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>
(cherry picked from commit 0a3c71e7e5)
This commit is contained in:
Lucas Wilkinson
2026-01-30 21:15:06 -07:00
committed by khluu
parent 5f45b0b7e0
commit d984d664cc
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__(

12
vllm/v1/worker/gpu_model_runner.py
View File

@@ -1382,12 +1382,14 @@ class GPUModelRunner(
num_scheduled_tokens: dict[str, int],
kv_cache_spec: KVCacheSpec,
num_reqs: int,
for_cudagraph_capture: bool = False,
) -> tuple[torch.Tensor | None, np.ndarray | None]:
if not isinstance(kv_cache_spec, CrossAttentionSpec):
return None, None
# Zero out buffer for padding requests that are not actually scheduled (CGs)
self.encoder_seq_lens.np[:num_reqs] = 0
# Build encoder_seq_lens array mapping request indices to
# encoder lengths for inputs scheduled in this batch
for req_id in num_scheduled_tokens:
@@ -1404,6 +1406,15 @@ class GPUModelRunner(
feature.mm_position.length for feature in req_state.mm_features
)
self.encoder_seq_lens.np[req_index] = encoder_input_tokens
if for_cudagraph_capture:
# During CUDA graph capture, we need to use realistic encoder lengths
# so that max_seqlen_k is captured with the correct value.
max_encoder_len = getattr(
self.model_config.hf_config,
"max_source_positions",
self.max_encoder_len,
)
self.encoder_seq_lens.np[:num_reqs] = max_encoder_len
self.encoder_seq_lens.copy_to_gpu(num_reqs)
encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs]
@@ -1821,6 +1832,7 @@ class GPUModelRunner(
num_scheduled_tokens or {},
kv_cache_group.kv_cache_spec,
num_reqs_padded,
for_cudagraph_capture=for_cudagraph_capture,
)
if kv_cache_gid > 0:
cm.block_table_tensor = _get_block_table(kv_cache_gid)