Name and Version

I was able to narrow it down to as of commit 7acb4e8cd2ce21f457d1298e75fad729520d263c, prefill performance with unsloth/GLM-5.1-GGUF:UD-Q3_K_XL when using CUDA 13.3 and ZenDNN (Zen5) drops by ~50%.

~250-260 prefill drops to ~120-130 prefill.
Decode appears to be impacted as well, less so.

Operating systems

Linux

GGML backends

ZenDNN

Hardware

2 x Epyc 9115
1 x RTX Pro 6000
2 x RTX 5090

Models

unsloth/GLM-5.1-GGUF:UD-Q3_K_XL

Problem description & steps to reproduce

Run commit 7acb4e8cd2ce21f457d1298e75fad729520d263c, compare performance with commit 3ecfb150a4bd2d92b2a7974bb1af954c8a5e2985.

First Bad Commit

7acb4e8cd2ce21f457d1298e75fad729520d263c

Relevant log output

...

...