set AttentionCGSupport.UNIFORM_BATCH for fa2 on xpu (#46555)
Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Xinyu Chen committed
Jun 24, 2026 at 18:05 UTC
ede54b926ebeb1bbb2d2f622599157d868daf803
1 file changed
+1
-1
vllm/v1/attention/backends/flash_attn.py
+1
-1
@@ -312,7 +312,7 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
312
# https://github.com/vllm-project/vllm/issues/22945
313
_cudagraph_support = (
314
AttentionCGSupport.ALWAYS
315
- if get_flash_attn_version() == 3 or current_platform.is_xpu()
315
+ if get_flash_attn_version() == 3
316
else AttentionCGSupport.UNIFORM_BATCH
317
)
318
supports_update_block_table: bool = True