Hacker News Reader

what is the ballpark vram / gpu requirement to run this ?

For just the model itself: 4 x params at F32, 2 x params at F16/BF16, or 1 x params at F8, e.g. 685GB at F8. It will be smaller for quantizations, but I'm not sure how to estimate those.

For a Mixture of Experts (MoE) model you only need to have the memory size of a given expert. There will be some swapping out as it figures out which expert to use, or to change expert, but once that expert is loaded it won't be swapping memory to perform the calculations.

You'll also need space for the context window; I'm not sure how to calculate that either.

I think your idea of MoE is incorrect. Despite the name they're not "expert" at anything in particular, used experts change more or less on each token -- so swapping them into VRAM is not viable, they just get executed on CPU (llama.cpp).

A common pattern is to offload (most of) the expert layers to the CPU. This combination is still quite fast even with slow system ram, though obviously inferior to a pure VRAM loading

I think your understanding of MoE is wrong. Depending on the settings, each token can actually be routed to multiple experts, called experts choice architecture. This makes it easier to parallelize the inference (each expert on a different device for example), but it's not simply just keeping one expert in memory.

Reply on news.ycombinator.com