inference-deployment
Run and debug ms-swift inference, app, deployment, backend selection, OpenAI-compatible serving, multimodal requests, adapters, and result/logprob workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- inference-deployment
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Run and debug ms-swift inference, app, deployment, backend selection, OpenAI-compatible serving, multimodal requests, adapters, and result/logprob workflows.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/api-reference.md9.5 KB
- 参考文档references/backend-compatibility.md8.9 KB
- 参考文档references/troubleshooting.md9.0 KB
- 参考文档references/workflows.md10.8 KB
- 脚本scripts/build_inference_command.py9.4 KB
- 脚本scripts/smoke_openai_client.py4.8 KB
03 SKILL.md 正文
Inference and Deployment
Use this sub-skill when an agent needs to run swift infer, swift app, or swift deploy, choose an inference backend, call a local OpenAI-compatible server, use swift.infer_engine Python APIs, or debug serving/runtime failures.
Route first
- Use
references/workflows.mdfor concrete CLI, app, deploy, client, batch, multimodal, LoRA, logprob, and result-saving recipes. - Use
references/api-reference.mdforInferArguments,DeployArguments,AppArguments,InferRequest,RequestConfig, engines, OpenAI endpoints, and Python client shapes. - Use
references/backend-compatibility.mdbefore selectingtransformers,vllm,sglang, orlmdeploy. - Use
references/troubleshooting.mdfor backend/model mismatches, optional dependency failures, QLoRA limitations, multimodal OOM,vllm_limit_mm_per_prompt, result/logprob surprises, adapter loading, and unreachable local servers. - Use
scripts/build_inference_command.pyto print safeswift infer,swift deploy, orswift appcommand skeletons without launching models. - Use
scripts/smoke_openai_client.pyto test a local OpenAI-compatible/v1/chat/completionsserver; it targets localhost by default and does not use external network resources.
Use cases covered here
- Interactive text or multimodal CLI inference with
swift infer. - Dataset or batch inference with saved JSONL results and optional metrics/logprobs.
- Local Gradio-style inference UI with
swift app, either self-hosting a model or connecting to an existing base URL. - OpenAI-compatible serving with
swift deploy, including/v1/models,/v1/chat/completions,/v1/completions,/v1/embeddings, health/ping, and Swift client calls. - Backend selection across
transformers,vllm,sglang, andlmdeploy, including optional dependency caveats. - Full checkpoint, LoRA adapter, merged LoRA, and multiple named LoRA serving plans.
Reroute boundaries
- Training command construction, checkpoint creation, and LoRA training arguments belong to the training sub-skill.
- Dataset schema design and custom dataset registration belong to data/model customization.
- Export, merge-only packaging, quantization, and evaluation pipelines belong to export/evaluation.
- GRPO rollout, Ray, and Megatron serving/sampling internals belong to advanced RL/distributed workflows.
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).