llama.cpp user asks whether MoE experts can be streamed from SSD instead of RAM

Ok_Warning2146 · reddit · 2026-07-23

Can llama.cpp offload non-shared MoE experts to SSD?

The poster asks whether llama.cpp can go beyond its current -cmoe setup, which loads non-shared experts of MoE models into RAM while keeping other weights and KV cache on GPU VRAM.

The question

They want to know if the same idea can be extended so that:

Why it matters

The motivation is simple: RAM is expensive, and MoE models keep getting larger. The question is essentially about more aggressive tiered memory management for local inference.

Original post →

More from coding & agent

coding & agent channel →