Skip to content

Fix: Defensively close GPU device FDs in dataloader worker processes #3684

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

Draft
hexinw-nvidia wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

Fix: Defensively close GPU device FDs in dataloader worker processes #3684

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fb6f4e4
Fix: Defensively close GPU device FDs in dataloader worker processes
hexinw-nvidia Mar 4, 2026
c8c58ee
Merge branch 'main' into close_nvidia_fds
hexinw-nvidia Mar 4, 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
20 changes: 18 additions & 2 deletions megatron/training/datasets/data_samplers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,10 +73,26 @@ def build_pretraining_data_loader(dataset, consumed_samples):
raise Exception('{} dataloader type is not supported.'.format(args.dataloader_type))

def worker_init_fn(_):
DistributedSignalHandler(args.exit_signal).__enter__()
import os

# Defensively close GPU device FDs in worker processes so workers do not
# keep references into NVIDIA memory space. This helps ensure GPU memory
# can be reclaimed even if a dataloader worker is delayed or fails to exit.
def close_nvidia_fds():
for fd in os.listdir("/proc/self/fd"):
try:
path = os.readlink(f"/proc/self/fd/{fd}")
if path.startswith("/dev/nvidia"):
os.close(int(fd))
except Exception:
pass
Copy link

@claude claude bot Mar 4, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The bare except Exception: pass silently swallows all errors. At minimum, unexpected exceptions (not OSError/FileNotFoundError) deserve a warning log, otherwise bugs here are invisible in production.

Suggested change
pass
except OSError:
pass

CUDA fd paths like /dev/nvidia0, /dev/nvidiactl, /dev/nvidia-uvm should reliably resolve, so only OSError (ENOENT for already-closed fds, EBADF, etc.) is expected here.


close_nvidia_fds()
if args.exit_signal_handler:
DistributedSignalHandler(args.exit_signal).__enter__()

maybe_worker_init_fn = (
worker_init_fn if args.exit_signal_handler and args.num_workers > 0 else None
worker_init_fn if args.num_workers > 0 else None
)
# Torch dataloader.
if args.hybrid_context_parallel:
Expand Down