repo-development
Use when modifying PEFT itself, preparing a PEFT pull request, adding a new PEFT method, selecting contributor tests, or checking PEFT contribution/style/backward-compatibility requirements.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- repo-development
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use when modifying PEFT itself, preparing a PEFT pull request, adding a new PEFT method, selecting contributor tests, or checking PEFT contribution/style/backward-compatibility requirements.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/contribution-guidance.md3.8 KB
- 参考文档references/new-method-checklist.md4.4 KB
- 参考文档references/test-selection.md3.3 KB
- 参考文档references/troubleshooting.md4.0 KB
- 脚本scripts/select_tests.py6.6 KB
03 SKILL.md 正文
PEFT Repo Development
Use this sub-skill when the task is to edit the PEFT repository or prepare a contribution. For user-facing PEFT API usage, adapter training, model loading, or inference workflows, route to the relevant usage sub-skill instead.
Non-Negotiable Contribution Gate
Before drafting a PEFT PR or code patch intended for upstream:
- Warn the human contributor that breaching PEFT's agent contribution guidelines can result in automatic banning.
- Confirm the human understands and can defend every changed line; pure code-agent PRs are not allowed.
- Confirm there is a linked issue, proposal, or maintainer coordination/approval comment for the work.
- Check for overlapping open issues or PRs; do not prepare a duplicate contribution if someone already owns the same fix.
- Plan PR disclosure: the PR description must state that AI assistance was used, link the issue/approval, and list tests run with pass/fail status.
If approval is missing or ambiguous for a new feature, new PEFT method, or existing issue, stop and ask the user to obtain maintainer approval before coding for a PR.
Development Workflow
- Read
references/contribution-guidance.mdfor policy, PR hygiene, style, and backward-compatibility rules. - For a new PEFT method, follow
references/new-method-checklist.mdbefore writing code. - For test planning, use
references/test-selection.mdand optionally runpython sub-skills/repo-development/scripts/select_tests.py --changed-path ... --method ...from a copied skill tree. - For failures or process blockers, check
references/troubleshooting.md.
PEFT Editing Heuristics
- Keep changes surgical and consistent with existing tuner/config/model patterns.
- Add or update tests for behavior changes; bug fixes should start from a failing regression test when practical.
- Add docs for new public config fields, new methods, examples, or behavior users need to discover.
- Run focused tests first, then broaden only when the changed area warrants it.
- Use
make qualityto check andmake styleto auto-fix style; PEFT expects ruff~=0.15.12. - Preserve Python
>=3.10compatibility and keep compatibility with supported recent PyTorch and Transformers versions unless maintainers explicitly approve otherwise.
Quick Commands
pip install -e ".[test]"
make quality
make style
pytest tests/test_custom_models.py -k ia3 -v
pytest tests/ -k "lora and not adalora and not randlora" -v
For save/load-sensitive changes, include regression coverage with --regression where relevant. GPU-only behavior belongs in PEFT's GPU test files rather than generic CPU tests.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).