biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[BugFix] Fix whisper FA2 + full cudagraphs (#33360)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Lucas Wilkinson
2026-01-30 21:15:06 -07:00
committed by GitHub
parent 29fba76781
commit 0a3c71e7e5
2 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/v1/attention/backends/flash_attn.py
View File

@@ -263,18 +263,6 @@ class FlashAttentionMetadataBuilder(AttentionMetadataBuilder[FlashAttentionMetad
vllm_config: "VllmConfig",
kv_cache_spec: "AttentionSpec",
) -> AttentionCGSupport:
# FA2 does not support CUDA graphs with encoder-decoder models due to
# accuracy issues reported in https://github.com/vllm-project/vllm/issues/33091
if (
vllm_config.model_config.is_encoder_decoder
and get_flash_attn_version() == 2
):
logger.warning_once(
"FlashAttention2 does not support CUDA graphs with "
"encoder-decoder models due to accuracy issues reported in #33091. "
"Disabling CUDA graph."
)
return AttentionCGSupport.NEVER
return cls._cudagraph_support
def __init__(