Regression? Inference gone slow since April (with GPU+CPU offloading)

[Autonom]

My previous compile of llama.cpp was in April. Today, I recompiled with master (b3130) and ran with the same model (miqu-1-70b.q4_k_m.gguf) and there is a significant drop in inference speed 0.5tk/s and stalls after one or two lines of text. Before it was running at 2-4tk/s. That's ~ 4X-8X slowdown.

I reverted to an older version (b2766) and recompiled and the normal speed is back.

Is it accepted that newer features slow inference down or does this qualify as a bug?

[ggerganov]

Without repro instructions this qualifies just as user error. Provide the commands and more information about your system in order to get some support