][0][0].shape[2]
+        if "inputs_embeds" in model_kwargs:
+            cur_len = model_kwargs["inputs_embeds"].shape[1]
+        else:
+            cur_len = input_ids.shape[-1]
+        model_kwargs["cache_position"] = torch.arange(past_length, cur_len, device=input_ids.device)
+        return model_kwargs
+
     @torch.no_grad()
     def generate(
         self,
@@ -1560,6 +1576,8 @@ class GenerationMixin:
                 raise ValueError("assisted generate is only supported for batch_size = 1")
             if not model_kwargs["use_cache"]:
                 raise ValueError("assisted generate requires `use_cache=True`")
+            if generation_config.cache_implementation == "static":
+                raise ValueError("assisted generate is not supported with `static_cache`")
 
             # 11. Get the candidate generator, given the parameterization
             candidate_generator = self._get_candidate_generator(
@@ -2024,11 +2042,9 @@ class GenerationMixin:
             )
 
         # keep track of which sequences are already finished
-        batch_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
+        batch_size = input_ids.shape[0]
         unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=input_ids.device)
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         this_peer_finished = False
 
@@ -2495,12 +2511,10 @@ class GenerationMixin:
             )
 
         # keep track of which sequences are already finished
-        batch_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
+        batch_size = input_ids.shape[0]
         this_peer_finished = False
         unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=input_ids.device)
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         while self._has_unfinished_sequences(this_peer_finished, synced_gpus, device=input_ids.device):
             # prepare model inputs
@@ -2792,12 +2806,10 @@ class GenerationMixin:
             )
 
         # keep track of which sequences are already finished
-        batch_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
+        batch_size = input_ids.shape[0]
         this_peer_finished = False
         unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=input_ids.device)
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         while self._has_unfinished_sequences(this_peer_finished, synced_gpus, device=input_ids.device):
             # prepare model inputs
@@ -3108,9 +3120,7 @@ class GenerationMixin:
         num_beams = beam_scorer.num_beams
 
         batch_beam_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         if num_beams * batch_size != batch_beam_size:
             raise ValueError(
@@ -3514,9 +3524,7 @@ class GenerationMixin:
         num_beams = beam_scorer.num_beams
 
         batch_beam_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         # init attention / hidden states / scores tuples
         scores = () if (return_dict_in_generate and output_scores) else None
@@ -3874,9 +3882,7 @@ class GenerationMixin:
         device = input_ids.device
 
         batch_beam_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         if return_dict_in_generate and output_scores:
             beam_indices = [tuple(() for _ in range(num_sub_beams * batch_size)) for _ in range(num_beam_groups)]
@@ -4292,9 +4298,7 @@ class GenerationMixin:
         num_beams = constrained_beam_scorer.num_beams
 
         batch_beam_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         if num_beams * batch_size != batch_beam_size:
             raise ValueError(
@@ -4655,11 +4659,9 @@ class GenerationMixin:
             )
 
         # keep track of which sequences are already finished
-        batch_size, cur_len = input_ids.shape
-        if "inputs_embeds" in model_kwargs:
-            cur_len = model_kwargs["inputs_embeds"].shape[1]
+        batch_size = input_ids.shape[0]
         unfinished_sequences = torch.ones(batch_size, dtype=torch.long, device=input_ids.device)
-        model_kwargs["cache_position"] = torch.arange(cur_len, device=input_ids.device)
+        model_kwargs = self._get_initial_cache_position(input_ids, model_kwargs)
 
         this_peer_finished = False
         while self._has_unfinished_sequences(this_peer_finished, synced_gpus, device=input_ids.device):
@@ -4679,20 +4681,21 @@ class GenerationMixin:
             # we use this forward pass to also pick the subsequent logits in the original model.
 
             # 2.1. Prepare the model inputs
-            model_kwargs = _prepare_attention_mask(
-                model_kwargs, candidate_input_ids.shape[1], self.config.is_encoder_decoder
+            candidate_kwargs = copy.copy(model_kwargs)
+            candidate_kwargs = _prepare_attention_mask(
+                candidate_kwargs, candidate_input_ids.shape[1], self.config.is_encoder_decoder
             )
-            model_kwargs = _prepare_token_type_ids(model_kwargs, candidate_input_ids.shape[1])
-            if "cache_position" in model_kwargs:
-                model_kwargs["cache_position"] = torch.cat(
+            candidate_kwargs = _prepare_token_type_ids(candidate_kwargs, candidate_input_ids.shape[1])
+            if "cache_position" in candidate_kwargs:
+                candidate_kwargs["cache_position"] = torch.cat(
                     (
-                        model_kwargs["cache_position"],
+                        candidate_kwargs["cache_position"],
                         torch.arange(cur_len, cur_len + candidate_length, device=input_ids.device, dtype=torch.long),
                     ),
                     dim=0,
                 )
 
-            model_inputs = self.prepare_inputs_for_generation(candidate_input_ids, **model_kwargs)
+            model_inputs = self.prepare_inputs_for_generation(candidate_input_ids, **candidate_kwargs)
             if "num_logits_to_keep" in model_inputs:
                 model_inputs["num_logits_to_keep"] = candidate_length + 1
 
@@ -4811,6 +4814,7 @@ class GenerationMixin:
                 outputs,
                 model_kwargs,
                 is_encoder_decoder=self.config.is_encoder_decoder,
+                num_new_tokens=n_matches + 1,
             )
 
             unfinished_sequences = unfinished_sequences & ~stopping_criteria(input_ids, scores)

src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py

@@ -1209,6 +1209,24 @@ class GPTBigCodeForCausalLM(GPTBigCodePreTrainedModel):
         )
         return model_inputs
 
+    def _get_initial_cache_position(self, input_ids, model_kwargs):
+        """
+        Calculates `cache_position` for the pre-fill stage based on `input_ids` and optionally past length.
+        Since gpt bigcode is special, the method is overridden here, other models use it from `generation.utils.py`.
+        """
+        past_length = 0
+        if "past_key_values" in model_kwargs:
+            if self.config.multi_query:
+                past_length = model_kwargs["past_key_values"][0].shape[1]
+            else:
+                past_length = model_kwargs["past_key_values"][0].shape[2]
+        if "inputs_embeds" in model_kwargs:
+            cur_len = model_kwargs["inputs_embeds"].shape[1]
+        else:
+            cur_len = input_ids.shape[-1]
+        model_kwargs["cache_position"] = torch.arange(past_length, cur_len, device=input_ids.device)
+        return model_kwargs
+
     @add_start_docstrings_to_model_forward(GPT_BIGCODE_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
         checkpoint=_CHECKPOINT_FOR_DOC,

src/transformers/models/jamba/modeling_jamba.py

@@ -231,6 +231,7 @@ class HybridMambaAttentionDynamicCache(DynamicCache):
         conv_kernel_size = config.mamba_d_conv
         self.conv_states = []
         self.ssm_states = []
+        self.transformer_layers = []
         for i in range(config.num_hidden_layers):
             if self.layers_block_type[i] == "mamba":
                 self.conv_states += [
@@ -242,6 +243,7 @@ class HybridMambaAttentionDynamicCache(DynamicCache):
             else:
                 self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
                 self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
+                self.transformer_layers.append(i)
 
         self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
         self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
@@ -276,6 +278,14 @@ class HybridMambaAttentionDynamicCache(DynamicCache):
             device = self.ssm_states[layer_idx].device
             self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))
 
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # take any layer that contains cache and not empty tensor
+        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
+        if len(self.key_cache) <= layer_idx:
+            return 0
+        return self.key_cache[layer_idx].shape[-2]
+
     def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
         raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")
 

tests/generation/test_utils.py

@@ -1091,8 +1091,9 @@ class GenerationTesterMixin:
             )
             self.assertListEqual(low_output.tolist(), high_output.tolist())
 
+    @parameterized.expand([("random",), ("same",)])
     @is_flaky()  # Read NOTE (1) below. If there are API issues, all attempts will fail.
-    def test_assisted_decoding_matches_greedy_search(self):
+    def test_assisted_decoding_matches_greedy_search(self, assistant_type):
         # This test ensures that the assisted generation does not introduce output changes over greedy search.
         # NOTE (1): The sentence above is true most of the time, there is a tiny difference in the logits due to matmul
         # shape differences -- and it may result in a different output. The input shape difference happens in the
@@ -1151,6 +1152,12 @@ class GenerationTesterMixin:
             }
             output_greedy = model.generate(input_ids, attention_mask=attention_mask, **generation_kwargs)
 
+            # test with the same assistant model or randomly init one
+            # in the first case all candidate tokens are accepted, in the second none is accepted
+            # case when some are accepted and some not is hard to reproduce, so let's hope this catches most errors :)
+            if assistant_type == "random":
+                assistant_model = model_class(config).to(torch_device).eval()
+            else:
                 assistant_model = model
             assistant_model.generation_config.num_assistant_tokens = 2  # see b)
             assistant_model.generation_config.num_assistant_tokens_schedule = "constant"  # see b)