finetuning-quantization
Route Qwen SFT, LoRA, Q-LoRA, DeepSpeed, adapter merge, and GPTQ quantization workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- finetuning-quantization
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Route Qwen SFT, LoRA, Q-LoRA, DeepSpeed, adapter merge, and GPTQ quantization workflows.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/data-format.md1.7 KB
- 参考文档references/deepspeed-and-hardware.md2.0 KB
- 参考文档references/finetuning-workflows.md2.7 KB
- 参考文档references/quantization.md1.4 KB
- 参考文档references/troubleshooting.md1.6 KB
- 脚本scripts/build_finetune_command.py2.4 KB
- 脚本scripts/validate_finetune_data.py2.2 KB
03 SKILL.md 正文
Qwen Fine-Tuning and Quantization
Use this sub-skill when the user wants to prepare supervised fine-tuning data, run full-parameter fine-tuning, LoRA, Q-LoRA, merge adapters, or quantize a fine-tuned Qwen checkpoint with GPTQ.
Safe start
- Use
scripts/validate_finetune_data.pyto validate the JSON training data shape before launching training. - Use
scripts/build_finetune_command.pyto print a command plan without starting training or loading a model. - Treat GPU, DeepSpeed, PEFT, and AutoGPTQ requirements as explicit prerequisites, not as implied by a Transformers import.
Routes
| User request | Read |
|---|---|
Training data shape, finetune.py arguments, masking, dataset loading, or file format issues |
references/data-format.md |
| Full fine-tuning, LoRA, Q-LoRA, DeepSpeed, FSDP, multinode, and merge/save behavior | references/finetuning-workflows.md |
| GPTQ quantization, output conversion, model copy steps, or AutoGPTQ compatibility | references/quantization.md |
| GPU memory, ZeRO stages, precision, optional deps, checkpoint naming, or training failure modes | references/deepspeed-and-hardware.md |
| Validation, optional dependency, adapter, GPTQ, or data-format errors | references/troubleshooting.md |
Boundaries
- For ordinary inference of the resulting checkpoint, use
../inference-model-loading/SKILL.md. - For serving the fine-tuned model, use
../serving-deployment/SKILL.md. - For function-calling training samples or ChatML/tokenizer details, use
../prompting-tool-use-tokenization/SKILL.md. - For evaluation scripts, use
../evaluation-reproduction/SKILL.md.
Operating rules
- Keep data validation separate from training. A valid JSON sample does not prove that training will fit in memory or that the chosen ZeRO stage is compatible.
- For Q-LoRA, the repository expects the Int4 chat checkpoint and fp16, not a BF16 base checkpoint.
- Do not recommend merging Q-LoRA adapters into a standalone model; that path is only documented for LoRA.
- Do not run
torchrun,deepspeed, GPTQ quantization, or adapter merge as a smoke test unless the user explicitly wants the side effect.
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).