Show HN: GPU Kill – A CLI tool to kill stuck GPU jobs without rebooting (github.com)

gpukill (gpu-kill) is a lightweight Rust CLI that bills itself as a cross‑vendor alternative to nvidia-smi: it lists and watches GPU usage, memory, temperature and processes; lets you gracefully (or forcefully) kill GPU processes; reset individual or all GPUs; filter and batch operations; and audit usage history with JSON export and reports. It supports NVIDIA (requires NVIDIA drivers), AMD (ROCm), Intel (intel-gpu-tools) and Apple Silicon (macOS M1–M4), and runs on Linux, macOS and Windows. Typical workflows shown include real‑time monitoring, killing stuck training jobs, batch‑killing Python processes, exporting stats for logging, and querying per‑user/process audit history. Technically, gpukill is a cargo-built Rust tool with vendor-specific backends that call existing low‑level utilities (nvidia-smi, rocm‑smi, intel_gpu_top, system_profiler). It stores audit data locally (~/.local/share/gpukill or macOS app support), may require sudo for privileged ops, and exposes automation-friendly features (filters, --watch, --output json) suitable for ML clusters and CI. Important implications: it can simplify GPU ops in heterogeneous environments and enable reproducible usage auditing, but GPU resets and force kills require caution and proper drivers/permissions. The project is MIT/FSL licensed and open to contributions.