training-and-configs
Assemble verl PPO, GRPO, SFT, and on-policy distillation training commands, Hydra overrides, backend strategy switches, and config troubleshooting guidance.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training-and-configs
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Assemble verl PPO, GRPO, SFT, and on-policy distillation training commands, Hydra overrides, backend strategy switches, and config troubleshooting guidance.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/configuration.md6.8 KB
- 参考文档references/training-workflows.md9.8 KB
- 参考文档references/troubleshooting.md5.8 KB
- 脚本scripts/build_ppo_command.py7.3 KB
03 SKILL.md 正文
training-and-configs
Use this sub-skill when a coding agent needs to build or debug verl training commands and Hydra config overrides for PPO-like RL, GRPO, SFT, or on-policy distillation.
Route by Task
- PPO / critic training: Use
references/training-workflows.mdfor thepython -m verl.trainer.main_ppocommand shape, required actor/critic/model/data/logger overrides, and FSDP or Megatron examples. - GRPO / critic-less RL: Use
references/training-workflows.mdforalgorithm.adv_estimator=grpo, grouped rolloutn, KL-loss settings, and no-critic command patterns. - SFT: Use
references/training-workflows.mdfortorchrun -m verl.trainer.sft_trainer,engine=fsdp, dataset message keys, LoRA/PEFT, and sequence-parallel switches. - On-policy distillation: Use
references/training-workflows.mdfordistillation.enabled=True, teacher resource sizing, teacher model overrides, and distillation loss modes. - Hydra and backend switches: Use
references/configuration.mdbefore changingmodel_engine, FSDP/FSDP2, Megatron, VeOmni, TorchTitan, rollout backend, batch, checkpoint, logger, or generated-reference config fields. - Failures and validation errors: Use
references/troubleshooting.mdfor overlong prompts, OOM, Hydra override syntax, config-doc CI, rollout backend mismatch, Ray/GPU resource, and distillation teacher sizing problems.
Bundled Tool
scripts/build_ppo_command.pyassembles a dry-runpython -m verl.trainer.main_ppocommand from common data, model, rollout, backend, batch, and logger options. It prints only; it never imports verl, starts Ray, downloads models, or runs training.
Boundaries
- Dataset schema details, reward/tool internals, backend installation, and checkpoint merge/export operations belong to other verl sub-skills or repo-level references.
- Do not point future agents at source checkout examples or scripts as runtime dependencies; adapt their patterns into commands or use the bundled script here.
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).