# Is there an alternative to base llama.cpp for RDNA2 gpus in general with local LLM usage ? — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/088ac40e0d748a19

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-19 13:06 ingested 2026-10-20 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1x9we92/is_there_an_alternative_to_base_llamacpp_for/)

## retained text



```
Most of the stuff built for amd gpus posted here or ai in general depends on WMMA which exists only from RDNA3 and up. In image-video generation I can bypass it with triton some of the time. But here, a thing works or not on rdna2. Something specific like strata for example has a rdna2 path which worked with the coder model I've tried but that is only for one specific huge model which has limited usage. 

In short, do people have apps - forks other than base llama.cpp which performs better with RDNA2 / on windows if possible/
```



document fetched 2026-10-20 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1x9we92/is_there_an_alternative_to_base_llamacpp_for
