Skip to main content
Basalt Compute

Contribute model ports

Last updated

Models that run well on alternative accelerators need engineers to port and tune them. Basalt Compute does that work in the open, in a public registry on GitHub.

How it works

  1. A bounty names the target. Maintainers list open bounties in the registry, with the target hardware, the numerical parity a port must meet, and the latency and throughput it must reach.
  2. You submit a bring-up package. A port is a small overlay on a pinned engine, submitted as a pull request, not a monolithic image.
  3. It is checked on the target hardware. A maintainer-triggered run builds the image, pulls the model weights and runs security scans, parity checks against a reference baseline and performance benchmarks on real hardware.
  4. Passing work is published. It is merged with a signed commit, and the image is published with an immutable digest attributed to the contributor.

What this is for

The point of the programme is to make more open models run well on more kinds of accelerator, with the work and the checks visible to everyone. It does not need an account on this site.

What is not here yet

A bounty board, submission tracking and anything beyond the registry itself are not part of this site. Everything happens on GitHub today.

If you port or tune models and want to talk, tell us what you work on. Use the contact form on the "Model porting" track. Email is the only required field; a GitHub or GitLab handle is optional. We reply by email.

Talk to us