A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 11:00 UTC
[Bugfix][MM encoder] Fix ViT attention backend resolving for Turing GPU (#29614)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Nov 28, 2025 at 03:17 UTC
38658ec6f3b3a09a6cd205bab23a550b3d3f8c0e
1 file changed
+9
-8
vllm/platforms/cuda.py
+9
-8
@@ -264,14 +264,15 @@ class CudaPlatformBase(Platform):
264
cls, head_size: int, dtype: torch.dtype
265
) -> "AttentionBackendEnum":
266
# Try FlashAttention first
267
- try:
268
- backend_class = AttentionBackendEnum.FLASH_ATTN.get_class()
269
- if backend_class.supports_head_size(
270
- head_size
271
- ) and backend_class.supports_dtype(dtype):
272
- return AttentionBackendEnum.FLASH_ATTN
273
- except ImportError:
274
- pass
267
+ if (cc := cls.get_device_capability()) and cc.major >= 8:
268
+ try:
269
+ backend_class = AttentionBackendEnum.FLASH_ATTN.get_class()
270
+ if backend_class.supports_head_size(
271
+ head_size
272
+ ) and backend_class.supports_dtype(dtype):
273
+ return AttentionBackendEnum.FLASH_ATTN
274
+ except ImportError:
275
+ pass
276
277
return AttentionBackendEnum.TORCH_SDPA
278