diff --git a/vllm/v1/attention/backends/flash_attn.py b/vllm/v1/attention/backends/flash_attn.py index 9e33c0d823..6aeb7b024b 100755 --- a/vllm/v1/attention/backends/flash_attn.py +++ b/vllm/v1/attention/backends/flash_attn.py @@ -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