Update: llama.cpp for Radeon VII / MI50 / MI60 — +14% PP, +9% long-context fill vs upstream + adaptive Flash Attention
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一个基于gfx906的llama.cpp新分支，专为Radeon VII / MI50 / MI60设计，显示出显著的性能提升。与当前的llama.cpp主线版本相比，该分支在首批PP处理速度上提高了14.1%（379.2 vs 332.3 t/s），在120k上下文填充速度上提高了9.3%（252.6 vs 231.1 t/s）。深层上下文TG性能基本持平，为13.6 vs 13.5 t/s，而DFlash接受度保持不变，为0.691。该分支还确保了确定性输出的字节级匹配。
I posted a new gfx906 based llama.cpp fork a few days ago. One of the main points of critique was that i did not provide sufficient numbers for the gains to be achieved.
TL;DR: After switching our Qwen 3.8 27B production setup to DFlash2, several of the old gfx906 optimizations turned out to be neutral or outright regressions. We went back through the existing gfx906 work, isolated the problem areas, reworked the small-Q Flash Attention path and added adaptive native/convert selection.
Against current llama.cpp mainline, the resulting fork is now +14.1% in first-batch PP (379.2 vs 332.3 t/s) and +9.3% in 120k-context fill (252.6 vs 231.1 t/s), while deep-context TG is effectively tied at 13.6 vs 13.5 t/s. DFlash acceptance is identical at 0.691, and deterministic output matches byte-for-byte.
Our thread is here:
https://forum.level1techs.com/t/glm-and-i-created-a-llama-cpp-fork-optimized-for-amd-gfx906-mi50-mi60-radeon-vii-gcn-hip/254257/3
This is the github for it:
https://github.com/milpster/gfx906-llama-cpp