Generate: SinkCache can handle iterative prompts (#27907)

src/transformers/cache_utils.py

@@ -38,6 +38,21 @@ class Cache:
         """Returns the sequence length of the cached states. A layer index can be optionally passed."""
         raise NotImplementedError("Make sure to implement `get_seq_length` in a subclass.")
 
+    def get_max_length(self) -> Optional[int]:
+        """Returns the maximum sequence length of the cached states, if there is any."""
+        raise NotImplementedError("Make sure to implement `get_max_length` in a subclass.")
+
+    def get_usable_length(self, new_seq_length: int, layer_idx: Optional[int] = 0) -> int:
+        """Given the sequence length of the new inputs, returns the usable length of the cache."""
+        # Cache without size limit -> all cache is usable
+        # Cache with size limit -> if the length cache plus the length of the new inputs is larger the maximum cache
+        #   length, we will need to evict part of the cache (and thus not all cache is usable)
+        max_length = self.get_max_length()
+        previous_seq_length = self.get_seq_length(layer_idx)
+        if max_length is not None and previous_seq_length + new_seq_length > max_length:
+            return max_length - new_seq_length
+        return previous_seq_length
+
 
 class DynamicCache(Cache):
     """
@@ -120,6 +135,10 @@ class DynamicCache(Cache):
             return 0
         return self.key_cache[layer_idx].shape[-2]
 
+    def get_max_length(self) -> Optional[int]:
+        """Returns the maximum sequence length of the cached states. DynamicCache does not have a maximum length."""
+        return None
+
     def reorder_cache(self, beam_idx: torch.LongTensor):
         """Reorders the cache for beam search, given the selected beam indices."""
         for layer_idx in range(len(self.key_cache)):
@@ -209,8 +228,11 @@ class SinkCache(Cache):
         # Workaround to make 'key_states.shape[-2] + past_key_value.get_seq_length(self.layer_idx)' <= window_length
         if len(self.key_cache) <= layer_idx:
             return 0
-        cache_length = self.key_cache[layer_idx].shape[-2]
+        return self.key_cache[layer_idx].shape[-2]
-        return min(cache_length, self.window_length - 1)
+
+    def get_max_length(self) -> Optional[int]:
+        """Returns the maximum sequence length of the cached states."""
+        return self.window_length
 
     def update(
         self,
@@ -267,7 +289,9 @@ class SinkCache(Cache):
 
             # On RoPE models, we need to recompute the Key rotation as the tokens are shifted
             if using_rope:
-                rerotation_cos, rerotation_sin = self._get_rerotation_cos_sin(key_states, cos, sin)
+                rerotation_cos, rerotation_sin = self._get_rerotation_cos_sin(
+                    key_states, cos[: self.window_length], sin[: self.window_length]
+                )
                 if partial_rotation_size is not None:
                     keys_to_keep, keys_pass = (
                         keys_to_keep[..., :partial_rotation_size],

src/transformers/models/llama/modeling_llama.py

@@ -398,7 +398,7 @@ class LlamaAttention(nn.Module):
                     "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                     "with a layer index."
                 )
-            kv_seq_len += past_key_value.get_seq_length(self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
 
@@ -503,7 +503,7 @@ class LlamaFlashAttention2(LlamaAttention):
 
         kv_seq_len = key_states.shape[-2]
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_seq_length(self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
@@ -910,7 +910,7 @@ class LlamaModel(LlamaPreTrainedModel):
             use_legacy_cache = not isinstance(past_key_values, Cache)
             if use_legacy_cache:
                 past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-            past_key_values_length = past_key_values.get_seq_length()
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
 
         if position_ids is None:
             device = input_ids.device if input_ids is not None else inputs_embeds.device
@@ -1127,8 +1127,10 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
             if isinstance(past_key_values, Cache):
                 cache_length = past_key_values.get_seq_length()
                 past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
             else:
                 cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
 
             # Keep only the unprocessed tokens:
             # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
@@ -1142,10 +1144,13 @@ class LlamaForCausalLM(LlamaPreTrainedModel):
                 input_ids = input_ids[:, past_length:]
             # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
 
-            # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
-            # older attention values, as their corresponding values are not part of the input.
+            if (
-            if cache_length < past_length and attention_mask is not None:
+                max_cache_length is not None
-                attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
 
         position_ids = kwargs.get("position_ids", None)
         if attention_mask is not None and position_ids is None:

src/transformers/models/mistral/modeling_mistral.py

@@ -268,7 +268,7 @@ class MistralAttention(nn.Module):
                     "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                     "with a layer index."
                 )
-            kv_seq_len += past_key_value.get_seq_length(self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
 
@@ -363,7 +363,7 @@ class MistralFlashAttention2(MistralAttention):
 
         kv_seq_len = key_states.shape[-2]
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_seq_length(self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
 
         # Because the input can be padded, the absolute sequence length depends on the max position id.
         rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
@@ -850,15 +850,13 @@ class MistralModel(MistralPreTrainedModel):
         else:
             raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
 
-        seq_length_with_past = seq_length
         past_key_values_length = 0
 
         if use_cache:
             use_legacy_cache = not isinstance(past_key_values, Cache)
             if use_legacy_cache:
                 past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-            past_key_values_length = past_key_values.get_seq_length()
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
-            seq_length_with_past = seq_length_with_past + past_key_values_length
 
         if position_ids is None:
             device = input_ids.device if input_ids is not None else inputs_embeds.device
@@ -1092,8 +1090,10 @@ class MistralForCausalLM(MistralPreTrainedModel):
             if isinstance(past_key_values, Cache):
                 cache_length = past_key_values.get_seq_length()
                 past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
             else:
                 cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
 
             # Keep only the unprocessed tokens:
             # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
@@ -1107,10 +1107,13 @@ class MistralForCausalLM(MistralPreTrainedModel):
                 input_ids = input_ids[:, past_length:]
             # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
 
-            # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
-            # older attention values, as their corresponding values are not part of the input.
+            if (
-            if cache_length < past_length and attention_mask is not None:
+                max_cache_length is not None
-                attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
 
         position_ids = kwargs.get("position_ids", None)
         if attention_mask is not None and position_ids is None:

src/transformers/models/persimmon/modeling_persimmon.py

@@ -295,7 +295,7 @@ class PersimmonAttention(nn.Module):
                     "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                     "with a layer index."
                 )
-            kv_seq_len += past_key_value.get_seq_length(self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         # Partial rotary embedding
@@ -612,7 +612,7 @@ class PersimmonModel(PersimmonPreTrainedModel):
             use_legacy_cache = not isinstance(past_key_values, Cache)
             if use_legacy_cache:
                 past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-            past_key_values_length = past_key_values.get_seq_length()
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
             seq_length_with_past = seq_length_with_past + past_key_values_length
 
         if position_ids is None:
@@ -831,8 +831,10 @@ class PersimmonForCausalLM(PersimmonPreTrainedModel):
             if isinstance(past_key_values, Cache):
                 cache_length = past_key_values.get_seq_length()
                 past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
             else:
                 cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
 
             # Keep only the unprocessed tokens:
             # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
@@ -846,10 +848,13 @@ class PersimmonForCausalLM(PersimmonPreTrainedModel):
                 input_ids = input_ids[:, past_length:]
             # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
 
-            # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
-            # older attention values, as their corresponding values are not part of the input.
+            if (
-            if cache_length < past_length and attention_mask is not None:
+                max_cache_length is not None
-                attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
 
         position_ids = kwargs.get("position_ids", None)
         if attention_mask is not None and position_ids is None:

src/transformers/models/phi/modeling_phi.py

@@ -334,7 +334,7 @@ class PhiAttention(nn.Module):
                     "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                     "with a layer index."
                 )
-            kv_seq_len += past_key_value.get_seq_length(self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         # Partial rotary embedding
@@ -444,7 +444,7 @@ class PhiFlashAttention2(PhiAttention):
 
         kv_seq_len = key_states.shape[-2]
         if past_key_value is not None:
-            kv_seq_len += past_key_value.get_seq_length(self.layer_idx)
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         # Partial rotary embedding
@@ -855,15 +855,13 @@ class PhiModel(PhiPreTrainedModel):
         else:
             raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
 
-        seq_length_with_past = seq_length
         past_key_values_length = 0
 
         if use_cache:
             use_legacy_cache = not isinstance(past_key_values, Cache)
             if use_legacy_cache:
                 past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-            past_key_values_length = past_key_values.get_seq_length()
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
-            seq_length_with_past = seq_length_with_past + past_key_values_length
 
         if position_ids is None:
             device = input_ids.device if input_ids is not None else inputs_embeds.device
@@ -1085,8 +1083,10 @@ class PhiForCausalLM(PhiPreTrainedModel):
             if isinstance(past_key_values, Cache):
                 cache_length = past_key_values.get_seq_length()
                 past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
             else:
                 cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
 
             # Keep only the unprocessed tokens:
             # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
@@ -1100,10 +1100,13 @@ class PhiForCausalLM(PhiPreTrainedModel):
                 input_ids = input_ids[:, past_length:]
             # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
 
-            # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
-            # older attention values, as their corresponding values are not part of the input.
+            if (
-            if cache_length < past_length and attention_mask is not None:
+                max_cache_length is not None
-                attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
 
         position_ids = kwargs.get("position_ids", None)
         if attention_mask is not None and position_ids is None:

tests/test_cache_utils.py

@@ -187,3 +187,45 @@ class CacheIntegrationTest(unittest.TestCase):
         gen_out = model.generate(**inputs, do_sample=False, max_new_tokens=3000, past_key_values=cache)
         decoded = tokenizer.batch_decode(gen_out, skip_special_tokens=True)
         self.assertTrue(decoded[0].endswith("to perform a variety of tasks. The Transformer is a neural network"))
+
+    def test_sink_cache_iterative_prompts(self):
+        """Tests that SinkCache supports more than one new token at once, when shifting the cache"""
+        tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
+        model = AutoModelForCausalLM.from_pretrained(
+            "HuggingFaceH4/zephyr-7b-beta", device_map="auto", torch_dtype=torch.float16
+        )
+        prompt = (
+            "Compose an engaging travel blog post about a recent trip to Hawaii, highlighting cultural experiences "
+            "and must-see attractions."
+        )
+
+        # Prepare generation settings
+        cache = SinkCache(window_length=256, num_sink_tokens=4)
+        input_ids = torch.tensor([], device=model.device, dtype=torch.int)
+        for _ in range(3):
+            # Tokenize the prompt with the correct chat template
+            chat = [{"role": "user", "content": prompt}]
+            tokenized_chat = tokenizer.apply_chat_template(chat, return_tensors="pt", add_generation_prompt=True).to(
+                model.device
+            )
+            input_ids = torch.cat((input_ids, tokenized_chat), dim=1)
+
+            # Perform the generation
+            gen_out = model.generate(
+                input_ids, do_sample=False, max_new_tokens=100, past_key_values=cache, use_cache=True
+            )
+            input_ids = gen_out
+
+        # We went well beyond the cache length
+        self.assertTrue(input_ids.shape[1] > cache.get_max_length() * 1.5)
+
+        # And it still produces a coherent english
+        decoded = tokenizer.batch_decode(input_ids, skip_special_tokens=True)
+        last_output = (
+            "<|assistant|>\nAs the sun began to set over the Pacific Ocean, I found myself standing on the shores of "
+            "Waikiki Beach, my heart filled with awe and wonder. I had just returned from a two-week journey to the "
+            "beautiful island of Hawaii, and it had been an unforgettable experience filled with cultural experiences "
+            "and must-see attractions that left me breathless.\n\nOne of the most memorable experiences of my trip "
+            "was visiting the historic district of Honolulu. Here,"
+        )
+        self.assertTrue(decoded[0].endswith(last_output))