Has HIP performance gotten worse for everyone else?

[Diablo-D3]

Using llama-cli.exe -hf unsloth/Qwen3.6-27B-MTP-GGUF:Q5_K_XL -ngl all --fit on --no-mmap --spec-type draft-mtp --spec-draft-n-max n --cache-type-k f16 --cache-type-v f16 --spec-draft-type-k f16 --spec-draft-type-v f16 -fa on (and omitting --spec-* when ntp) on a 7900XTX in Windows 11. All test cases fit model and context into VRAM.

HIP, fa on

draft	pp	tg
ntp	106.1 t/s	19.9 t/s
mtp n=1	130.6 t/s	25.7 t/s
mtp n=2	186.2 t/s	11.0 t/s
mtp n=3	183.9 t/s	12.2 t/s
mtp n=4	172.1 t/s	14.0 t/s

vulkan, fa on

draft	pp	tg
ntp	660.3 t/s	27.4 t/s
mtp n=1	553.1 t/s	52.3 t/s
mtp n=2	592.4 t/s	62.8 t/s
mtp n=3	605.7 t/s	66.4 t/s
mtp n=4	549.6 t/s	59.2 t/s

fa off

draft	pp	tg
hip ntp	125.7 t/s	23.6 t/s
vulkan ntp	337.3 t/s	25.7 t/s
vulkan mtp n=3	502.7 t/s	61.3 t/s

[Ezzz-dev]

ROCm uses more memory than Vulkan, I have your same setup, and you're also trying to run with F16, I can't imagine the amount that it's trying to fit, you must define a specific context size manually, under that space with that model, I'd say you can only fit about 30k context at much.

Also, llama.cpp has a problem with Windows where its performance throttles down over time until you make a restart ( not saying this is the case with you ).

[Diablo-D3]

Vulkan, afiak, currently has a memory tracking bug. So, I don't think it actually uses less RAM, it just doesn't account for MTP correctly. See #24159. All of the test cases are at least 8k context and do fit.

I'm aware of the Windows problem, but it is complex and hard to reproduce. I rebooted the machine before performing this test just to make sure I wasn't currently being effected by it.