Home/Contact
Project · Contact
Contact
Corrections, contributions, measured benchmark results, and questions the FAQ does not cover.
Send a message
Direct
Corrections
Include the page URL and, if you can, a minimal reproduction. A single layer with a surprising error number is worth more than a paragraph of description.
Benchmark results
We will publish measured results with a reproducible method and full configuration disclosure. We will not publish vendor-supplied numbers without one.
Writing for the site
Guides are welcome, especially on topics we have avoided: activation quantization, rotation methods, and low-precision training with microscaling formats.
What we cannot help with
Library-specific API questions are better asked of that library. We keep this site deliberately free of API surface that changes every release.
Before you write about a quality regression
Four of these five catch the problem before you need us at all.
- Confirm the bf16 baseline on the same harness and seed
- Confirm blocks run along the reduction axis
- Confirm the exclusion list covers head, embeddings, norms and routers
- Confirm a double round trip is a no-op
- Have the per-layer relative Frobenius error report to hand