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

Where it sits

Reports nvidia-smi, Blackwell hints, and whether an inference port is already taken.

flowchart LR
  agent["dsh agent"] --> tool["gpu_doctor"] --> gpu["nvidia-smi / VRAM"]

Suite diagram and version snapshot: dsh-wsl-kit. This plugin is 0.2.2 (full; also in llm). Do not copy that matrix into this README.

Compatibility

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