training
Build, debug, and explain ms-swift pre-training and supervised fine-tuning workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Build, debug, and explain ms-swift pre-training and supervised fine-tuning workflows.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/cli-reference.md7.8 KB
- 参考文档references/troubleshooting.md8.6 KB
- 参考文档references/workflows.md9.4 KB
- 脚本scripts/build_training_command.py9.0 KB
- 脚本scripts/validate_training_config.py15.8 KB
03 SKILL.md 正文
ms-swift Training
Use this sub-skill when the task is about swift pt, swift sft, training configs, LoRA/QLoRA/full tuning, distributed training flags, multimodal training memory controls, checkpoint resume semantics, or handing trained checkpoints/adapters to inference/export workflows.
Route First
- Use
swift ptfor continued pre-training (CPT) or plain generative-language-model training where the chat template should be disabled; it is equivalent toswift sft --use_chat_template false --loss_scale all. - Use
swift sftfor supervised fine-tuning, chat/instruction tuning, LoRA/QLoRA/full tuning, multimodal SFT, and most training examples. - Reroute dataset schema design, custom dataset registration, and column mapping to
data-model-customization. - Reroute runtime inference, deployment servers, and generation backends to
inference-deployment. - Reroute RLHF/GRPO/GKD/PPO/KTO/DPO, Ray rollout, and Megatron distributed training depth to
advanced-rl-distributed. - Reroute export, quantization, model push, LoRA merge as a primary workflow, and standalone evaluation to
export-evaluation.
Read These References
- Training workflows for command patterns, YAML/JSON configs, tuning modes, checkpoints, multimodal options, and training-to-inference handoff.
- CLI reference for high-value
swift sftandswift ptarguments and safe option combinations. - Troubleshooting for download/offline behavior, CUDA/NPU setup, OOM, DDP, DeepSpeed/FSDP, LoRA/QLoRA, and template issues.
Bundled Scripts
- Use
scripts/build_training_command.py --helpto print a safe command skeleton forswift sftorswift ptwithout launching training. - Use
scripts/validate_training_config.py CONFIG.yamlorCONFIG.jsonto catch high-risk config mistakes before running training; it parses YAML/JSON and does not import heavy model code.
Quick Patterns
swift sft \
--model Qwen/Qwen2.5-7B-Instruct \
--dataset ./train.jsonl \
--split_dataset_ratio 0.05 \
--train_type lora \
--output_dir output/qwen2_5_lora_sft \
--max_length 2048 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 16 \
--learning_rate 1e-4 \
--save_steps 500
swift pt \
--model Qwen/Qwen2.5-7B \
--dataset ./corpus.jsonl \
--train_type lora \
--loss_scale all \
--use_chat_template false \
--output_dir output/qwen2_5_cpt
Minimum Review Checklist
- Confirm
model,datasetorcached_dataset,output_dir, tuning mode,max_length, batch/accumulation, dtype, and save/eval cadence. - Decide whether the run needs ModelScope or HuggingFace sources; use local model/dataset paths plus
--check_model falsefor fully offline runs. - For LoRA outputs, hand off inference with
--adapters checkpoint_dir; for full fine-tuning outputs, hand off with--model checkpoint_dir. - For vLLM/SGLang/LMDeploy acceleration after LoRA, plan a merge/export path; do not promise adapter-only QLoRA merge acceleration.
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).