dsh-wsl-gpu

The subtler failure mode in WSL2 local inference is not a wrong model pick but resource contention. On a single 16 GB-class RTX card, running Ollama, llama-server, and vLLM side by side means any extra few hundred MB of VRAM can push the stack into an OOM, and the crash often does not surface until the next large GGUF load. gpu_doctor is a diagnostic built exactly for that moment: it pulls GPU visibility, VRAM headroom, live utilization, and inference port occupancy into one report so you know how much the card can still take before you hit swap.

Under the hood it parses each GPU row for name, driver version, VRAM used vs. total, and compute capability; adds Blackwell / RTX 50-specific hints around sm_120 and CUDA 12.8+ or 13.x compatibility; scans the common inference ports 11434, 1234, 8000, and 8080 to flag whether Ollama, vLLM, or Unsloth Desktop are competing for the same device; and hands off to companion tools like host_reach or docker_doctor focus=vllm when the bottleneck sits at the host network or container layer.

It is aimed at developers doing local LLM inference on a Windows + WSL2 stack with a single NVIDIA card, particularly those who juggle Ollama, vLLM, and llama.cpp, or who hit load failures after a driver update or a CUDA rebuild. Run it once after a driver update, once after a CUDA build failure, and once before loading another large model, and a two-second output tells you which engine to close or whether it is time to reach for a bigger card.