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>
This commit is contained in:
Xinyu Chen
2026-06-24 18:05:02 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent 52fbe12283
commit ede54b926e
+1 -1
View File
@@ -312,7 +312,7 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
# https://github.com/vllm-project/vllm/issues/22945
_cudagraph_support = (
AttentionCGSupport.ALWAYS
if get_flash_attn_version() == 3 or current_platform.is_xpu()
if get_flash_attn_version() == 3
else AttentionCGSupport.UNIFORM_BATCH
)
supports_update_block_table: bool = True