Skip to main content

Several people, one GPU card

HAMi splits the card. gShare places the session, bills it, and takes the card back. All of it from one console.

Get startedUser manual
The administrator dashboard: utilization, GPU devices, node health and storage servers

Interactive GPU sessions

Jupyter, VS Code or a terminal on a slice of a card, opened from the console. No API key and no batch queue to learn.

The session list: mode, GPU spec, uptime, estimated cost and state

Credits, limits, tenants

A session holds credits before it runs, consumes them as it runs, and settles when it ends. Every move is audited.

One control plane, many clusters

Attach a second GPU cluster with one script. Sessions land where their data lives, on storage pools with a measured capacity.

Fractional cards, honest accounting

A tier is a fraction of one card, and the bill follows the occupancy it actually takes. Idle GPUs are paused and handed back.

The session wizard picking a GPU tier, each one showing the share it bills

Look around without a GPU

Postgres, Redis, the API and the console come up together. Running a real session needs a GPU cluster.

$ git clone https://github.com/boanlab/gshare
$ cd gshare
$ make compose-up
$ make smoke

Console on localhost:8000, API on localhost:8080. Tear it down with make compose-down.