Name and Version

build/bin/llama-cli --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon Graphics (RADV GFX1151) (radv) | uma: 1 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
version: 7684 (53eb9435d)
built with GNU 14.2.1 for Linux x86_64

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-bench, llama-server

Command line

build/bin/llama-bench -m ../models/unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF/Qwen3-Next-80B-A3B-Instruct-UD-Q8_K_XL-00001-of-00002.gguf -p 4096 -ub 256,512,1024,2048,4096 -n 0 -fa 1 -r 3 --mmap 0 -ngl 0,999

Problem description & steps to reproduce

Qwen3-Next behaves a little strange with respect to ubatch sizes. Normally I expect performance to raise up to point, then plateau (or decrease slightly). With this model (and hardware) performance almost halves with ubatch sizes > 512.

Also PP on Vulkan is slower than CPU. The avx 512 path with repacking seems to work very well.

First Bad Commit

No response

Relevant log output

No response