llama.cpp issue #27097: Vulkan slow token generation with AMD dGPU Resizable BAR disabled

Fields

id

sources/noncuda-llamacpp-27097-rebar-vulkan-slow

URL

https://github.com/ggml-org/llama.cpp/issues/27097

Publisher

ggml-org/llama.cpp

Kind

forum

Published

2026-08-15

Accessed

2026-10-03

notes

RX 7900 XTX with ReBAR off in BIOS: Qwen3.8-27B Q4_K_XL decode 13.89 t/s at n=128 and 25.28 t/s at d=4096; setting GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1 raises these to 37.09 and higher. A commenter measured 40 -> ~6 t/s on the same card. A patch that auto-detects a small host-visible VRAM heap is linked (craig-b/llama.cpp 6c1281da3). Issue open, labelled bug-unconfirmed.

Notes

RX 7900 XTX with ReBAR off in BIOS: Qwen3.8-27B Q4_K_XL decode 13.89 t/s at n=128 and 25.28 t/s at d=4096; setting GGML_VK_DISABLE_HOST_VISIBLE_VIDMEM=1 raises these to 37.09 and higher. A commenter measured 40 -> ~6 t/s on the same card. A patch that auto-detects a small host-visible VRAM heap is linked (craig-b/llama.cpp 6c1281da3). Issue open, labelled bug-unconfirmed.

Cross-references

This record references no other records.