Gpu Facts

This host's accelerators as swamp data — every GPU (vendor, name, VRAM, driver, compute capability), the total VRAM, system RAM, usable CPU cores and the compute runtime (CUDA / ROCm / Metal / none) — read straight off the machine with no network and no SSH. WHAT IT DOES

Answers "can this model run here, and how?" without guessing. It probes the machine's own accelerators (NVIDIA via nvidia-smi, AMD via rocm-smi, Apple via system_profiler) and writes one gpu resource: each device's VRAM and driver, the total VRAM, plus system RAM and CPU cores. A consumer — a Hugging Face recommender choosing a quantisation, or an Ollama Modelfile renderer setting num_gpu and num_ctx — reads those fields instead of hard-coding them, so the same workflow adapts to a 4090 box, a Jetson or a Mac. Prefer it over shelling out to nvidia-smi inside a workflow: it parses every vendor's output into one stable, aligned shape. RAM is read via Deno's systemMemoryInfo (which works under swamp's method sandbox) with /proc/sysctl/free fallbacks. A GPU whose VRAM is not queryable — NVIDIA GB10 / Jetson unified memory, where nvidia-smi reports N/A — records totalVramMiB: 0; pass an explicit vramMiB budget to a consumer in that case.

INSTALL

swamp extension pull @svendowideit/gpu-facts

DEPENDENCIES

None. It runs nvidia-smi, rocm-smi, system_profiler, sysctl or nproc only when they are present, reads /proc/meminfo on Linux, and needs no network, no credentials and no other extension.

RUN

# Detect this host's accelerators and record them (auto-creates the model
# instance `gpu-host` on first use). Run once per host; the `gpu` resource
# is what a recommender or Modelfile renderer reads.
swamp model @svendowideit/gpu-facts method run collect gpu-host

# Gate before collecting: report whether a gpu record already exists.
swamp model @svendowideit/gpu-facts method run status gpu-host

CONFIGURE

Global arguments (set at creation with --global-arg, or per call with --input):

vendorOverride  auto | nvidia | amd | apple | none  (default auto)
deviceIndex     Keep one 0-based device index       (default -1 = all)

UPGRADE

This extension installs nothing on the host, so there is no on-host state to reconcile — upgrade the code alone and collect records fresh facts on its next run:

swamp extension update @svendowideit/gpu-facts

WHAT IT INSTALLS

Nothing on the host — no services, daemons, cron triggers or webhooks. It reads local commands/files and writes swamp resources (gpu, status).