verl-project
- 12 skills
- 0 followers
- 5 hours ago last updated
- ▌ Ascend CI Case Diff Scan 2 · verl-projectForward to the repo-local shared skill implementation at `.agent/skills/ascend-ci-case-diff-scan` for scanning an external verl repository for Ascend/NPU CI coverage gaps.
- ▌ Manage Wheelhouse Components · verl-projectUpgrade a pinned component version or add a brand-new CUDA wheel component to the verl-wheelhouse build pipeline. Covers editing versions.yaml, writing ci/build_scripts/<builder>.sh, and creating/copying a .github/workflows/build-<component>.yml trigger workflow. Use when asked to bump/update/pin a submodule version, add a new wheel/component to the wheelhouse, or edit versions.yaml in this repo.
- ▌ Ascend CI Case Diff Scan · verl-project bundleUse when auditing Ascend/NPU CI test coverage in a verl repository, checking whether CPU/GPU workflow test cases have corresponding NPU workflow coverage, or identifying test cases missing from NPU CI workflows
- ▌ Pr · verl-project bundlePrepare, create, or update a verl-vla pull request following repository conventions. Use when the user asks to open, draft, publish, or revise a PR; inspect the complete branch diff and existing GitHub edits, require final confirmation before creating a branch, pushing, or changing a PR.
- ▌ Issue · verl-project bundlePrepare, create, or update verl-vla GitHub issues following repository conventions. Use when the user asks to draft, file, open, or revise a bug report or feature request; search for duplicates, gather the issue-template details, and require final confirmation before changing GitHub.
- ▌ Commit · verl-project bundleSynchronize the current tracked branch, then prepare and create a local Git commit following verl-vla conventions. Use when the user asks to commit current changes, organize a coherent local commit, or validate a proposed commit message; preserve local work while pulling and resolving conflicts, require final confirmation before staging or committing, and do not push.
- ▌ Upgrade Deps · verl-projectUpgrade, downgrade, or re-pin a Python dependency (vllm, sglang, torch, transformers, flash-attn, etc.) in verl's universal uv.lock flow.
- ▌ Self Review · verl-projectReview your own verl-omni branch against the project rubric before opening or updating a PR. Use before submitting a contribution, or whenever asked to self-review. Report-only: it groups findings by severity, gives a READY / NEEDS CHANGES verdict, and never edits files. Mirrors the checks a maintainer applies, over your whole diff.
- ▌ Add Pipeline · verl-projectRouter for adding a diffusion or omni pipeline to verl-omni. Classifies the task (new architecture vs new algorithm; policy-gradient vs direct-preference) and points at the authoritative guide under docs/contributing/. Use when integrating a new model, or a new model+algorithm pair such as flow-GRPO / DPO / NFT / DanceGRPO.
- ▌ Commit And Pr · verl-projectverl-omni commit message + PR conventions and the mandatory contribution policy. MUST load before any git commit or PR creation -- enforces the [{modules}] {type}: {description} title format, commit trailers, duplicate-work checks, and AI-assistance disclosure.
- ▌ Run Cpu Tests · verl-projectHow to write and run verl-omni CPU tests (test_*_on_cpu.py) that exercise adapters, rewards, and configs without a GPU or model weights. Use when adding tests, reproducing a failure locally, or producing the test evidence a PR body requires.
- ▌ Add Reward Score · verl-projectGuide for adding a new reward scorer to verl-omni and wiring it into a run. Use when adding a reward function or reward model for image, video, or multimodal RL (flow-GRPO, DanceGRPO, DPO), including preference models and remote HTTP scorers.