EN

Recipes

Agent: llm-action + chitu, lmdeploy, opencode, vLLM

README 링크에서 발견 · 2 days ago

Software & DevOps

스택: chitu, llm-action, lmdeploy, opencode, vllm

What they did

이 레시피는 chitu, llm-action, lmdeploy, opencode, vllm을 함께 묶습니다. 이들이 registry README의 상호 링크에서 함께 등장하기 때문입니다. 이는 유지관리자(또는 커뮤니티)가 이들을 하나의 스택으로 취급한다는 신호입니다. 아래에 링크된 앵커 저장소(chitu)의 README에서 시작하세요. 지원 도구: llm-action, lmdeploy, opencode, vllm. 각 도구의 runchr 페이지에서 설치한 다음, 앵커 프로젝트의 문서를 따라 오케스트레이션을 구성하세요.

How it went

통합 작업을 예상하세요: README 링크는 동시 출현을 보여줄 뿐, 바로 사용할 수 있는 튜토리얼은 아닙니다. 각 도구를 업스트림 문서에 따라 설치하고 구성하면 일관된 스택을 얻게 됩니다.

Worth knowing

여기서의 연결은 검증된 프로덕션 실행이 아니라 링크 패턴에서 추론한 것입니다. 의존하기 전에 자신의 repo에서 인증, 시크릿, 모델 선택을 검증하세요.

Wiring

레지스트리 도구 간 README 상호 링크에서 유추한 내용입니다. 각 도구를 설치한 뒤, 연결 세부 사항은 기본 리포지토리의 링크된 README를 따르세요.

Try it yourself

I can't fetch the source link or review the anchor README from here, so I can't tell you which of those tools I "already use" — that would be a guess, not a fact. Here's what I can do honestly:

**What I can't confirm**
- The contents of the anchor README (no link was included in your message, and I have no browsing in this context).
- Which of chitu, llm-action, lmdeploy, opencode, vllm are already in use — that depends on your environment, not mine.

**Minimal wiring plan (tool-agnostic, adjust once the README is in hand)**

*Install order (dependency-first):*
1. Python env + `uv` or `pip` (base for vllm, lmdeploy, chitu).
2. `vllm` — primary inference server; install first so you have a working OpenAI-compatible endpoint.
3. `lmdeploy` — only if you need its quantization/Turbomind path; install after vllm to avoid CUDA/torch version clashes.
4. `chitu` — install last among inference tools; it pins specific torch/CUDA builds and is the most likely to conflict.
5. `opencode` — client-side; install after the server is stable.
6. `llm-action` — CI/automation layer; install last, once the endpoint and client both work.

*Env vars (minimum set):*
- `OPENAI_BASE_URL=http://localhost:8000/v1`
- `OPENAI_API_KEY=sk-local` (dummy for local vllm)
- `HF_HOME=/path/to/models`
- `CUDA_VISIBLE_DEVICES=0`
- `VLLM_LOGGING_LEVEL=INFO`
- Tool-specific: `LMDEPLOY_*`, `CHITU_*` — fill in from each project's README once confirmed.

*First end-to-end task:*
Serve one small model via vllm, hit it from opencode with a single prompt, and assert the response in a `llm-action` step — i.e., "prompt → model → structured output → CI check passes." That validates the whole chain before adding lmdeploy/chitu.

If you paste the README text or the link contents, I'll reconcile this plan against it and tell you exactly which tools are already covered vs. new.

Source

English

Tools in this stack