Reenable SDPA's FA2 During Training with torch.compile (#30442)

* Reenable SDPA's FA2 during training with torch.compile

* fix Olmo's SDPA FA2 dispatching too

* update formatting

* improved SDPA comment

* formatting and explanatory comment

* is_causal if statement to one-liner

src/transformers/modeling_attn_mask_utils.py

@@ -240,6 +240,7 @@ class AttentionMaskConverter:
         inputs_embeds: torch.Tensor,
         past_key_values_length: int,
         sliding_window: Optional[int] = None,
+        is_training: bool = False,
     ) -> bool:
         """
         Detects whether the optional user-specified attention_mask & the automatically created causal mask can be ignored in case PyTorch's SDPA is used, rather relying on SDPA's `is_causal` argument.
@@ -263,11 +264,11 @@ class AttentionMaskConverter:
         if attention_mask is None:
             # TODO: When tracing with TorchDynamo with fullgraph=True, the model is recompiled depending on the input shape, thus SDPA's `is_causal` argument is rightfully updated (see https://gist.github.com/fxmarty/1313f39037fc1c112508989628c57363). However, when using `torch.export` or
             # or `torch.onnx.dynamo_export`, we must pass an example input, and `is_causal` behavior is hard-coded. If a user exports a model with q_len > 1, the exported model will hard-code `is_causal=True` which is in general wrong (see https://github.com/pytorch/pytorch/issues/108108).
-            # Thus, we currently can NOT set `ignore_causal_mask = True` here. We would need a `torch._dynamo.is_exporting()` flag.
+            # Thus, we only set `ignore_causal_mask = True` if the model is set to training.
             #
             # Besides, jit.trace can not handle the `q_len > 1` condition for `is_causal` (`TypeError: scaled_dot_product_attention(): argument 'is_causal' must be bool, not Tensor`).
             if (
-                not is_tracing
+                (is_training or not is_tracing)
                 and (query_length == 1 or key_value_length == query_length)
                 and (sliding_window is None or key_value_length < sliding_window)
             ):
@@ -279,7 +280,7 @@ class AttentionMaskConverter:
                     raise ValueError(
                         f"Incorrect 4D attention_mask shape: {tuple(attention_mask.shape)}; expected: {expected_shape}."
                     )
-            elif not is_tracing and torch.all(attention_mask == 1):
+            elif (is_training or not is_tracing) and torch.all(attention_mask == 1):
                 if query_length == 1 or key_value_length == query_length:
                     # For query_length == 1, causal attention and bi-directional attention are the same.
                     ignore_causal_mask = True

src/transformers/models/cohere/modeling_cohere.py

@@ -590,15 +590,17 @@ class CohereSdpaAttention(CohereAttention):
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
 
-        # In case we are not compiling, we may set `causal_mask` to None, which is required to dispatch to SDPA's Flash Attention 2 backend, rather
-        # relying on the `is_causal` argument.
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this if statement instead of an
+        # inline conditional assignment to support both torch.compile's `dynamic=True` and `fullgraph=True`
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query_states,
             key_states,
             value_states,
             attn_mask=causal_mask,
             dropout_p=self.attention_dropout if self.training else 0.0,
-            is_causal=causal_mask is None and q_len > 1,
+            is_causal=is_causal,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -996,7 +998,10 @@ class CohereModel(CoherePreTrainedModel):
             # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument,
             # in order to dispatch on Flash Attention 2.
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
-                attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
             ):
                 return None
 

src/transformers/models/gemma/modeling_gemma.py

@@ -570,15 +570,17 @@ class GemmaSdpaAttention(GemmaAttention):
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
 
-        # In case we are not compiling, we may set `causal_mask` to None, which is required to dispatch to SDPA's Flash Attention 2 backend, rather
-        # relying on the `is_causal` argument.
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this if statement instead of an
+        # inline conditional assignment to support both torch.compile's `dynamic=True` and `fullgraph=True`
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query_states,
             key_states,
             value_states,
             attn_mask=causal_mask,
             dropout_p=self.attention_dropout if self.training else 0.0,
-            is_causal=causal_mask is None and q_len > 1,
+            is_causal=is_causal,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -982,7 +984,10 @@ class GemmaModel(GemmaPreTrainedModel):
             # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument,
             # in order to dispatch on Flash Attention 2.
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
-                attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
             ):
                 return None
 

src/transformers/models/llama/modeling_llama.py

@@ -666,15 +666,17 @@ class LlamaSdpaAttention(LlamaAttention):
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
 
-        # In case we are not compiling, we may set `causal_mask` to None, which is required to dispatch to SDPA's Flash Attention 2 backend, rather
-        # relying on the `is_causal` argument.
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this if statement instead of an
+        # inline conditional assignment to support both torch.compile's `dynamic=True` and `fullgraph=True`
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query_states,
             key_states,
             value_states,
             attn_mask=causal_mask,
             dropout_p=self.attention_dropout if self.training else 0.0,
-            is_causal=causal_mask is None and q_len > 1,
+            is_causal=is_causal,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -1074,7 +1076,10 @@ class LlamaModel(LlamaPreTrainedModel):
             # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument,
             # in order to dispatch on Flash Attention 2.
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
-                attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
             ):
                 return None
 

src/transformers/models/olmo/modeling_olmo.py

@@ -647,13 +647,17 @@ class OlmoSdpaAttention(OlmoAttention):
             key_states = key_states.contiguous()
             value_states = value_states.contiguous()
 
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this if statement instead of an
+        # inline conditional assignment to support both torch.compile's `dynamic=True` and `fullgraph=True`
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
         attn_output = torch.nn.functional.scaled_dot_product_attention(
             query_states,
             key_states,
             value_states,
             attn_mask=causal_mask,
             dropout_p=self.attention_dropout if self.training else 0.0,
-            is_causal=causal_mask is None and q_len > 1,
+            is_causal=is_causal,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -1057,7 +1061,10 @@ class OlmoModel(OlmoPreTrainedModel):
             # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument,
             # in order to dispatch on Flash Attention 2.
             if AttentionMaskConverter._ignore_causal_mask_sdpa(
-                attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
             ):
                 return None