AGENT SKILL

rl-and-rewards

Guides agents configuring Axolotl GRPO, EBFT, custom reward functions, vLLM training servers, NeMo Gym, Hatchery, and async online RL checks.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
rl-and-rewards
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Guides agents configuring Axolotl GRPO, EBFT, custom reward functions, vLLM training servers, NeMo Gym, Hatchery, and async online RL checks.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

RL and Rewards

Use this sub-skill when the task involves Axolotl online RL or reward-driven training: GRPO, GDPO-style reward aggregation, custom reward functions, axolotl vllm-serve for training, EBFT feature-matching rewards, NeMo Gym environment rewards, Hatchery-style remote RL flows, or async rollout debugging.

Route By Task

  • For GRPO configs, vLLM serving, async prefetch, replay, re-roll, and importance sampling, read references/grpo.md.
  • For EBFT structured or strided feature-matching runs, read references/ebft.md.
  • For reward function signatures, deterministic checks, multiple rewards, rollout functions, NeMo Gym rewards, and Hatchery reward hooks, read references/reward-functions.md.
  • For symptom-driven fixes around rewards, vLLM, async rollouts, EBFT mode selection, and resource constraints, read references/troubleshooting.md.
  • Before training, validate a local reward function with scripts/validate_reward_function.py: python scripts/validate_reward_function.py rewards.py:accuracy_reward.

Quick Workflow

  1. Classify the method first: rl: grpo for programmatic verifiers or environment rewards, rl: ebft for feature matching against ground-truth completions, and route DPO/KTO/ORPO/SimPO preference data to preference-tuning.
  2. For GRPO, define the YAML, reward module, and vLLM plan together; trl.reward_funcs, trl.num_generations, trl.max_completion_length, trl.use_vllm, and vllm.host/vllm.port must agree.
  3. For EBFT, choose ebft.mode before copying settings: structured mode uses prompt plus ground_truth and usually vLLM, while strided mode uses tokenized document anchors and does not require vLLM.
  4. Run local deterministic reward checks before axolotl train; then use axolotl preprocess config.yaml --debug only for data/tokenization visibility and axolotl train config.yaml only after the user accepts runtime cost.
  5. For server-backed RL, start or verify the generation/environment services before training; axolotl vllm-serve config.yaml must serve the same base model as the training config.

Boundaries

  • This sub-skill owns GRPO/GDPO online RL, reward functions, vLLM serving relationships, async rollout features, EBFT structured/strided setup, NeMo Gym reward routing, Hatchery reward hooks, and safe local reward validation.
  • Route static DPO, KTO, ORPO, SimPO, reward-model, and process-reward-model preference workflows to preference-tuning.
  • Route generic dataset format mapping, chat_template, schema discovery, and YAML-only validation to data-and-configs.
  • Route cluster launch, DeepSpeed/FSDP tuning, GPU placement, and throughput engineering to distributed-and-performance.
  • Route model architecture quirks, tokenizer templates, LoRA/QLoRA target modules, and adapter loading to model-loading-and-adapters.
  • Route CLI mechanics, fetch/docs commands, inference, and merge-lora operations to cli-and-operations.

Evidence Notes

This guidance is distilled from Axolotl GRPO, EBFT, vLLM serving, RLHF, schema, integration, example, and unit-test evidence. It does not claim live training, vLLM, GPU, model-loading, or environment-service verification.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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).

★ 351