Skip to content

[BugFix] Potential Fix for FA3 full-cudagraph IMA #25490

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
WoosukKwon merged 3 commits into from
Sep 24, 2025
Merged

[BugFix] Potential Fix for FA3 full-cudagraph IMA #25490

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e1c19ca
fix
LucasWilkinson Sep 23, 2025
91f28b4
Merge branch 'main' into lwilkinson/potential-full-CG-ima-fix
mgoin Sep 24, 2025
ad60fa8
fix CI
LucasWilkinson Sep 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 11 additions & 11 deletions vllm/v1/attention/backends/flash_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,10 +194,9 @@ def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],

self.use_full_cuda_graph = \
self.compilation_config.cudagraph_mode.has_full_cudagraphs()
self.max_cudagraph_size = self.compilation_config.max_capture_size

if self.use_full_cuda_graph and self.aot_schedule:
self.max_cudagraph_size = self.compilation_config.max_capture_size

if self.max_cudagraph_size > 992:
# This condition derives from FA3's internal heuristic.
# TODO(woosuk): Support larger cudagraph sizes.
Expand Down Expand Up @@ -259,6 +258,15 @@ def build(self,
self.aot_schedule = False
aot_schedule = False

max_num_splits = 0 # 0 means use FA3's heuristics, not CG compatible
if self.use_full_cuda_graph and \
num_actual_tokens <= self.max_cudagraph_size:
# NOTE(woosuk): Setting num_splits > 1 may increase the memory
# usage, because the intermediate buffers of size [num_splits,
# num_heads, num_tokens, head_size] are allocated. Therefore,
# we only set num_splits when using cuda graphs.
max_num_splits = self.max_num_splits

def schedule(batch_size, cu_query_lens, max_query_len, seqlens,
max_seq_len, causal):
cache_dtype = self.cache_config.cache_dtype
Expand All @@ -281,7 +289,7 @@ def schedule(batch_size, cu_query_lens, max_query_len, seqlens,
page_size=self.block_size,
causal=causal,
window_size=self.aot_sliding_window,
num_splits=self.max_num_splits,
num_splits=max_num_splits,
)
return None

Expand Down Expand Up @@ -322,7 +330,6 @@ def schedule(batch_size, cu_query_lens, max_query_len, seqlens,
max_seq_len=max_seq_len,
causal=causal)
# For FA3 + full cudagraph
max_num_splits = 0
if self.use_full_cuda_graph and scheduler_metadata is not None:
n = scheduler_metadata.shape[0]
self.scheduler_metadata[:n] = scheduler_metadata
Expand All @@ -333,13 +340,6 @@ def schedule(batch_size, cu_query_lens, max_query_len, seqlens,
self.scheduler_metadata[n:] = 0
scheduler_metadata = self.scheduler_metadata[:n]

if num_actual_tokens <= self.max_cudagraph_size:
# NOTE(woosuk): Setting num_splits > 1 may increase the memory
# usage, because the intermediate buffers of size [num_splits,
# num_heads, num_tokens, head_size] are allocated. Therefore,
# we only set num_splits when using cuda graphs.
max_num_splits = self.max_num_splits

attn_metadata = FlashAttentionMetadata(
num_actual_tokens=num_actual_tokens,
max_query_len=max_query_len,
Expand Down