Skip to content

[None][fix] Always sync local ranks after prefetch in HfWeightLoader #13556

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
lancelly wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[None][fix] Always sync local ranks after prefetch in HfWeightLoader #13556

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 6 additions & 2 deletions tensorrt_llm/_torch/models/checkpoints/hf/weight_loader.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,8 +85,12 @@ def load_weights(self, checkpoint_dir: str,
f"Prefetching {prefetch_size / (1024**3):.2f}GB checkpoint files."
)
self.prefetch_files(weight_files)
# Ensure that all local ranks have finished prefetching before loading weights
local_mpi_barrier()
# Sync all local ranks unconditionally. `enable_prefetch` depends on
# `psutil.virtual_memory().available`, a per-rank volatile value, so
# different ranks may take different branches; gating the barrier on
# it would deadlock between ranks that prefetched and ranks that
# skipped. Ranks that didn't prefetch reach the barrier immediately.
local_mpi_barrier()

return self._load_weights_in_parallel(
weight_files, self._load_safetensors_file,
Expand Down
Loading