Revert "Re-enable SDPA's FA2 path (#30070)" (#30314)

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

This reverts commit 05bdef16.

* Revert "Fix quality Olmo + SDPA (#30302)"

This reverts commit ec92f983.

src/transformers/modeling_attn_mask_utils.py

@@ -234,59 +234,6 @@ class AttentionMaskConverter:
 
         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],
@@ -358,6 +305,7 @@ 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.
@@ -368,12 +316,37 @@ def _prepare_4d_causal_attention_mask_for_sdpa(
         or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
     )
 
-    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,
-    )
+    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
 
     if ignore_causal_mask:
         expanded_4d_mask = None

src/transformers/models/cohere/modeling_cohere.py

@@ -590,15 +590,12 @@ 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.
         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()
@@ -911,7 +908,9 @@ class CohereModel(CoherePreTrainedModel):
         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)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_seen_tokens + inputs_embeds.shape[1]
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -975,31 +974,16 @@ class CohereModel(CoherePreTrainedModel):
             attentions=all_self_attns,
         )
 
-    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
-
+    # 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):
         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]
@@ -1007,9 +991,7 @@ class CohereModel(CoherePreTrainedModel):
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1]
-                if isinstance(attention_mask, torch.Tensor)
-                else past_seen_tokens + sequence_length + 1
+                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_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

@@ -570,15 +570,12 @@ 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.
         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()
@@ -891,7 +888,9 @@ class GemmaModel(GemmaPreTrainedModel):
         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)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_seen_tokens + inputs_embeds.shape[1]
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -961,31 +960,16 @@ class GemmaModel(GemmaPreTrainedModel):
             attentions=all_self_attns,
         )
 
-    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
-
+    # 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):
         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]
@@ -993,9 +977,7 @@ class GemmaModel(GemmaPreTrainedModel):
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1]
-                if isinstance(attention_mask, torch.Tensor)
-                else past_seen_tokens + sequence_length + 1
+                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_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

@@ -656,6 +656,7 @@ class LlamaSdpaAttention(LlamaAttention):
         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]]
 
@@ -666,15 +667,12 @@ 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.
         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()
@@ -989,7 +987,9 @@ class LlamaModel(LlamaPreTrainedModel):
         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)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_seen_tokens + inputs_embeds.shape[1]
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1053,31 +1053,16 @@ class LlamaModel(LlamaPreTrainedModel):
             attentions=all_self_attns,
         )
 
-    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
-
+    # 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):
         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]
@@ -1085,9 +1070,7 @@ class LlamaModel(LlamaPreTrainedModel):
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1]
-                if isinstance(attention_mask, torch.Tensor)
-                else past_seen_tokens + sequence_length + 1
+                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_length + 1
             )
 
         causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)

src/transformers/models/olmo/modeling_olmo.py

@@ -653,7 +653,6 @@ class OlmoSdpaAttention(OlmoAttention):
             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()
@@ -971,7 +970,9 @@ class OlmoModel(OlmoPreTrainedModel):
         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)
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_seen_tokens + inputs_embeds.shape[1]
+        )
 
         # embed positions
         hidden_states = inputs_embeds
@@ -1035,32 +1036,17 @@ class OlmoModel(OlmoPreTrainedModel):
             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
     # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
-    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
-
+    def _update_causal_mask(self, attention_mask, input_tensor, cache_position, current_length):
         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]
@@ -1068,9 +1054,7 @@ class OlmoModel(OlmoPreTrainedModel):
             target_length = self.config.max_position_embeddings
         else:  # dynamic cache
             target_length = (
-                attention_mask.shape[-1]
-                if isinstance(attention_mask, torch.Tensor)
-                else past_seen_tokens + sequence_length + 1
+                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_length + 1
             )
 
         causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)

tests/test_modeling_common.py

@@ -3772,42 +3772,6 @@ class ModelTesterMixin:
 
                 self.assertTrue(len(fail_cases) == 0, "\n".join(fail_cases))
 
-    @require_torch_sdpa
-    @require_torch_gpu
-    @slow
-    def test_sdpa_can_dispatch_on_flash(self):
-        compute_capability = torch.cuda.get_device_capability()
-        major, _ = compute_capability
-
-        if not torch.version.cuda or major < 8:
-            self.skipTest("This test requires an NVIDIA GPU with compute capability >= 8.0")
-
-        for model_class in self.all_model_classes:
-            if not model_class._supports_sdpa:
-                self.skipTest(f"{model_class.__name__} does not support SDPA")
-
-            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-            if config.model_type in ["llava", "llava_next", "vipllava"]:
-                self.skipTest("Llava-like models currently (transformers==4.39.1) requires an attention_mask input")
-            if config.model_type in ["idefics"]:
-                self.skipTest("Idefics currently (transformers==4.39.1) requires an image_attention_mask input")
-            model = model_class(config)
-
-            with tempfile.TemporaryDirectory() as tmpdirname:
-                model.save_pretrained(tmpdirname)
-                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, attn_implementation="sdpa")
-                model.to(torch_device)
-
-                inputs_dict.pop("attention_mask", None)
-                inputs_dict.pop("decoder_attention_mask", None)
-
-                for name, inp in inputs_dict.items():
-                    if isinstance(inp, torch.Tensor) and inp.dtype in [torch.float32, torch.float16]:
-                        inputs_dict[name] = inp.to(torch.float16)
-
-                with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
-                    _ = model(**inputs_dict)
-
     @require_torch_sdpa
     @slow
     def test_eager_matches_sdpa_generate(self):