Salomi, a research repo on extreme low-bit transformer quantization

12 points
1/21/1970
a day ago
by Edward9055

Comments


Edward9055

I’ve been working on a research repository called SALOMI focused on extreme low-bit transformer quantization and inference.

The main thing I’d emphasize is that the most useful result was not a clean “1-bit works” claim. The stronger takeaway was that correlation-based reconstruction metrics can look promising while end-to-end perplexity still collapses, and that strict bits-per-parameter accounting changes a lot of early sub-1-bit conclusions.

The repo includes:

quantization and runtime code tests and evaluation tooling backend/runtime experiments historical drafts and corrected findings

I tried to preserve both the promising directions and the places where the original story did not hold up under stricter evaluation.

Would especially appreciate feedback from people who have worked on quantization, inference systems, or evaluation methodology.

Repo: https://github.com/OrionsLock/SALOMI

a day ago

rdos

Was any text in the repo NOT written by AI?

a day ago

Edward9055

I used AI tools during development, same as most people writing code right now. The research direction, experiments, and conclusions are mine I read the papers, designed the experiments, ran them, and documented where things broke. The repo includes 60+ experiment iterations, result logs showing failures, and documentation that corrects earlier optimistic claims. That's not a pattern you'd get from prompting a model to generate a project. I'm one person, so yes, AI helped with implementation. The research was mine.

14 hours ago

kevmo314

Was this shoveled out with Claude/Codex to try to ride off the Bonsai release?

a day ago