计划执行与检查节点
executing-plans
按既定实施计划完成工作，保持过程检查和反馈。
探索技能
效率与协作 · SKILL.md
post-training-rl-and-multimodal
post-training-rl-and-multimodal【效率与协作】Route optional Megatron ModelOpt, RL/GRPO, VLM, and multimodal/MIMO workflows with explicit dependency and hardware limits.
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
技能说明
GitHubname: post-training-rl-and-multimodal description: "Route optional Megatron ModelOpt, RL/GRPO, VLM, and multimodal/MIMO workflows with explicit dependency and hardware limits." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION
post-training-rl-and-multimodal
Use this sub-skill for ModelOpt quantization/pruning/distillation/export, Megatron RL/GRPO, VLM, audio/video, and MIMO workflows. These surfaces are optional and usually need more than the minimal Megatron Core install.
Read first
- For ModelOpt and post-training operations, read references/post-training-workflows.md.
- For RL/GRPO concepts, rollout/inference handles, packed data, and configs, read references/rl-workflows.md.
- For VLM/multimodal/MIMO data and model-provider boundaries, read references/multimodal-workflows.md.
- For missing optional packages, data/checkpoint mismatch, GPU memory, and config failures, read references/troubleshooting.md.
Route by task
| Task | Route |
|---|---|
| Quantize/prune/export a model | Confirm ModelOpt install, checkpoint/model compatibility, and target artifact before running. |
| Run GRPO/RL post-training | Confirm reward/environment configs, rollout backend, packed sequence schema, and GPU topology. |
| Train a VLM/MIMO model | Confirm media manifest/schema, encoder/model-provider, tokenizer, and heterogeneous topology. |
| Only run a standard GPT pretrain | Route to ../training-cli-and-data/SKILL.md. |
| Missing ModelOpt/vision/audio dependency | Route to ../install-and-environment/SKILL.md for a minimal extra install; do not install all dev deps blindly. |
Safety
- Do not start long RL training, model conversion, or server processes as a probe.
- Use
--help, import checks, and tiny mock fixtures first. - Treat external reward services, HF tokens, media datasets, and model checkpoints as user-provided resources; never embed credentials.
- Expect H100/Blackwell and large-memory assumptions for many current examples; document unverified hardware rather than substituting a CPU check.
Boundaries
- Core model/topology decisions belong to ../core-models-and-parallelism/SKILL.md.
- Generic environment and optional dependency installation belongs to ../install-and-environment/SKILL.md.
- Checkpoint format/conversion belongs to ../checkpointing-and-conversion/SKILL.md.