Re-enable SDPA's FA2 path (#30070)

* tentatively re-enable FA2 + SDPA

* better comment

* _ignore_causal_mask_sdpa as staticmethod

* type hints

* use past_seen_tokens instead

* enable copied from for sdpa

* ruff

* llama simplifications on review

* remove unnecessary self.is_causal check

* fix copies

* cleaning

* precise message

* better doc

* add test

* simplify

* Update src/transformers/models/llama/modeling_llama.py

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

* Update src/transformers/models/llama/modeling_llama.py

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

* Update src/transformers/models/llama/modeling_llama.py

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

* style

---------

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

src/transformers/modeling_attn_mask_utils.py

@@ -234,6 +234,59 @@ def _unmask_unattended(
 
         return expanded_mask.mul(~torch.all(expanded_mask == min_dtype, dim=-1, keepdim=True))
 
+    @staticmethod
+    def _ignore_causal_mask_sdpa(
+        attention_mask: Optional[torch.Tensor],
+        inputs_embeds: torch.Tensor,
+        past_key_values_length: int,
+        sliding_window: Optional[int] = None,
+    ) -> 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.
+
+        In case no token is masked in the `attention_mask` argument, if `query_length == 1` or
+        `key_value_length == query_length`, we rather rely on SDPA `is_causal` argument to use causal/non-causal masks,
+        allowing to dispatch to the flash attention kernel (that can otherwise not be used if a custom `attn_mask` is passed).
+        """
+
+        batch_size, query_length = inputs_embeds.shape[0], inputs_embeds.shape[1]
+        key_value_length = query_length + past_key_values_length
+
+        is_tracing = (
+            torch.jit.is_tracing()
+            or isinstance(inputs_embeds, torch.fx.Proxy)
+            or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
+        )
+
+        ignore_causal_mask = False
+
+        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.
+            #
+            # 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 sliding_window is None or key_value_length < sliding_window:
+                ignore_causal_mask = not is_tracing
+        elif sliding_window is None or key_value_length < sliding_window:
+            if len(attention_mask.shape) == 4:
+                expected_shape = (batch_size, 1, query_length, key_value_length)
+                if tuple(attention_mask.shape) != expected_shape:
+                    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):
+                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
+
+                # Unfortunately, for query_length > 1 and key_value_length != query_length, we cannot generally ignore the attention mask, as SDPA causal mask generation
+                # may be wrong. We will set `is_causal=False` in SDPA and rely on Transformers attention_mask instead, hence not setting it to None here.
+                # Reference: https://github.com/pytorch/pytorch/issues/108108
+                # TODO: maybe revisit this with https://github.com/pytorch/pytorch/pull/114823 in PyTorch 2.3.
+
+        return ignore_causal_mask
+
 
 def _prepare_4d_causal_attention_mask(
     attention_mask: Optional[torch.Tensor],
@@ -305,7 +358,6 @@ def _prepare_4d_causal_attention_mask_for_sdpa(
     attn_mask_converter = AttentionMaskConverter(is_causal=True, sliding_window=sliding_window)
 
     key_value_length = input_shape[-1] + past_key_values_length
-    _, query_length = input_shape
 
     # torch.jit.trace, symbolic_trace and torchdynamo with fullgraph=True are unable to capture the controlflow `is_causal=attention_mask is None and q_len > 1`
     # used as an SDPA argument. We keep compatibility with these tracing tools by always using SDPA's `attn_mask` argument in case we are tracing.
@@ -316,37 +368,12 @@ def _prepare_4d_causal_attention_mask_for_sdpa(
         or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
     )
 
-    ignore_causal_mask = False
-
-    if attention_mask is None:
-        if sliding_window is None or key_value_length < sliding_window:
-            ignore_causal_mask = not is_tracing
-    elif sliding_window is None or key_value_length < sliding_window:
-        # 4d mask is passed through
-        if len(attention_mask.shape) == 4:
-            expected_shape = (input_shape[0], 1, input_shape[1], key_value_length)
-            if tuple(attention_mask.shape) != expected_shape:
-                raise ValueError(
-                    f"Incorrect 4D attention_mask shape: {tuple(attention_mask.shape)}; expected: {expected_shape}."
-                )
-            else:
-                # if the 4D mask has correct shape - invert it and fill with negative infinity
-                inverted_mask = 1.0 - attention_mask.to(inputs_embeds.dtype)
-                attention_mask = inverted_mask.masked_fill(
-                    inverted_mask.to(torch.bool), torch.finfo(inputs_embeds.dtype).min
-                )
-                return attention_mask
-
-        elif not is_tracing and torch.all(attention_mask == 1):
-            if query_length == 1:
-                # For query_length == 1, causal attention and bi-directional attention are the same.
-                ignore_causal_mask = True
-            elif key_value_length == query_length:
-                ignore_causal_mask = True
-
-            # Unfortunately, for query_length > 1 and key_value_length != query_length, we cannot generally ignore the attention mask, as SDPA causal mask generation
-            # may be wrong. We will set `is_causal=False` in SDPA and rely on Transformers attention_mask instead, hence not setting it to None here.
-            # Reference: https://github.com/pytorch/pytorch/issues/108108
+    ignore_causal_mask = AttentionMaskConverter._ignore_causal_mask_sdpa(
+        attention_mask=attention_mask,
+        inputs_embeds=inputs_embeds,
+        past_key_values_length=past_key_values_length,
+        sliding_window=sliding_window,
+    )
 
     if ignore_causal_mask:
         expanded_4d_mask = None

src/transformers/models/cohere/modeling_cohere.py

@@ -590,12 +590,15 @@ def forward(
             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.
         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,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -908,9 +911,7 @@ 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_seen_tokens + inputs_embeds.shape[1]
-        )
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_seen_tokens)
 
         # embed positions
         hidden_states = inputs_embeds
@@ -974,24 +975,41 @@ def forward(
             attentions=all_self_attns,
         )
 
-    # 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.
-    # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
-    # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
-    def _update_causal_mask(self, attention_mask, input_tensor, cache_position, current_length):
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_seen_tokens: int,
+    ):
+        # 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.
+        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
         if self.config._attn_implementation == "flash_attention_2":
             if attention_mask is not None and 0.0 in attention_mask:
                 return attention_mask
             return None
 
+        if self.config._attn_implementation == "sdpa":
+            # 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
+            ):
+                return None
+
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
         if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"):  # static cache
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_length + 1
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
             )
 
         causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)

src/transformers/models/gemma/modeling_gemma.py

@@ -570,12 +570,15 @@ def forward(
             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.
         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,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -888,9 +891,7 @@ 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_seen_tokens + inputs_embeds.shape[1]
-        )
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_seen_tokens)
 
         # embed positions
         hidden_states = inputs_embeds
@@ -960,24 +961,41 @@ def forward(
             attentions=all_self_attns,
         )
 
-    # 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.
-    # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
-    # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
-    def _update_causal_mask(self, attention_mask, input_tensor, cache_position, current_length):
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_seen_tokens: int,
+    ):
+        # 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.
+        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
         if self.config._attn_implementation == "flash_attention_2":
             if attention_mask is not None and 0.0 in attention_mask:
                 return attention_mask
             return None
 
+        if self.config._attn_implementation == "sdpa":
+            # 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
+            ):
+                return None
+
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
         if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"):  # static cache
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_length + 1
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
             )
 
         causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)

src/transformers/models/llama/modeling_llama.py

@@ -656,7 +656,6 @@ def forward(
         value_states = repeat_kv(value_states, self.num_key_value_groups)
 
         causal_mask = attention_mask
-        # if attention_mask is not None and cache_position is not None:
         if attention_mask is not None:
             causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
 
@@ -667,12 +666,15 @@ def forward(
             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.
         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,
         )
 
         attn_output = attn_output.transpose(1, 2).contiguous()
@@ -987,9 +989,7 @@ 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_seen_tokens + inputs_embeds.shape[1]
-        )
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_seen_tokens)
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1053,24 +1053,41 @@ def forward(
             attentions=all_self_attns,
         )
 
-    # 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.
-    # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
-    # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
-    def _update_causal_mask(self, attention_mask, input_tensor, cache_position, current_length):
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_seen_tokens: int,
+    ):
+        # 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.
+        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
+        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
+
         if self.config._attn_implementation == "flash_attention_2":
             if attention_mask is not None and 0.0 in attention_mask:
                 return attention_mask
             return None
 
+        if self.config._attn_implementation == "sdpa":
+            # 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
+            ):
+                return None
+
         dtype, device = input_tensor.dtype, input_tensor.device
         min_dtype = torch.finfo(dtype).min
         sequence_length = input_tensor.shape[1]
         if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"):  # static cache
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_length + 1
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
             )
 
         causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)