gpu-smi (Python wrapper)

Friendly Python access to gpu-smi — stdlib only, no compiler, no dependencies.

You still need the gpu-smi binary itself (this package calls it for you). Windows: it's bundled — pip install is all you need. Linux: get it from releases or cargo build --release, then put it on PATH or set GPU_SMI_BIN (a Linux binary will be bundled in a later release the same way).

pip install gpu-smi
# from source: pip install ./python

Beginner (no JSON needed)

import gpu_smi

for gpu in gpu_smi.gpus():
    print(gpu.summary())
    # [1] AMD Radeon RX 9070 XT | VRAM 2513/16304 MB (15%) | 39C | 22% util | 8.2W

gpu = gpu_smi.first()          # single-GPU shortcut (raises NoGpuError if none)
print(gpu.name, gpu.temp_c, gpu.util_percent)
print(gpu.vram_free_mb, gpu.usage_ratio)
for p in gpu.processes:        # pid, name, mem_mb, kind ("C"/"G"), is_compute
    print(p.pid, p.name, p.mem_mb)

python -m gpu_smi prints the same report with zero code (--raw for JSON).

Power user

import gpu_smi

rows = gpu_smi.query_raw()                          # raw list[dict], untouched
for snap in gpu_smi.watch(interval=0.5, count=10):  # polling generator
    ...

# Long-running HTTP API (one persistent process, Prometheus included):
with gpu_smi.serve(port=8080, token=True) as api:   # token=True -> auto 256-bit token via mode-600 tempfile
    print(api.gpus()[0].summary())
    print(api.metrics())                            # raw Prometheus exposition text

api = gpu_smi.ServeClient("http://ml-box:8080", token=open("token.txt").read().strip())
api.health()  # {"status": "ok"} — open even with auth on

Binary lookup order: explicit arg > GPU_SMI_BIN env > PATH > bundled gpu_smi/bin/. Errors: BinaryNotFoundError, QueryError, AuthError (bad bearer token), NoGpuError — all subclass GpuSmiError.

License

MIT, same as gpu-smi.