torpedo

A persistent HTTP telemetry server for Tenstorrent cards. It keeps one device-discovery context alive, samples `tt-smi` 6.3.0 telemetry on a background thread, and serves the latest snapshot as JSON at `/v1/devices` with a `/healthz` check, rediscovering devices with exponential backoff after a reset. It marks a device as held when other processes appear in `/proc/driver/tenstorrent/<n>/pids`, and reads allocated GDDR passively from tt-metal's `/dev/shm` allocator region without opening a Metal context. Written in Python with `luwen` (default) and `umd` backends and a systemd unit; the repo also includes `watchgpu`, a shell script that shows CUDA, Tenstorrent and Apple Silicon hosts side by side over SSH.