serving-apis
Serve LMDeploy models through OpenAI-compatible, Responses-compatible, Anthropic-compatible, proxy, and client integration APIs.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- serving-apis
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Serve LMDeploy models through OpenAI-compatible, Responses-compatible, Anthropic-compatible, proxy, and client integration APIs.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-reference.md6.2 KB
- 参考文档references/cli-reference.md5.9 KB
- 参考文档references/client-integration.md6.8 KB
- 参考文档references/troubleshooting.md6.6 KB
- 脚本scripts/check_server_contract.py7.3 KB
03 SKILL.md 正文
Serving APIs
Use this sub-skill when the task is to expose an LMDeploy model as a network service or connect an OpenAI/Responses/Anthropic-style client to that service.
Route Here
- Build
lmdeploy serve api_servercommands, including--server-name,--server-port,--model-name,--backend,--tp, cache/session flags,--api-keys,--ssl, docs, and parser flags. - Operate
lmdeploy serve proxyfor multipleapi_servernodes and route OpenAI-compatible traffic through the proxy. - Call
/v1/models,/v1/chat/completions,/v1/completions,/v1/responses,/v1/messages, and/v1/messages/count_tokens. - Use
lmdeploy.serve.openai.api_client.APIClient, the OpenAI Python SDK, Codex Responses integration, Claude Code Anthropic integration, orcurl. - Configure
--tool-call-parserand--reasoning-parserfor tool-call and reasoning output extraction.
Route Elsewhere
- Offline local text inference belongs in
pipeline-inference. - Multimodal message/image/video payload details belong in
vision-language. - Creating AWQ/GPTQ/SmoothQuant/other quantized artifacts belongs in
quantization. - Deep backend performance tuning, new-model support, kernel behavior, or backend config internals belong in
backend-extension.
Quick Start
lmdeploy serve api_server <model-or-path> \
--backend pytorch \
--server-name 0.0.0.0 \
--server-port 23333 \
--model-name lmdeploy-model \
--tp 1
Validate the exposed model and chat endpoint:
curl http://127.0.0.1:23333/v1/models
curl http://127.0.0.1:23333/v1/chat/completions \
-H "content-type: application/json" \
-d '{"model":"lmdeploy-model","messages":[{"role":"user","content":"Reply with pong"}],"max_tokens":32}'
If launched with --api-keys my-secret, add -H "Authorization: Bearer my-secret" to client requests.
References
references/cli-reference.md: command construction forapi_serverandproxy.references/api-reference.md: endpoint contracts and request/response notes.references/client-integration.md:APIClient, OpenAI SDK, Codex, Claude Code, and validation curls.references/troubleshooting.md: symptoms, causes, and recovery steps.scripts/check_server_contract.py: offline/optional-live contract checker for base URLs and endpoint shape.
Validation Pattern
- Run
lmdeploy serve api_server --helpandlmdeploy serve proxy --helpto verify installed CLI flags. - Start
api_serverwith an explicit--model-nameand any needed parser/auth flags. - Run
python scripts/check_server_contract.py --base-url http://127.0.0.1:23333 --model <model-name>for offline request preview. - Add
--probeto perform live/v1/models, chat, completions, Responses, and Anthropic contract probes.
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).