The only legit PR I can find is this [0] and it's still open.
There's currently a lot of rejected vibe-coded PRs: [1] (violation of AI policy).
The OP's PR says it was generated with Claude Code so it has a very low chance of getting merged upstream.
[0] https://github.com/ggml-org/llama.cpp/pull/21089
[1] https://github.com/ggml-org/llama.cpp/pulls?q=Turboquant+is%...
I find it quite exciting to read some results in an effort to understand if TurboQuant main ideas can be applied to model weights. There are other similar projects, so we’ll see, but it seems some of this fork results look promising.