gpu-lease

Original：🇺🇸 English

Translated

Use before running local GPU workloads such as PyTorch, SGLang serving, Ray clusters, CUDA benchmarks, or scripts that use GPUs; instructs agents to wrap commands with gpu-lease run so CUDA_VISIBLE_DEVICES is set through a lease.

1installs

Sourcereyoung/gpu-lease

Added on2026-06-25

NPX Install

npx skill4agent add reyoung/gpu-lease gpu-lease

SKILL.md Content

View Translation Comparison →

GPU Lease

Use this skill before running local GPU workloads from Codex or another code agent.

GPU workloads include PyTorch training or inference, SGLang serving, Ray workers or clusters, CUDA benchmarks, and scripts that import GPU frameworks or launch GPU-serving processes.

Workflow

Use the machine daemon through the default socket
```
/var/run/gpu-lease.sock
```
. Do not start a new daemon for routine GPU work. Do not pass
```
--socket
```
or set
```
GPU_LEASE_SOCKET
```
unless the user explicitly provides another socket.
Wrap every GPU command with
```
gpu-lease run
```
. By default, request the number of GPUs you need with
```
--count
```
and include
```
--wait
```
so the command starts when GPUs are ready:
bash
```
gpu-lease run --count 2 --wait -- python train.py --batch-size 8
```
Use exact GPU IDs only when the user specifically requires fixed devices:
bash
```
gpu-lease run --ids 0,1 -- python train.py --batch-size 8
```
Let
```
gpu-lease run
```
own
```
CUDA_VISIBLE_DEVICES
```
. Do not set it separately unless you are intentionally composing with another scheduler.
Keep the GPU process as the direct child of
```
gpu-lease run
```
. The lease is released when that command exits.

Examples

bash

gpu-lease run --count 1 --wait -- python -m torch.distributed.run --nproc_per_node=1 train.py
gpu-lease run --count 2 --wait -- python -m sglang.launch_server --model-path ./model
gpu-lease run --count 4 --wait -- ray start --head --num-gpus=4