llama.cpp issue #27761: ROCm flash-attention only works with q8_0/q8_0 and q4_0/q4_0 KV cache

Fields

id

sources/noncuda-llamacpp-27761-rocm-fa-kv-graph-splits

URL

https://github.com/ggml-org/llama.cpp/issues/27761

Publisher

ggml-org/llama.cpp

Kind

forum

Published

2026-08-26

Accessed

2026-10-03

notes

RX 7900 XTX (gfx1100), build 10639 win-rocm-7.14. q5_0 in either K or V explodes the graph from 2 splits to 34 and moves ~550 MiB of compute to the CPU host buffer: pp512 957.11 -> 300.10 t/s (about -68%). The same build on Vulkan handles q5_0 with 2 splits and no regression. Labelled bug-unconfirmed and stale.

Notes

RX 7900 XTX (gfx1100), build 10639 win-rocm-7.14. q5_0 in either K or V explodes the graph from 2 splits to 34 and moves ~550 MiB of compute to the CPU host buffer: pp512 957.11 -> 300.10 t/s (about -68%). The same build on Vulkan handles q5_0 with 2 splits and no regression. Labelled bug-unconfirmed and stale.

Cross-references

This record references no other records.