Fix llama model sdpa attention forward function masking bug when outp…

…ut_attentions=True (#30652)

* Fix llama model forward function with attention=True, same-length encoded sequence.

* Fix style

* propagate fix to modeling_cohere, gemma, dbrx, and olmo (which copy the same sdpa masking logic from llama)

* Fix style

* ignore unnecessary sdpa mask converter when output_attentions=True

* add tests checking sdpa and eager outputs match when output_attentions=True

* Split if statements in two lines

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Fix formatting

* Add fix to new jetmoe model

* Add missing output_attentions argument to jetmoe mask creation

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/models/cohere/modeling_cohere.py

@@ -889,7 +889,9 @@ def forward(
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
 
-        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -958,6 +960,7 @@ def _update_causal_mask(
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: Cache,
+        output_attentions: bool,
     ):
         # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
         # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
@@ -974,7 +977,9 @@ def _update_causal_mask(
         # to infer the attention mask.
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
-        if self.config._attn_implementation == "sdpa" and not using_static_cache:
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
                 attention_mask,
                 inputs_embeds=input_tensor,
@@ -1020,6 +1025,7 @@ def _update_causal_mask(
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
             and attention_mask.device.type == "cuda"
+            and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.

src/transformers/models/dbrx/modeling_dbrx.py

@@ -1123,7 +1123,10 @@ def forward(
 
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
-        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1204,6 +1207,7 @@ def _update_causal_mask(
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: Cache,
+        output_attentions: bool,
     ):
         # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
         # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
@@ -1220,7 +1224,9 @@ def _update_causal_mask(
         # to infer the attention mask.
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
-        if self.config._attn_implementation == "sdpa" and not using_static_cache:
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
                 attention_mask,
                 inputs_embeds=input_tensor,
@@ -1266,6 +1272,7 @@ def _update_causal_mask(
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
             and attention_mask.device.type == "cuda"
+            and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.

src/transformers/models/gemma/modeling_gemma.py

@@ -873,7 +873,9 @@ def forward(
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
 
-        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -948,6 +950,7 @@ def _update_causal_mask(
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: Cache,
+        output_attentions: bool,
     ):
         # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
         # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
@@ -964,7 +967,9 @@ def _update_causal_mask(
         # to infer the attention mask.
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
-        if self.config._attn_implementation == "sdpa" and not using_static_cache:
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
                 attention_mask,
                 inputs_embeds=input_tensor,
@@ -1008,6 +1013,7 @@ def _update_causal_mask(
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
             and attention_mask.device.type == "cuda"
+            and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.

src/transformers/models/jetmoe/modeling_jetmoe.py

@@ -1103,7 +1103,9 @@ def forward(
                     " this may lead to unexpected behaviour for Flash Attention version of JetMoe. Make sure to "
                     " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                 )
-        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
 
         hidden_states = inputs_embeds
 
@@ -1178,6 +1180,7 @@ def _update_causal_mask(
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: Cache,
+        output_attentions: bool,
     ):
         # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
         # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
@@ -1194,7 +1197,9 @@ def _update_causal_mask(
         # to infer the attention mask.
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
-        if self.config._attn_implementation == "sdpa" and not using_static_cache:
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
                 attention_mask,
                 inputs_embeds=input_tensor,
@@ -1240,6 +1245,7 @@ def _update_causal_mask(
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
             and attention_mask.device.type == "cuda"
+            and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.

src/transformers/models/llama/modeling_llama.py

@@ -967,7 +967,9 @@ def forward(
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
 
-        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1036,6 +1038,7 @@ def _update_causal_mask(
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: Cache,
+        output_attentions: bool,
     ):
         # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
         # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
@@ -1052,7 +1055,9 @@ def _update_causal_mask(
         # to infer the attention mask.
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
-        if self.config._attn_implementation == "sdpa" and not using_static_cache:
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
                 attention_mask,
                 inputs_embeds=input_tensor,
@@ -1098,6 +1103,7 @@ def _update_causal_mask(
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
             and attention_mask.device.type == "cuda"
+            and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.

src/transformers/models/olmo/modeling_olmo.py

@@ -945,7 +945,9 @@ def forward(
         if position_ids is None:
             position_ids = cache_position.unsqueeze(0)
 
-        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1015,6 +1017,7 @@ def _update_causal_mask(
         input_tensor: torch.Tensor,
         cache_position: torch.Tensor,
         past_key_values: Cache,
+        output_attentions: bool,
     ):
         # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
         # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
@@ -1031,7 +1034,9 @@ def _update_causal_mask(
         # to infer the attention mask.
         past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
         using_static_cache = isinstance(past_key_values, StaticCache)
-        if self.config._attn_implementation == "sdpa" and not using_static_cache:
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
                 attention_mask,
                 inputs_embeds=input_tensor,
@@ -1077,6 +1082,7 @@ def _update_causal_mask(
             self.config._attn_implementation == "sdpa"
             and attention_mask is not None
             and attention_mask.device.type == "cuda"
+            and not output_attentions
         ):
             # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
             # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.