core-runtime
Guides LazyLLM installation, optional extras, CLI commands, configuration, launcher, components, prompters, and base runtime diagnostics.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- core-runtime
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Guides LazyLLM installation, optional extras, CLI commands, configuration, launcher, components, prompters, and base runtime diagnostics.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/config-cli-components.md4.0 KB
- 参考文档references/troubleshooting.md1.9 KB
- 脚本scripts/config_cli_smoke.py2.3 KB
03 SKILL.md 正文
LazyLLM Core Runtime
Use this sub-skill for base LazyLLM setup and runtime questions: package installation, optional extras, CLI command routing, lazyllm.config, namespace behavior, launchers, components/registries, prompters, dependency diagnostics, and safe source-checkout smoke tests.
Start here when
- The user asks to install LazyLLM or choose between base,
rag,standard,full,dev, or backend-specific extras. - An import error names missing optional packages or
lazyllm install <extra>. - The task mentions
lazyllmCLI commands:install,deploy,run,skills,review, orreview-local. - The task concerns
lazyllm.config, environment-backed options,lazyllm.namespace,ComponentBase, component registration, prompt templates, or launcher/service scaffolding. - You need a no-network check before moving to RAG, agents, models, writer, or flows.
Files to read
- config-cli-components.md for CLI command families, config behavior, prompters, components, and launcher notes.
- troubleshooting.md for core install/import/CLI/config failure modes.
- Root Installation and optional extras for extra selection across workflows.
- scripts/config_cli_smoke.py for a safe local smoke check of config, prompt, component registration, and CLI command table assumptions.
Core workflow checklist
- Identify the selected workflow. If the task is only config/CLI/import, keep it here. If it needs a domain feature, route onward after checking dependencies:
- RAG/document import → rag-document-processing
- Agents/tools/MCP → agents-tools
- Model serving/provider/fine-tuning → model-deployment
- Flow composition → flow-orchestration
- Writer/review → writer-review
- Check Python and install state. LazyLLM 0.7.5 expects Python
>=3.10,<3.14. - Run a safe import/CLI diagnostic when environment state is unclear:
python ../../scripts/check_lazyllm_env.py --check-cli python scripts/config_cli_smoke.py - Install the smallest extra named by the failure. For example, RAG import failures usually require
lazyllm install rag, notfull. - Avoid side effects. Core runtime checks should not download models, call providers, start external services, or post reviews.
CLI command families
The dispatcher recognizes these top-level commands:
lazyllm install <extra1> <extra2> <pkg1> ...
lazyllm deploy modelname
lazyllm deploy mcp_server <command> [args ...] [options]
lazyllm run graph.json
lazyllm run chatbot
lazyllm run rag
lazyllm skills init/list/info/delete/add/import/install
lazyllm review ...
lazyllm review-local ...
A bare lazyllm --help can print usage and exit non-zero. Use concrete subcommands such as lazyllm skills list for smoke checks.
Config and namespace facts
lazyllm.configsupports key lookup and mutation with option validation, aliases, default fallback for empty strings, and post-action hooks.- Environment-backed values can be refreshed, and namespace contexts can redirect environment prefixes, for example
lazyllm.namespace("my"). - Tests verify
LAZYLLM_GPU_TYPE,LAZYLLM_DISPLAY, alias normalization, strict options, and namespace isolation behavior.
Component and prompt facts
lazyllm.components.registercan create groups and register callables/classes under lowercase and capitalization-preserving names.fc_registerbelongs to agents/tools, but simple tool registration metadata is often debugged from core runtime imports.Prompter,AlpacaPrompter, andChatPromptergenerate string prompts and OpenAI-stylemessagespayloads with history support.
Review before handing off
Before routing to a heavier sub-skill, leave the user with:
- active Python interpreter and LazyLLM version,
- selected optional extra or reason none is needed,
- safe command to reproduce the import/CLI/config state,
- explicit note if a provider, GPU, model download, external DB, parser service, MCP process, or remote review action is required.
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).