From Simon Willison: “Mistral tweet a link to a 281GB magnet BitTorrent of Mixtral 8x22B—their latest openly licensed model release, significantly larger than their previous best open model Mixtral 8x7B. I’ve not seen anyone get this running yet but it’s likely to perform extremely well, given how good the original Mixtral was.”
That’s huge, I’m guessing we’ll need to use a giant swap file?
You’re right, but the model is also not quantized so is likely to be in 16bit floats. If you quantize it you can get substantially smaller models which run faster though may be somewhat less accurate.
Knowing that the 4 bit quantized 8x7B model gets downscaled to 4.1GB, this might be roughly 3 times larger? So maybe 12GB? Let’s see.Edit: sorry those numbers were for Mistral 7B, not mixtral. For Mixtral, the quantized model size is 26GB (4 bits), so triple that would be roughly 78 GB. Luckily, being an MoE, not all of it has to be loaded simultaneously to the GPU.
From what I recall, it only uses 13B parameters at once, so if we compare that to codellama 13B, quantized to 4 bits, that is 7.4GB, so triple that would be 22GB, so would require a 24GB GPU. Someone double check if I misunderstood something.
24GB GPUs include the AMD 7900 XTX and the nvidia RTX 4090 (Ti), non-mobile.
I thought MoEs had to be loaded entirely in the (V)RAM and the inference speedup was because you only need to use a fraction of layers to compute the next token (but the choice of layers can be different for each token, so you need them all ready; or keep moving data between the disk <-> RAM <-> VRAM and get reduced performance).