Kalsa
Engines / SelfRef / SelfRef/llama.cpp-rdna

SelfRef/llama.cpp-rdna

AMDForkMIT

Shrinks Qwen3.8-27B with the ROCmFPX format and decodes it faster, in a llama.cpp fork for AMD GPUs. It is a fork of ggml-org/llama.cpp.

What it does better

The ROCmFPX 4-bit format makes Qwen3.8-27B 14.63 GiB instead of 16.69 GiB for unsloth UD-Q4_K_XL, a 12% saving, on a Radeon 8060S

Trade-off: perplexity on wikitext-2 rises from 5.8965 to 5.9842 (+1.5%, the error bars overlap).

Source

Benchmarks

  • On Radeon 8060S (Strix Halo), RADV, idle GPU with Qwen3.8-27B (ROCmFPX-MQ-Q4 (Q4_0_ROCMFP4)), it decodes at 12.56 tok/s, against 11.71 tok/s for unsloth UD-Q4_K_XL (+7%). Source Note: Token generation at 128 tokens, from the author's working notes dated 2026-08-21. A different weight format against the usual one on the same model and machine; the notes do not state which engine ran the baseline, so the baseline engine is not stated. Perplexity on wikitext-2 rises from 5.8965 to 5.9842 (+1.5%, the error bars overlap).

Listing

Listed: Passed all seven checks.

Security review

Not reviewed yet.

Install

To install it, follow the project's own instructions on GitHub. The engine runs with your user's permissions. macOS doesn't sandbox it, so install only engines you trust.

Something wrong or out of date? Write to help@kalsa.io, or ask us to remove this page.