Generate: Mistral/Mixtral FA2 cache fix when going beyond the context window (#28037)

Author: gante

src/transformers/models/mistral/modeling_mistral.py

@@ -363,6 +363,12 @@ def forward(
 
         kv_seq_len = key_states.shape[-2]
         if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "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_usable_length(kv_seq_len, self.layer_idx)
 
         # Because the input can be padded, the absolute sequence length depends on the max position id.
@@ -385,11 +391,16 @@ def forward(
 
         if past_key_value is not None:
             # Activate slicing cache only if the config has a value `sliding_windows` attribute
-            if getattr(self.config, "sliding_window", None) is not None and kv_seq_len > self.config.sliding_window:
+            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+            if (
+                getattr(self.config, "sliding_window", None) is not None
+                and kv_seq_len > self.config.sliding_window
+                and cache_has_contents
+            ):
                 slicing_tokens = 1 - self.config.sliding_window
 
-                past_key = past_key_value[0]
-                past_value = past_key_value[1]
+                past_key = past_key_value[self.layer_idx][0]
+                past_value = past_key_value[self.layer_idx][1]
 
                 past_key = past_key[:, :, slicing_tokens:, :].contiguous()
                 past_value = past_value[:, :, slicing_tokens:, :].contiguous()
@@ -400,8 +411,6 @@ def forward(
                         f" {past_key.shape}"
                     )
 
-                past_key_value = (past_key, past_value)
-
                 if attention_mask is not None:
                     attention_mask = attention_mask[:, slicing_tokens:]
                     attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)

src/transformers/models/mixtral/modeling_mixtral.py

@@ -414,6 +414,12 @@ def forward(
 
         kv_seq_len = key_states.shape[-2]
         if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "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_usable_length(kv_seq_len, self.layer_idx)
 
         # Because the input can be padded, the absolute sequence length depends on the max position id.
@@ -436,11 +442,16 @@ def forward(
 
         if past_key_value is not None:
             # Activate slicing cache only if the config has a value `sliding_windows` attribute
-            if getattr(self.config, "sliding_window", None) is not None and kv_seq_len > self.config.sliding_window:
+            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+            if (
+                getattr(self.config, "sliding_window", None) is not None
+                and kv_seq_len > self.config.sliding_window
+                and cache_has_contents
+            ):
                 slicing_tokens = 1 - self.config.sliding_window
 
-                past_key = past_key_value[0]
-                past_value = past_key_value[1]
+                past_key = past_key_value[self.layer_idx][0]
+                past_value = past_key_value[self.layer_idx][1]
 
                 past_key = past_key[:, :, slicing_tokens:, :].contiguous()
                 past_value = past_value[:, :, slicing_tokens:, :].contiguous()
@@ -451,8 +462,6 @@ def forward(
                         f" {past_key.shape}"
                     )
 
-                past_key_value = (past_key, past_value)
-
                 if attention_mask is not None:
                     attention_mask = attention_mask[:, slicing_tokens:]
                     attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)