dsh-wsl-gpu
Install set: part of dsh-wsl-kit. Prefer KIT_SET=daily | llm | github | full (see kit README). Fault tree: TROUBLESHOOTING.md.
DeepSeek Harness tool: gpu_doctor — WSL nvidia-smi, VRAM pressure, Blackwell/5080 hints, and whether Ollama / vLLM / Unsloth Desktop ports compete on one GPU.
Part of dsh-wsl-kit.
中文说明 → README.zh.md
Compatibility
| Field | Value |
|---|
| Plugin | dsh-wsl-gpu 0.2.1 |
| Minimum dsh | ≥ 0.1.2 (web UI one-shot ?token= on Windows relay :3081) |
| Latest verified | See dsh-wsl-kit Compatibility (currently 0.1.5-rc.1) — single source of truth for the suite |
| Kit set | llm / full (some also useful alone) |
| Cloud Flash | Use model id deepseek-flash (V4.1 Flash) in ~/.dsh/settings.yaml / llm-deepseek — not configured by this plugin |
| Agent Teams | Upstream experimental; not required here |
Suite floor versions: kit check-plugin-versions.sh. Fault tree: TROUBLESHOOTING.md.
Why
Local inference needs the Windows NVIDIA driver to expose GPUs into WSL2. On a single ~16GB card (e.g. RTX 5080), opening Ollama and llama-server and vLLM at once is a common OOM path. This tool reports visibility, VRAM, and open inference ports together.
Install
curl -fsSL https://raw.githubusercontent.com/173787247/dsh-wsl-kit/master/install.sh | KIT_SET=llm bash
# or:
dsh plugin --profile web add github:173787247/dsh-wsl-gpu
Ask: “Run gpu_doctor” after driver updates, CUDA build failures, or before loading another large GGUF.
What you get
- Parsed GPU rows: name, driver, VRAM used/total, util, compute capability
- Blackwell / RTX 50 tips (
sm_120, CUDA 12.8+/13.x)
- Inference port scan:
11434 / 1234 / 8000 / 8080
- Pointers to
host_reach and docker_doctor focus=vllm
Config
- id: dsh-wsl-gpu
name: dsh-wsl-gpu
config:
timeoutMs: 20000
probeTimeoutMs: 1200
probeInference: true
Test
npm test
License
MIT
