Example#1
Conversation
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 62b440b22a
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| case "$1" in | ||
| head) run_head ;; | ||
| worker) run_worker ;; | ||
| submit) run_submit ;; | ||
| *) echo "Unknown command: $1"; exit 1 ;; |
There was a problem hiding this comment.
Implement print-ip command in 30B launcher script
scripts/launch-p2p-rdma.sh always probes node IPs by running ${TRAIN_SCRIPT} print-ip in get_container_ip, but this script only handles head|worker|submit, so selecting TRAIN_SCRIPT=scripts/run-qwen3-30B-A3B-p2p.sh makes start_ray fail with Unknown command: print-ip before cluster startup. Add a print-ip branch (as done in the 235B script) or make the launcher use a fallback probe for scripts without that subcommand.
Useful? React with 👍 / 👎.
| --update-weight-transfer-mode p2p | ||
| --update-weight-buffer-size "$((1024 * 1024 * 1024))" | ||
| --check-weight-update-equal | ||
| "${extra_mismatch_args[@]}" |
There was a problem hiding this comment.
Guard optional extra_mismatch_args under nounset
With set -euo pipefail enabled, expanding "${extra_mismatch_args[@]}" aborts with an unbound-variable error unless that array is defined, and this name is not initialized in this script or in the sourced scripts/models/qwen3-30B-A3B.sh; as a result, the default submit path fails before ray job submit. Use a safe expansion (for example with a default-empty array) or initialize it explicitly.
Useful? React with 👍 / 👎.
| -e HSA_FORCE_FINE_GRAIN_PCIE=1 \ | ||
| -e HSA_NO_SCRATCH_RECLAIM=1 \ | ||
| -e HF_HOME=/root/.cache/huggingface \ | ||
| -e WANDB_KEY=cd411df8b73eb3f5c1ae1220cc1ec4e3c9d1f86e \ |
There was a problem hiding this comment.
Remove hardcoded WANDB API key from launcher
This commit embeds a concrete Weights & Biases API key directly in the script and injects it into every container, which is a credential leak and allows unintended writes against that account by anyone who can read or run the script. Load the key from a secure external source (environment/secret store) instead of committing it in plaintext.
Useful? React with 👍 / 👎.
No description provided.