← sources

Splitting Strata across two devices?

source: reddit (reddit:LocalLLaMA)published 2026-10-30 23:03ingested 2026-10-31 02:40

retained text

I’m trying to run Qwen3.8-Flash-Next (125B-A6B) with Strata but trying to see if I can use existing compute instead of spending buckets of money. 

My main machine has.. Ryzen 7 7800X3D, Radeon RX 9070 XT (16 GB), 32 GB DDR5, Ubuntu 26.04. I currently run Qwen3.8-27B (IQ3\_M) on llama.cpp (Vulkan and ROCm 10.1) at about 65 tok/s.

I also have a second PC with 32 GB of DDR4. Is there any way to use that RAM too, or does Strata only work on a single machine? If not, is spreading experts across two machines on a LAN with something like llama.cpp RPC realistic, or too slow?

Any real numbers on tok/s and first-token time with experts in system RAM would help.

document fetched 2026-10-31 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xju6sj/splitting_strata_across_two_devices