Skip to content

About

Who

llama-cmpv is built and maintained by Joseph Stackhouse at Stack-Tech.

What this is

A research fork of llama.cpp carrying a small series of patches that make NVIDIA CMP 100-210 mining cards usable for LLM inference, plus the full record of the experiments behind them.

It is not a distribution and not a product. It is a patch series meant to be rebased onto upstream repeatedly, kept small enough that rebasing stays possible, with activation behavior documented per mechanism. Some patches use runtime selectors; others change a supported source specialization.

What it is not

Not a fork of upstream's direction

The fork preserves upstream llama.cpp source and its build system. Source modifications also live in the affected upstream files; the complete patch series and public documentation live under cmp/. Build, publishing and research tooling remain private.

Not benchmarks of llama.cpp

Nothing on this site is a claim about llama.cpp's performance in general. Every number is measured on one unusual machine — Volta silicon behind a Gen1 x1 link with no peer path — and several of the mechanisms here would be worthless or negative on ordinary hardware. That is stated on the pages, not buried in a footnote.

Licence and attribution

The patches are offered against upstream llama.cpp under its MIT licence. Upstream's copyright and licensing are unchanged by anything in this repository.

This documentation is authored by Joseph Stackhouse, Stack-Tech. If you use the published findings and their documented limitations, an attribution link is appreciated.

Contributing

The rules for anyone working on this fork, human or agent, live in a single rules file in the working repository, kept in one file so they cannot drift between copies. Patches here are offered upstream where they are general enough; where they are specific to these cards they stay here.

Contact

Issues and pull requests on the repository are the best route. github.com/Josephur