[V1] V1 FlashInfer Attention (vllm-project#16684)

Signed-off-by: mgoin <[email protected]>
Co-authored-by: Aurick Qiao <[email protected]>

Author: 2 people

tests/v1/e2e/test_cascade_attention.py

@@ -1,13 +1,21 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import pytest
+
 from vllm import LLM, SamplingParams
 
+from ...utils import fork_new_process_for_each_test
+
 
-def test_cascade_attention(example_system_message, monkeypatch):
+@fork_new_process_for_each_test
+@pytest.mark.parametrize("attn_backend",
+                         ["FLASH_ATTN_VLLM_V1", "FLASHINFER_VLLM_V1"])
+def test_cascade_attention(example_system_message, monkeypatch, attn_backend):
     prompt = "\n<User>: Implement fibonacci sequence in Python.\n<Claude>:"
 
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
+        m.setenv("VLLM_ATTENTION_BACKEND", attn_backend)
 
         llm = LLM(model="Qwen/Qwen2-1.5B-Instruct")
         sampling_params = SamplingParams(temperature=0.0, max_tokens=100)

vllm/engine/arg_utils.py

@@ -1474,7 +1474,7 @@ def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
                                    recommend_to_remove=False)
                 return False
 
-        # No FlashInfer or XFormers so far.
+        # No XFormers so far.
         V1_BACKENDS = [
             "FLASH_ATTN_VLLM_V1", "FLASH_ATTN", "PALLAS", "PALLAS_VLLM_V1",
             "TRITON_ATTN_VLLM_V1", "TRITON_MLA", "FLASHMLA", "IPEX_V1"

vllm/platforms/cuda.py

@@ -213,6 +213,9 @@ def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
                         return ("vllm.attention.backends."
                                 "flashmla.FlashMLABackend")
         if use_v1:
+            if selected_backend == _Backend.FLASHINFER:
+                logger.info_once("Using FlashInfer backend on V1 engine.")
+                return "vllm.v1.attention.backends.flashinfer.FlashInferBackend"
             if selected_backend == _Backend.TRITON_ATTN_VLLM_V1:
                 logger.info_once("Using Triton backend on V1 engine.")
                 return ("vllm.v1.attention.backends."

vllm/v1/attention/backends/flash_attn.py

@@ -64,10 +64,6 @@ def get_kv_cache_shape(
             raise ValueError("Block size must be a multiple of 16.")
         return (2, num_blocks, block_size, num_kv_heads, head_size)
 
-    @staticmethod
-    def use_cascade_attention(*args, **kwargs) -> bool:
-        return use_cascade_attention(*args, **kwargs)
-
 
 @dataclass
 class FlashAttentionMetadata:
@@ -402,6 +398,9 @@ def schedule(cu_query_lens, max_query_len, seqlens, max_seq_len,
         )
         return attn_metadata
 
+    def use_cascade_attention(self, *args, **kwargs) -> bool:
+        return use_cascade_attention(*args, **kwargs)
+
 
 class FlashAttentionImpl(AttentionImpl):