From 791f92135a1429b3fdacffebf33c67c35188a76e Mon Sep 17 00:00:00 2001 From: Giancarlo Delfin <32987265+TheEpicDolphin@users.noreply.github.com> Date: Tue, 13 Oct 2026 16:22:52 -0700 Subject: [PATCH] [Attention] Set ALWAYS cudagraph support for FA4 (#60908) Signed-off-by: Giancarlo Delfin Co-authored-by: Woosuk Kwon --- vllm/v1/attention/backends/flash_attn.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/v1/attention/backends/flash_attn.py b/vllm/v1/attention/backends/flash_attn.py index 6e906bc520..bb19f5c804 100755 --- a/vllm/v1/attention/backends/flash_attn.py +++ b/vllm/v1/attention/backends/flash_attn.py @@ -593,7 +593,7 @@ def _maybe_symmetrize_window( class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetadata]): - # FA3: + # FA3/FA4: # Supports full cudagraphs for all cases. # # FA2: @@ -613,7 +613,7 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad # https://github.com/vllm-project/vllm/issues/22945 _cudagraph_support = ( AttentionCGSupport.ALWAYS - if get_flash_attn_version() == 3 + if get_flash_attn_version() in (3, 4) else AttentionCGSupport.UNIFORM_BATCH ) supports_update_block_table: bool = True