I just wanted to say I really appreciate this project!

[Drael64]

I've got the pre-compiled rocm version of llama.cpp, and I compiled this with 6.4.2 and community libraries (for gfx1032), and whereas llama.cpp actually kind of blows on my card using rocm (arguably worse than vulkan somehow), this backend roars. I get 600 t/s ish PP, and 20-30 t/s for gen on a 8b qwen3 (maybe a little slower than vulkan on kobold, but the PP difference is large), which honestly feels more in the realm of how text gen is supposed to feel, despite me having a potato gpu. I think this is in part due to how the optimisations are exposed/work in kobold, although probably the HIP version plays in as well.

Anyway, thank you. Genuinely.

[gloowa]

I'd like to join in on the thanks. I would love to contribute, but I don't have much experience with C/C++

<3

Any chance for a new release to bring in latest changes from koboldcpp? I would love to try out gemma 4...