Skip to content

docs(k8s): filter trainer torchrun logs to rank 0 instead of silencing #2552

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
samsja wants to merge 2 commits into
base: main
Choose a base branch
from
Draft

docs(k8s): filter trainer torchrun logs to rank 0 instead of silencing #2552

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ccfde8f
docs(k8s): filter trainer torchrun logs to rank 0 instead of silencing
samsja May 19, 2026
52f9a20
docs(k8s): point helm values.yaml comment at the multi-GPU torchrun p…
samsja May 19, 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
10 changes: 10 additions & 0 deletions docs/kubernetes.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,9 +258,19 @@ torchrun \
--node-rank=$RANK \
--nproc-per-node=8 \
--rdzv-endpoint=my-exp-trainer-0.$HEADLESS_SERVICE:29501 \
--log-dir=/data/outputs/logs/trainer/torchrun \
--local-ranks-filter=0 \
--redirect=3 \
--tee=3 \
src/prime_rl/trainer/sft/train.py @ configs/train.toml
```

`--local-ranks-filter=0 --tee=3` keeps only rank 0's stdout/stderr on the pod's
console (so Loki/the dashboard see each log line once instead of N times for an
N-GPU pod), while `--redirect=3 --log-dir=...` still writes every rank's
stdout/stderr to per-rank files under the mounted PVC for debugging. This
matches what the launcher does on single-node / SLURM deployments.

## Troubleshooting

### Can't access shared storage
Expand Down
6 changes: 5 additions & 1 deletion k8s/prime-rl/values.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -99,7 +99,11 @@ trainer:

# Auto-start configuration (set to false to use sleep infinity for debugging)
autoStart: false
command: "" # e.g., "uv run trainer @ /app/examples/reverse_text/rl/train.toml --output-dir /data/outputs"
# Single GPU: "uv run trainer @ /app/examples/reverse_text/rl/train.toml --output-dir /data/outputs"
# Multi-GPU: use torchrun and pass --local-ranks-filter=0 --tee=3 --redirect=3 --log-dir=... so
# only rank 0's stdout reaches the pod console (Loki/dashboard see each line once) while every
# rank's stdout/stderr is still written to per-rank files. See docs/kubernetes.md.
command: ""

# Helps reduce CUDA memory fragmentation with PyTorch allocator
pytorchCudaAllocConf: "expandable_segments:True"
Expand Down
Loading