Skip to content

[Bugfix][Core] Sync reused pinned input buffers under PP batch queue (fixes IMA with sparse MLA + PP) #47644

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

Open
thomaslwang wants to merge 2 commits into
base: main
Choose a base branch
from
+6 −3
Open

[Bugfix][Core] Sync reused pinned input buffers under PP batch queue (fixes IMA with sparse MLA + PP) #47644

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d6d15a0
[Bugfix][Core] Sync reused pinned input buffers under PP batch queue
thomas-l-wang Jul 5, 2026
79dba7a
Merge branch 'main' into fix-pp-pinned-buffer-race
thomaslwang Jul 6, 2026
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
9 changes: 6 additions & 3 deletions vllm/v1/worker/gpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -713,11 +713,14 @@ def __init__(
# Separate cuda stream for overlapping transfer of sampled token ids from
# GPU to CPU when async scheduling is enabled.
self.async_output_copy_stream: torch.cuda.Stream | None = None
# cuda event to synchronize use of reused CPU tensors between steps
# when async scheduling is enabled.
self.prepare_inputs_event: torch.Event | None = None
if self.use_async_scheduling:
self.async_output_copy_stream = torch.cuda.Stream()
# cuda event to synchronize use of reused CPU tensors between steps
# whenever steps overlap (async scheduling, or the PP batch queue):
# input prep for step N+1 runs while step N's non_blocking H2D copies
# from the same pinned buffers may still be pending.
self.prepare_inputs_event: torch.Event | None = None
if self.vllm_config.max_concurrent_batches > 1:
self.prepare_inputs_event = torch.Event()

# self.cudagraph_batch_sizes sorts in ascending order.
Expand Down
Loading