cp-to-release
Skill2.1k starsChanged 20 days agoApache-2.0View on GitHub
Manually cherry-pick a merged NVIDIA-NeMo/RL PR onto the current release branch (r0.X.0) and open a cp PR with matching CI labels and /ok to test.
- Reads credentials
- Commits and pushes
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 0 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 22
Run from the NeMo-RL checkout.
- Very specific: line 29
Stop if merged is false. Keep title, merge_commit_sha, and the labels starting with CI:.
≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
cp-to-release
Input: the original PR number N (already merged). Optional: a release branch name; otherwise use the newest origin/rX.Y.Z.
Release flow since r0.8.0: PRs merge into the release branch first, and a label auto cherry-picks them back to main. A manual cp is needed when a PR was merged into main instead of the release branch first. (Up to r0.7.0 it was the reverse: main first, auto cp to release.)
The user invoking this skill authorizes the cherry-pick commit, push, PR, label, and
/ok to test — no pause-before-commit for the clean path. Do not assign reviewers; the user does.
Steps
Run from the NeMo-RL checkout. $GITHUB_TOKEN (public_repo scope) must be set in every Bash call that uses it; <user> is the user's GitHub branch prefix.
- Read the original PR.
Stop ifcurl -sH "Authorization: Bearer $GITHUB_TOKEN" https://api.github.com/repos/NVIDIA-NeMo/RL/pulls/<N> \ | jq '{title, merged, merge_commit_sha, base: .base.ref, labels: [.labels[].name]}'mergedis false. Keeptitle,merge_commit_sha, and the labels starting withCI:. - Pick the release branch (unless given):
Also checkgit fetch origin git branch -r | grep -E '^ *origin/r[0-9]+\.[0-9]+\.[0-9]+$' | sort -V | tail -1git ls-remote origin refs/heads/<user>/cp-<N>is empty; if the branch exists, ask. - Branch + cherry-pick in a worktree, so the user's main checkout (often dirty, on another
branch) is untouched. Put it under a temp directory:
git worktree add -b <user>/cp-<N> <tmpdir>/cp-<N> origin/<rel> cd <tmpdir>/cp-<N> && git cherry-pick -s <merge_commit_sha>-sis required: the commit must carry the user's ownSigned-off-byalongside the author's. Forgetting it means amend + force-push + a second/ok to test. - Conflicts: stop and hand back. Show the conflicting files; do not auto-resolve. After the
user resolves and continues, verify before pushing:
git log -1— the cherry-pick committed, and has the user'sSigned-off-by.grep -n "^<<<<<<< \|^=======$\|^>>>>>>> " <files>— no leftover markers.- Per file, compare the
+lines with the original:diff <(git diff HEAD~1..HEAD -- <f> | grep '^+' | grep -v '^+++' | sort) <(git show <merge_sha> -- <f> | grep '^+' | grep -v '^+++' | sort)and report any extra/missing lines.
- Push and open a PR targeting the release branch, never
main:
POSTgit push -u origin <user>/cp-<N>/repos/NVIDIA-NeMo/RL/pullswith"base": "<rel>","head": "<user>/cp-<N>".- Title:
cp: cherry-pick #<N> - <original title verbatim>(keeps semantic-pull-request CI happy). - Body, one line:
- Clean:
manually cp #<N> into <rel> branch since it was merged into main instead of <rel> first. - Conflict: append the conflict, e.g.
... first. There's a conflict in<file>(<why, e.g. depends on #M which is not in <rel>>).
- Clean:
- Title:
- Labels + CI. Copy the original's
CI:*labels (POST /issues/<pr>/labels). Do not copycherry-pick-to-main(it makes the bot cherry-pick the PR back to main) or release labels such asr0.8.0. Then post/ok to test <full head sha>as an issue comment. - Report: PR number (bare
#N), base, commit sha, labels, that/ok to testwas posted, and the worktree path (git worktree remove <path>to clean up).
Reference PRs
- #4412 (r0.8.0, clean, new flow): body
manually cp #4182 into r0.8.0 branch since it was merged into main instead of r0.8.0 first. - r0.7.0, old flow: #3211 (clean, empty body); conflicts: #3153
manually cp #3147 into r0.7.0 branch since there's a conflict indocs/guides/environments.md., #3195, #3214.
Trending
More talk lately.about the whole repository
Who's talking about NVIDIA-NeMo/RL
- vllm-project/vllm“[RFC]: Native Weight Syncing APIs”GitHub · Jan 7, 2026 · 14 reactions
- @sungkim.bsky.social“Nvidia has open-sourced the Llama-Nemotron-Super-V1.5 post-training Synthetic Data: …”Bluesky · Jul 31, 2025 · 4 likes
- vllm-project/vllm“[Bug]: Accuracy Discrepancy in Qwen-7B-Chat Evaluation: vLLM Logprobs Differ from SGLANG, Impacting …”GitHub · Jun 25, 2025 · 3 reactions
- huggingface/trl“Async-GRPO: Decouple rollout generation and reward computation/model update for faster training”GitHub · Nov 27, 2025 · 2 reactions
- Alvorecer721/Nemo-RL“fix(data): eval-path processors render a duplicated default system block (two chat-template calls)”GitHub · Sep 17, 2026 · 1 reactions
The biggest 5 of 52 mentions linking to the repository that our crawl found. Each line is their own headline.
See all 52
Talked about on GitHub and Bluesky (52 times).
On GitHub · 50
- Jan 7, 2026[RFC]: Native Weight Syncing APIs — vllm-project/vllm14 reactions · 11 comments
- Jun 25, 2025[Bug]: Accuracy Discrepancy in Qwen-7B-Chat Evaluation: vLLM Logprobs Differ from SGLANG, Impacting Performance — vllm-project/vllm3 reactions · 7 comments
- Nov 27, 2025Async-GRPO: Decouple rollout generation and reward computation/model update for faster training — huggingface/trl2 reactions · 3 comments
- Sep 17, 2026fix(data): eval-path processors render a duplicated default system block (two chat-template calls) — Alvorecer721/Nemo-RL1 reaction
- Jul 15, 2026docs(rl): add ModelExpress weight refit guide — ai-dynamo/dynamo1 reaction · 1 comment
45 more on GitHub
- Jul 9, 2026[roadmap] verl 26Q3 roadmap — verl-project/verl1 reaction · 5 comments
- Jan 8, 2026[Feat][RL][1/2] Native Weight Syncing API: NCCL — vllm-project/vllm1 reaction · 23 comments
- Oct 29, 2026feat: support FP32 LM-head output for NeMo RL / vLLM parity — NVIDIA-NeMo/Automodel0 reactions
- Oct 29, 2026Track softmax collapse and cancellation fixes in logprobs and losses — swiss-ai/Nemo-RL0 reactions
- Oct 27, 2026Support DAPO in async GRPO training — swiss-ai/Nemo-RL0 reactions · 1 comment
- Oct 25, 2026feat: preserve score heads — NVIDIA-NeMo/Gym0 reactions · 4 comments
- Oct 20, 2026Fix potential positional embedding tensor reuse for train/rollout 0-mismatch RL — NVIDIA/Megatron-LM0 reactions · 19 comments
- Oct 15, 2026feat(speechlm): support native RL checkpoints and frozen CTC export — NVIDIA-NeMo/Speech0 reactions · 20 comments
- Oct 15, 2026feat(audio): align ASR scoring and support explicit Ray launches — NVIDIA-NeMo/Skills0 reactions · 1 comment
- Oct 15, 2026feat(freezing): select exact parameters with FreezeConfig — NVIDIA-NeMo/Automodel0 reactions · 10 comments
- Oct 14, 2026fix(inference): support batch-invariant Mamba prefix caching — NVIDIA/Megatron-LM0 reactions · 7 comments
- Oct 11, 2026Latest 20 Papers - October 12, 2026 — xllm-io/daily-papers0 reactions
- Oct 8, 2026Latest 20 Papers - October 09, 2026 — xllm-io/daily-papers0 reactions
- Oct 8, 2026No VLLM DP path for EP>1 — swiss-ai/Nemo-RL0 reactions · 1 comment
- Oct 7, 2026Latest 20 Papers - October 08, 2026 — xllm-io/daily-papers0 reactions
- Oct 7, 2026Latest 15 Papers - October 08, 2026 — Hdksg10/DailyArxiv0 reactions
- Oct 7, 2026Support exporting MXFP8 values and scales from TE quantized tensors — NVIDIA/TransformerEngine0 reactions
- Oct 7, 2026Latest 15 Papers - October 07, 2026 — Ponkux/DailyArXiv-cp0 reactions
- Oct 7, 2026Latest 15 Papers - October 07, 2026 — ever-xing/DailyArXiv0 reactions
- Oct 2, 2026feat(swe): forward per-rollout sampling seeds to OpenHands — NVIDIA-NeMo/Gym0 reactions · 1 comment
- Oct 2, 2026feat: support episode-derived NeMo Gym request seeds — sdevare-nv/nv-OpenHands0 reactions
- Sep 28, 2026pjin/main-dev-20260928 — NVIDIA-NeMo/Gym0 reactions · 1 comment
- Sep 23, 2026feat(token-capture): add direct ledger manifest reader — NVIDIA-NeMo/Gym0 reactions · 1 comment
- Sep 22, 2026fix(peft): expose gather_output on output-layer substitutes — NVIDIA-NeMo/Megatron-Bridge0 reactions · 1 comment
- Sep 17, 2026Support vLLM version 0.29.0 — NVIDIA-NeMo/Gym0 reactions
- Sep 15, 2026feat(checkpoint): prefix token level recovery — NVIDIA-NeMo/Gym0 reactions · 1 comment
- Sep 13, 2026feat(checkpoint): add durable turn-level rollout recovery — NVIDIA-NeMo/Gym0 reactions · 1 comment
- Sep 10, 2026perf(data-plane): avoid local batch copies — rohitrango/NeMo-RL0 reactions · 2 comments
- Sep 4, 2026LLM Post training - Question regarding seq-mask-tis and global_valid_toks denominator scaling. — mlcommons/training0 reactions · 4 comments
- Sep 1, 2026decouple force_on_policy_ratio from step0 reforward. — google/tunix0 reactions · 1 comment
- Aug 26, 2026chore(customizer): enable configuring grpo hyperparams — NVIDIA-NeMo/nemo-helix0 reactions · 2 comments
- Aug 24, 2026fix(llm): preserve nvext through tool-call jail — ai-dynamo/dynamo0 reactions · 7 comments
- Aug 24, 2026Batch invariant nvls — NVIDIA/Megatron-LM0 reactions · 4 comments
- Aug 13, 2026[bug] nonlearnable params exist on different device than learnable parameters — NVIDIA-NeMo/Megatron-Bridge0 reactions · 5 comments
- Aug 10, 2026fold_weight crashes with AttributeError: 'NoneType' object has no attribute 'data' on Megatron models with tied word embeddings — NVIDIA/Model-Optimizer0 reactions · 1 comment
- Aug 4, 2026default port range to the 5000-5999 band — NVIDIA-NeMo/Gym0 reactions · 1 comment
- Aug 4, 2026nemo-gym / nemo-rl をライブラリレジストリに登録する — airas-org/airas0 reactions
- Jul 31, 2026Async colo upstreaming candidates — anlthms/RL0 reactions
- Jul 30, 2026[VDR][v0.5.0] Tutorials fail when followed verbatim — NVIDIA-NeMo/Gym0 reactions
- Jul 28, 2026test(vllm): cover upstream token metadata fast path — NVIDIA-NeMo/Gym0 reactions · 1 comment
- Jul 22, 2026NeMoRL: Container & Examples — eth-cscs/alps-extended-images0 reactions
- Jul 16, 2026Investigate torch.compile numerical differences between grad and no-grad RMSNorm forward passes — NVIDIA-NeMo/Automodel0 reactions · 1 comment
- Jul 13, 2026Support agentic training for SWE with TRTLLM rollout — hchings/RL0 reactions
- Jul 8, 2026fix: deterministic implementation of _float32_rms_norm_fwd — NVIDIA-NeMo/Automodel0 reactions · 6 comments
- Jun 10, 2026Nemotron Super RL is using Nano Parser — NVIDIA-NeMo/Nemotron0 reactions
On Bluesky · 2
- Jul 31, 2025Nvidia has open-sourced the Llama-Nemotron-Super-V1.5 post-training Synthetic Data: huggingface.co/datasets/nvi... Human Data:… — @sungkim.bsky.social4 likes and reposts
- Aug 23, 2025📦 NVIDIA-NeMo / RL ⭐ 683 (+3) 🗒 Python Scalable toolkit for efficient model reinforcement — @github-trending.bsky.social1 likes and repost
In this repository21 other files it gives its agents.
- AGENTS.mdAGENTS.md
- build-and-dependencySkill
- cicdSkill
- config-conventionsSkill
- contributingSkill
- copyrightSkill
- cp-to-mainSkill
- error-handlingSkill
- launch-nemo-rlSkill
- linting-and-formattingSkill
- nemo-rl-auto-researchSkill
- nemo-rl-brev-etiquetteSkill
- nemo-rl-docsSkill
- nemo-rl-session-memorySkill
- pr-html-concise-publicSkill
- removing-outdated-configSkill
- review-prSkill
- review-pr-teamSkill
- review-pr-team-verify-followupSkill
- review-pr-team-with-htmlSkill
- testingSkill
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
Reviews can't be read right now.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

