[MM Encoder] Default to use TORCH_SDPA backend for ViT on Volta/Turing GPU (#36472)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

tests/kernels/attention/test_mha_attn.py

@@ -19,6 +19,7 @@ from vllm.model_executor.layers.attention import MMEncoderAttention
 from vllm.platforms import current_platform
 from vllm.platforms.cpu import CpuPlatform
 from vllm.platforms.cuda import CudaPlatform
+from vllm.platforms.interface import DeviceCapability
 from vllm.platforms.rocm import RocmPlatform
 from vllm.utils.torch_utils import set_default_torch_dtype, set_random_seed
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
@@ -83,6 +84,20 @@ def test_mha_attn_platform(default_vllm_config, device: str):
             attn = MMEncoderAttention(16, 72, scale=1)
             assert attn.attn_backend == AttentionBackendEnum.TRITON_ATTN
 
+        # Test Turing (pre-Ampere, sm_75): FlashAttention requires sm>=80,
+        # and Triton no longer supports MMA on Turing, so we expect that
+        # TORCH_SDPA is used for MMEncoderAttention.
+        with (
+            patch("vllm.model_executor.models.vision.current_platform", CudaPlatform()),
+            patch.object(
+                CudaPlatform,
+                "get_device_capability",
+                return_value=DeviceCapability(major=7, minor=5),
+            ),
+        ):
+            attn = MMEncoderAttention(16, 64, scale=1)
+            assert attn.attn_backend == AttentionBackendEnum.TORCH_SDPA
+
 
 def ref_attention(
     query: torch.Tensor,

vllm/platforms/cuda.py

@@ -413,12 +413,20 @@ class CudaPlatformBase(Platform):
 
     @classmethod
     def get_supported_vit_attn_backends(cls) -> list["AttentionBackendEnum"]:
-        return [
-            AttentionBackendEnum.FLASH_ATTN,
-            AttentionBackendEnum.TRITON_ATTN,
-            AttentionBackendEnum.TORCH_SDPA,
-            AttentionBackendEnum.FLASHINFER,
-        ]
+        if cls.has_device_capability(80):
+            return [
+                AttentionBackendEnum.FLASH_ATTN,
+                AttentionBackendEnum.TRITON_ATTN,
+                AttentionBackendEnum.TORCH_SDPA,
+                AttentionBackendEnum.FLASHINFER,
+            ]
+        else:
+            return [
+                AttentionBackendEnum.FLASH_ATTN,
+                AttentionBackendEnum.TORCH_SDPA,
+                AttentionBackendEnum.TRITON_ATTN,
+                AttentionBackendEnum.FLASHINFER,
+            ]
 
     @classmethod
     def get_vit_attn_backend(
@@ -438,7 +446,7 @@ class CudaPlatformBase(Platform):
         cc = cls.get_device_capability()
         for vit_attn_backend in cls.get_supported_vit_attn_backends():
             if vit_attn_backend == AttentionBackendEnum.TORCH_SDPA:
-                continue
+                return vit_attn_backend
             try:
                 backend_class = vit_attn_backend.get_class()
                 is_backend_supported = backend_class.supports_head_size(