AGENT SKILL

application-recipes

Navigate ColossalAI first-party application packages, environment isolation, data/model prerequisites, and command anatomy without assuming heavyweight app execution is verified.

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

01 前言 Front Matter

name
application-recipes
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Navigate ColossalAI first-party application packages, environment isolation, data/model prerequisites, and command anatomy without assuming heavyweight app execution is verified.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

Application Recipes

Use this sub-skill when the task names ColossalChat/Coati, Colossal-LLaMA, ColossalEval, ColossalQA, ColossalMoE, or asks which first-party ColossalAI application project to use.

Application packages are useful but broad. They often require separate installation, large models, datasets, credentials, CUDA kernels, or dependency pins that can conflict with the core ColossalAI package environment.

Route Here

  • Choose among first-party applications by task: RLHF/post-training, LLaMA pretraining/SFT, LLM evaluation, retrieval QA, or MoE training/inference.
  • Plan isolated environments and dependency constraints for application packages.
  • Explain command anatomy from application scripts without running heavy model/data workflows.
  • Diagnose app dependency conflicts, missing model weights, dataset format issues, and service/credential requirements.

Reroute

  • Core Booster plugin details: use ../booster-training/SKILL.md.
  • Core Colossal-Inference API and generation details: use ../inference-and-serving/SKILL.md.
  • Core launch/hostfile/NCCL issues: use ../installation-and-launch/SKILL.md.
  • ShardFormer or topology design: use ../parallelism-and-sharding/SKILL.md.

Application Map

  • ColossalChat / Coati: SFT, reward model, PPO, DPO, SimPO, ORPO, KTO, GRPO, LoRA, quantization, and post-training command families.
  • Colossal-LLaMA: continual pretraining, SFT, tokenizer/model/data preparation, training scripts, and LLaMA checkpoint inference examples.
  • ColossalEval: dataset preparation, inference configs, dataset/GPT evaluation configs, metrics, and leaderboard reproduction flow.
  • ColossalQA: LangChain-style retrieval conversation, vector stores, local/API LLMs, document loading, memory, and SQL/file paths.
  • ColossalMoE: MoE training and inference examples powered by ColossalAI.

References and Helper

  • references/application-overview.md gives a task-to-app routing table and environment isolation rules.
  • references/colossalchat.md covers ColossalChat/Coati stages and command families.
  • references/llama-eval-qa-moe.md covers Colossal-LLaMA, ColossalEval, ColossalQA, and ColossalMoE.
  • references/troubleshooting.md lists app-specific dependency, data, model, and credential failures.
  • scripts/application_env_matrix.py prints package names and notable dependency conflicts.

Operating Rules

  • Do not claim application imports or native tests are verified unless a separate app environment was prepared and checked.
  • Do not install app requirements into a stable core ColossalAI environment without checking version conflicts.
  • Treat shell scripts as command anatomy: adapt paths, model names, datasets, GPU counts, and secrets to the user's environment.

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