pipeline-inference
Use LMDeploy pipeline APIs and chat CLI for offline text LLM inference, streaming, sessions, generation settings, backend configs, and chat templates.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- pipeline-inference
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use LMDeploy pipeline APIs and chat CLI for offline text LLM inference, streaming, sessions, generation settings, backend configs, and chat templates.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-reference.md9.7 KB
- 参考文档references/chat-templates.md6.3 KB
- 参考文档references/troubleshooting.md7.8 KB
- 参考文档references/workflows.md10.9 KB
- 脚本scripts/inspect_pipeline_config.py6.8 KB
03 SKILL.md 正文
Pipeline Inference
Use this sub-skill when the task is offline text LLM inference through lmdeploy.pipeline, Pipeline, or lmdeploy chat.
Read First
references/api-reference.mdfor exact constructors, methods, response fields, and config defaults.references/workflows.mdfor copy-ready offline inference patterns, batch prompts, streaming, sessions, LoRA, logits, hidden states, and PPL.references/chat-templates.mdfor JSON chat templates, Python registration, prompt-format mismatch diagnosis, and safe inspection.references/troubleshooting.mdfor OOM, multiprocessing, missing TurboMind extension, stop-token, model download, and trust boundary failures.scripts/inspect_pipeline_config.pyfor no-download inspection of installed API signatures, dataclass defaults, chat-template registry names, and CLI availability.
Scope
Own these tasks:
- Create an offline text pipeline with
from lmdeploy import pipelineandTurbomindEngineConfigorPytorchEngineConfig. - Run single prompts, batches, OpenAI-style messages, streaming output, and multi-turn chat sessions.
- Tune
GenerationConfig,cache_max_entry_count,tp,session_len,max_batch_size, LoRA adapters, and chat templates for inference. - Inspect logits, last hidden states, token lengths, finish reasons, and PPL without designing a server.
- Use
lmdeploy chatfor local interactive text inference and CLI smoke checks.
Route elsewhere:
- Multimodal media loading, image/video prompt construction, and VLM-specific formatting:
vision-language. - REST/gRPC/OpenAI-compatible servers, proxy, API clients, and API-server auth:
serving-apis. - Creating AWQ/GPTQ/SmoothQuant artifacts with
lmdeploy lite:quantization. - New architecture patches, backend internals, custom kernels, or model support development:
backend-extension.
Fast Start
from lmdeploy import GenerationConfig, TurbomindEngineConfig, pipeline
backend_config = TurbomindEngineConfig(cache_max_entry_count=0.6, tp=1)
gen_config = GenerationConfig(max_new_tokens=128, top_p=0.9, temperature=0.7)
with pipeline("org/model-or-local-path", backend_config=backend_config) as pipe:
responses = pipe(["Explain LMDeploy in one sentence.", "List two use cases."], gen_config=gen_config)
for response in responses:
print(response.text)
For PyTorch backend, use PytorchEngineConfig(...). When tp > 1 with PyTorch in a script, wrap pipeline creation in if __name__ == "__main__":.
Validation
Use these checks before recommending or shipping an offline pipeline workflow:
python sub-skills/pipeline-inference/scripts/inspect_pipeline_config.py --include-cli
lmdeploy chat --help
These checks do not download a model. Model execution requires the user’s chosen model weights, backend dependencies, GPU/accelerator memory, and any required trust_remote_code decision.
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).