gpusched
github.com/ceruleane/gpusched- Category
- Developer Tools
- Rank
- No. 2913Tools index
Previous survey · No. 2805 ·
- Type
- TOOL
- Use case
- Models: Train & Run
- Latest release
- v0.4.0
- Date
About
gpusched is a GPU job scheduler built to run multiple model-training jobs on a single GPU without the out-of-memory crashes that occur when jobs are launched uncoordinated. It also exposes an interface meant for AI agents to dynamically request and allocate GPU resources so they can train models on their own.
At a glance
gpusched dispatches shell jobs to GPUs with enough reported free memory and records how much memory the jobs used. It is worth trying when you queue many experiments on one machine and estimate their memory requirements by hand.
Good fit
Researchers using one NVIDIA GPU host for a batch of training or evaluation jobs.
Consider something else when
Use a cluster scheduler when you need multi-user isolation, quotas or scheduling across hosts.
Start here
Read the worked example and start with two small jobs with known memory needs. Compare placement and reported peaks before sharing a GPU between larger jobs.
Cost and requirements
The project documents a Python tool with no Python dependencies. GPU placement relies on NVIDIA tooling and commands supplied in a job file.
Limits to know
Memory declarations inform placement. They do not cap the memory a process can allocate. Polling can also miss short spikes, so keep headroom or use exclusive scheduling for the tool’s own jobs.
Sources checked 2026-11-05. Source-based assessment.
Why it made the leaderboard
It gives coding agents a scriptable interface to allocate and schedule GPU training jobs without manually managing OOM-prone concurrent runs.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.