AMD Strix Halo user builds llama.cpp Laguna on ROCm and hits flash-attn limits

dbinnunE3 · reddit · 2026-07-22

A Reddit user documents a full build-and-benchmark attempt for the Laguna fork of llama.cpp on an AMD Strix Halo box using ROCm/HIP.

What they did

Built the poolsideai/llama.cpp laguna branch inside a Fedora toolbox container.
Used GPU passthrough and installed ROCm from Fedora’s native repos.
Had to add several missing -devel packages during CMake configuration.
Fixed a source bug in common/speculative.cpp by adding the missing <cmath> include.

What worked

The GPU was detected correctly as an AMD Radeon 8060S / gfx1151 device.
A llama-bench run without flash attention or KV quantization succeeded.
On a laguna-s-2.1-Q4KM.gguf model, the benchmark reported about 305.54 t/s for pp512 and 18.44 t/s for tg128.

What failed

A flash-attention run crashed because the HIP kernel had no compatible device code for HIP arch 1300.
Turning flash attention off broke quantized KV cache support.
Trying DFlash through llama-server later failed because dflash requires ctxother to be set.

The post is essentially a detailed field report on getting a large local model stack running on AMD hardware and where the current rough edges are.

More from coding & agent