← sources

q38fn on mixed gpus, which engine?

source: reddit (reddit:LocalLLaMA)published 2026-10-30 12:16ingested 2026-10-31 02:40

retained text

So i am running 2 radeon vii and a rtx 3080. Currently i am using strata which does not support mixed backends, so my 3080 is just expensive swap space.

Is there any inference engine that supports fast q38fn over mixed backends like rocm/cuda/vulkan but is also fast? Llama.cpp is way too slow sadly.

document fetched 2026-10-31 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1xjdrs7/q38fn_on_mixed_gpus_which_engine