lazy-llm
Guides LazyLLM low-code LLM application workflows including modules, flows, RAG, agents, tools, writer review, CLI, and deployment.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- lazy-llm
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides LazyLLM low-code LLM application workflows including modules, flows, RAG, agents, tools, writer review, CLI, and deployment.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/api-surface-map.md6.3 KB
- 参考文档references/installation-and-extras.md4.6 KB
- 参考文档references/repo-provenance.md3.1 KB
- 参考文档references/repo-routing-metadata.json484 B
- 参考文档references/troubleshooting.md4.5 KB
- 脚本scripts/check_lazyllm_env.py5.9 KB
- 脚本scripts/inspect_lazyllm_surface.py3.8 KB
03 SKILL.md 正文
LazyLLM Repo Skill
Use this skill when a task asks how to build, debug, test, or adapt LazyLLM applications: low-code LLM modules, composable flows, RAG/document pipelines, agents/tools/MCP, writer/review workflows, model deployment, online providers, or the lazyllm CLI.
First actions for any LazyLLM task
- Read Repository provenance if you need to decide whether this skill matches a current checkout or whether a refresh is required.
- Pick the focused sub-skill from the route map below before opening details.
- Check dependency class early. Base LazyLLM imports do not include every optional RAG, agent, model-serving, multimodal, tracing, or provider dependency.
- Prefer the bundled diagnostic scripts over ad hoc probes:
- scripts/check_lazyllm_env.py verifies importability, optional workflow groups, and safe CLI availability.
- scripts/inspect_lazyllm_surface.py prints verified signatures for the main public APIs without running models or network calls.
- Never assume a model, provider key, vector database, parser service, Kubernetes/Slurm cluster, npm MCP server, or GPU backend is available unless the user explicitly provided it.
Route map
| User task or signal | Read this |
|---|---|
Install LazyLLM, inspect extras, use lazyllm CLI, configure lazyllm.config, launcher/service basics, components/prompters, package import failures |
core-runtime |
Use TrainableModule, OnlineModule, OnlineChatModule, ServerModule, deploy/fine-tune models, choose local vs online vs multimodal backend, diagnose provider/model type behavior |
model-deployment |
Compose pipeline, parallel, diverter, switch, ifs, loop, bind, or traceable application graphs |
flow-orchestration |
Build RAG/document workflows with Document, DocNode, readers, transforms, retrievers, rerankers, BM25, vector stores, parser service, Milvus/OpenSearch |
rag-document-processing |
Register function-call tools, use ToolManager, build React/ReWOO/PlanAndSolve agents, use sandbox flags, SQL/HTTP/search tools, SkillManager, MCP |
agents-tools |
Use writer IR/artifacts, writer tools, revision/stream utilities, Feishu adapter boundaries, lazyllm review or local review workflows |
writer-review |
Installation and extras quick guide
Read Installation and optional extras for the full matrix. Typical choices:
python -m pip install lazyllm
lazyllm install rag # document/RAG imports
lazyllm install agent-advanced # MCP/advanced agent dependencies
lazyllm install standard # broad app stack; may install heavy model/vector packages
lazyllm install full # all optional stacks; use only when explicitly needed
For a source checkout, use an editable install with a supported Python (>=3.10,<3.14) and then install only the extras required by the selected workflow. Do not install full just to answer a base CLI/config/flow question.
Minimal no-network smoke check:
python scripts/check_lazyllm_env.py --require-rag --require-agent --require-writer
Run that command from inside this skill directory after copying or locating the script, or pass the script path explicitly from any working directory. It only imports LazyLLM modules and does not call a provider.
High-signal API surfaces
Read API surface map when a task names a class/function or when you need exact routing. Verified public entry points include:
lazyllm.pipeline,parallel,diverter,switch,ifs,loop,bindTrainableModule,OnlineModule,OnlineChatModule,ServerModule,ActionModuleDocument,Retriever,Reranker,DocNode, BM25 and store/index helpersfc_register,ToolManager,SkillManager,ReactAgent,ReWOOAgent,PlanAndSolveAgent- writer artifact models and
WriterToolBase - CLI commands:
install,deploy,run,skills,review,review-local
Backend and safety policy
- Safe by default: import checks, config/CLI inspection, flow graphs with Python callables, BM25/local RAG smoke, writer artifact round trips, tool schema registration.
- Ask or require explicit evidence: cloud provider calls, model downloads, local GPU serving/fine-tuning, multimodal generation, external vector databases, parser-service workers, MCP servers launched through
npx/npm, Kubernetes/Slurm/SCO jobs, PR posting. - Do not post or mutate remote systems from review, MCP, Feishu, database, or provider helpers unless the user asks for that side effect and provides credentials.
Troubleshooting shortcuts
Start with cross-cutting troubleshooting, then the owning sub-skill's troubleshooting file. Common LazyLLM symptoms:
ImportError: Missing package(s): ... You can install them by: lazyllm install rag→ read the RAG or install reference; install the exact extra, notfull, unless the task needs heavy backends.lazyllm --helpexits with a usage error → this CLI prints usage from its dispatcher; use concrete subcommands such aslazyllm skills listor inspect the command table in core-runtime.- Provider/API-key errors → route to model-deployment and keep provider tests optional.
- Flow result mismatch → route to flow-orchestration and verify
bind, skip/kept items, and input tuple handling.
When not to use this skill
Do not use this skill for generic LLM theory, non-LazyLLM frameworks, model architecture research unrelated to LazyLLM modules, or direct production deployment to a user cluster without first checking the model-deployment and core runtime dependency/back-end constraints.
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).