gpu-lock-client 0.3.0

Async and sync client + CLI for the gpu-lock service.

gpu-lock-client

Python client + CLI for the gpu-lock service — a FIFO priority mutex that coordinates shared GPU access between processes and Docker containers.

pip install gpu-lock-client

Quick use — CLI

export GPU_LOCK_URL=http://192.168.10.13:8090

# Acquire GPU 0, run a command, release on exit (sets CUDA_VISIBLE_DEVICES).
gpu-lock run --gpu 0 -- python train.py

# Let the server pick the least-busy GPU.
gpu-lock run --gpu auto -- bash

# Interactive priorities run ahead of the normal queue.
gpu-lock run --gpu auto --priority high -- python infer.py

# Low-priority batch work yields to everyone else.
gpu-lock run --gpu 0 --priority low -- python nightly_job.py

# Inspection
gpu-lock status
gpu-lock queue --gpu 0

Quick use — Python

from gpu_lock_client import gpu_lock, gpu_lock_sync, Priority

# Async
async with gpu_lock("my-service", gpu=0, ttl=300, priority=Priority.HIGH) as lease:
    if lease:
        os.environ["CUDA_VISIBLE_DEVICES"] = str(lease.gpu)
    ...

# Sync
with gpu_lock_sync("my-service", gpu="auto", ttl=300) as lease:
    ...

The context manager automatically heart-beats the lease in the background, so long-running jobs don't need to pick a huge TTL.

Passthrough mode

If GPU_LOCK_URL is empty or unset, every client call is a no-op: functions return None, context managers yield None, the CLI's run just execs the command. This lets the same code run on single-GPU dev machines without a gpu-lock service.

Auth

Set GPU_LOCK_TOKEN — the client sends it as Authorization: Bearer.

See the project README for the full protocol, server setup, and priority semantics.