Support encoder_only attention for FlexAttention (#22273)

Signed-off-by: Max de Bayser <mbayser@br.ibm.com>

Author: maxdebayser

tests/kernels/test_flex_attention.py

@@ -9,7 +9,9 @@
 import torch
 from packaging import version
 
-from vllm import LLM, SamplingParams
+from vllm import SamplingParams
+
+from ..models.utils import check_embeddings_close
 
 TORCH_VERSION = version.parse(torch.__version__)
 MINIMUM_TORCH_VERSION = version.parse("2.7.0")
@@ -28,15 +30,15 @@ def set_seed(seed):
     not torch.cuda.is_available() or TORCH_VERSION < MINIMUM_TORCH_VERSION,
     reason="CUDA not available or PyTorch version < 2.7",
 )
-def test_flex_attention_vs_default_backend(monkeypatch):
+def test_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
     """Test that FlexAttention produces the same outputs as the default backend.
 
     This test compares the outputs from the FlexAttention backend with
     the default backend, ensuring they are identical when using the same seed.
     """
     model_name = "Qwen/Qwen2.5-1.5B-Instruct"
     seed = 42
-    max_tokens = 32
+    max_tokens = 24
     prompts = [
         "Hello, my name is",
         "The president of the United States is",
@@ -54,39 +56,85 @@ def test_flex_attention_vs_default_backend(monkeypatch):
         m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
 
         set_seed(seed)
-
-        llm_flex = LLM(
-            model_name,
-            tensor_parallel_size=1,
-            num_gpu_blocks_override=128,
-            enforce_eager=True,
-        )
-        output_flex = llm_flex.generate(prompts, sampling_params)
+        with vllm_runner(model_name,
+                         runner="generate",
+                         tensor_parallel_size=1,
+                         num_gpu_blocks_override=128,
+                         enforce_eager=True) as llm_flex:
+            output_flex = llm_flex.generate(prompts, sampling_params)
 
     # Run with default backend
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
         set_seed(seed)
-        llm_default = LLM(
-            model_name,
-            tensor_parallel_size=1,
-            num_gpu_blocks_override=128,
-            enforce_eager=True,
-        )
-        output_default = llm_default.generate(prompts, sampling_params)
+        with vllm_runner(model_name,
+                         runner="generate",
+                         tensor_parallel_size=1,
+                         num_gpu_blocks_override=128,
+                         enforce_eager=True) as llm_default:
+            output_default = llm_default.generate(prompts, sampling_params)
 
     # Compare outputs from both backends
     for i, (flex_result,
             default_result) in enumerate(zip(output_flex, output_default)):
         prompt = prompts[i]
-        flex_text = flex_result.outputs[0].text
-        default_text = default_result.outputs[0].text
+        flex_text = flex_result[1][0]
+        default_text = default_result[1][0]
 
         assert flex_text == default_text, (
             f"FlexAttention output doesn't match default for: {prompt!r}\n"
             f"FlexAttention: {flex_text!r}\n"
             f"Default: {default_text!r}")
 
 
+@pytest.mark.skipif(
+    not torch.cuda.is_available() or TORCH_VERSION < MINIMUM_TORCH_VERSION,
+    reason="CUDA not available or PyTorch version < 2.7",
+)
+def test_encoder_flex_attention_vs_default_backend(vllm_runner, monkeypatch):
+    """Test that FlexAttention produces the same outputs as the default backend.
+
+    This test compares the outputs from the FlexAttention backend with
+    the default backend for encoder models.
+    """
+    model_name = "BAAI/bge-base-en-v1.5"
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+    ]
+
+    # Run with flex attention
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "1")
+        m.setenv("VLLM_ATTENTION_BACKEND", "FLEX_ATTENTION")
+        with vllm_runner(model_name,
+                         runner="pooling",
+                         dtype=torch.bfloat16,
+                         tensor_parallel_size=1,
+                         max_model_len=100,
+                         enforce_eager=True) as llm_flex:
+            flex_outputs = llm_flex.embed(prompts)
+
+    # Run with default backend
+    with monkeypatch.context() as m:
+        m.setenv("VLLM_USE_V1", "1")
+        with vllm_runner(model_name,
+                         runner="pooling",
+                         dtype=torch.bfloat16,
+                         tensor_parallel_size=1,
+                         max_model_len=100,
+                         enforce_eager=True) as llm_default:
+            default_outputs = llm_default.embed(prompts)
+
+    check_embeddings_close(
+        embeddings_0_lst=flex_outputs,
+        embeddings_1_lst=default_outputs,
+        name_0="flex",
+        name_1="default",
+        tol=1e-2,
+    )
+
+
 if __name__ == "__main__":
     pytest.main([__file__])

vllm/v1/attention/backends/flex_attention.py

@@ -148,6 +148,7 @@ def causal_mask_mod(b: torch.Tensor, h: torch.Tensor, q_idx: torch.Tensor,
 
 @dataclass
 class FlexAttentionMetadata:
+    causal: bool
     num_actual_tokens: int  # Number of tokens excluding padding.
     max_query_len: int
     query_start_loc: torch.Tensor
@@ -177,10 +178,9 @@ class FlexAttentionMetadata:
     num_blocks = 0
     block_mask: Optional[BlockMask] = None
     score_mod: Optional[_score_mod_signature] = None
-    mask_mod: Optional[_mask_mod_signature] = None
     logical_mask_mod: _mask_mod_signature = causal_mask_mod
 
-    def get_mask_mod(self) -> _mask_mod_signature:
+    def get_causal_mask_mod(self) -> _mask_mod_signature:
         """Creates the mask_mod function for FlexAttention.
 
         This function creates the combined mask mod function that handles:
@@ -233,14 +233,39 @@ def final_mask_mod(
 
         return final_mask_mod
 
+    def get_bidirectional_mask_mod(self) -> _mask_mod_signature:
+        """Creates the encoder mask_mod function for FlexAttention.
+
+        Since the encoder bidirectional attention doesn't run with 
+        KV cache, this function creates a mask based on the
+        packed query sequences.
+        """
+        # Create a lookup mapping from query indices -> request number
+        request_lookup = _offsets_to_doc_ids_tensor(self.query_start_loc)
+
+        def final_mask_mod(
+            b: torch.Tensor,
+            h: torch.Tensor,
+            q_idx: torch.Tensor,
+            kv_idx: torch.Tensor,
+        ) -> torch.Tensor:
+            return request_lookup[q_idx] == request_lookup[kv_idx]
+
+        return final_mask_mod
+
     def build_block_mask(self) -> BlockMask:
-        assert self.mask_mod is not None
+        if self.causal:
+            mask_mod = self.get_causal_mask_mod()
+            kv_len = self.total_cache_tokens
+        else:
+            mask_mod = self.get_bidirectional_mask_mod()
+            kv_len = self.num_actual_tokens
         return create_block_mask_compiled(
-            self.mask_mod,
+            mask_mod,
             None,
             None,
             self.num_actual_tokens,
-            self.total_cache_tokens,
+            kv_len,
             device=self.block_table.device,
         )
 
@@ -251,7 +276,6 @@ def __post_init__(self):
         assert self.prefix_kv_lens is None, "Not implemented yet."
         assert self.suffix_kv_lens is None, "Not implemented yet."
         self.num_blocks = self.total_cache_tokens // self.block_size
-        self.mask_mod = self.get_mask_mod()
         self.block_mask = self.build_block_mask()
 
 
@@ -306,6 +330,7 @@ def build(self,
             self.device, non_blocking=True)
 
         out = FlexAttentionMetadata(
+            causal=common_attn_metadata.causal,
             num_actual_tokens=num_actual_tokens,
             max_query_len=max_query_len,
             query_start_loc=query_start_loc,
@@ -350,6 +375,12 @@ def __init__(
         self.head_size = head_size
         self.scale = float(scale)
         self.num_kv_heads = num_kv_heads
+        self.attn_type = attn_type
+
+        if attn_type not in (AttentionType.ENCODER_ONLY,
+                             AttentionType.DECODER):
+            raise NotImplementedError(
+                f"FlexAttention does not support {attn_type} attention")
 
         if alibi_slopes is not None:
             raise NotImplementedError(
@@ -425,26 +456,38 @@ def forward(
 
         num_actual_tokens = attn_metadata.num_actual_tokens
 
-        key_cache, value_cache = kv_cache.unbind(0)
-
-        torch.ops._C_cache_ops.reshape_and_cache_flash(
-            key,
-            value,
-            key_cache,
-            value_cache,
-            attn_metadata.slot_mapping,
-            self.kv_cache_dtype,
-            layer._k_scale,
-            layer._v_scale,
-        )
+        if not attn_metadata.causal:
+            assert self.attn_type == AttentionType.ENCODER_ONLY
+
+            query, key_tensor, value_tensor = map(
+                lambda x: self.view_as_4d(x).permute(0, 2, 1, 3),
+                (query, key, value),
+            )
+
+        else:
+            assert self.attn_type == AttentionType.DECODER
+            key_cache, value_cache = kv_cache.unbind(0)
+
+            torch.ops._C_cache_ops.reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping,
+                self.kv_cache_dtype,
+                layer._k_scale,
+                layer._v_scale,
+            )
+
+            # View out the block_size dim
+            key_cache = key_cache.view(-1, self.num_kv_heads, self.head_size)
+            value_cache = value_cache.view(-1, self.num_kv_heads,
+                                           self.head_size)
+            query, key_tensor, value_tensor = map(
+                lambda x: self.view_as_4d(x).permute(0, 2, 1, 3),
+                (query, key_cache, value_cache),
+            )
 
-        # View out the block_size dim
-        key_cache = key_cache.view(-1, self.num_kv_heads, self.head_size)
-        value_cache = value_cache.view(-1, self.num_kv_heads, self.head_size)
-        query, key_cache, value_cache = map(
-            lambda x: self.view_as_4d(x).permute(0, 2, 1, 3),
-            (query, key_cache, value_cache),
-        )
         query = query[:, :, :num_actual_tokens, :]
         # Doesn't work for now -> constraint violation
         # torch._dynamo.try_mark_dynamic(query, 2)
@@ -465,8 +508,8 @@ def forward(
 
         out = flex_attention_compiled(
             query,
-            key_cache,
-            value_cache,
+            key_tensor,
+            value_tensor,
             attn_metadata.score_mod,
             attn_metadata.block_mask,
             self.scale,