CPU-MoE still helps with mmap. Should not overly hurt token-gen speed on the Mac since the CPU has access to most (though not all) of the unified memory bandwidth, which is the bottleneck.

For sure I was running on autopilot with that reply. Though in Q4 I would expect it to fit, as 24B-A4B Gemma model without CPU offloading got up to 18GB of VRAM usage

No - this model has the weights memory footprint of a 35B model (you do save a little bit on the KV cache, which will be smaller than the total size suggests). The lower number of active parameters gives you faster inference, including lower memory bandwidth utilization, which makes it viable to offload the weights for the experts onto slower memory. On a Mac, with unified memory, this doesn't really help you. (Unless you want to offload to nonvolatile storage, but it would still be painfully slow.)

All that said you could probably squeeze it onto a 36GB Mac. A lot of people run this size model on 24GB GPUs, at 4-5 bits per weight quantization and maybe with reduced context size.

I bet the poster just didn’t remember that important detail about Macs, it is kind of unusual from a normal computer point of view.

I wonder though, do Macs have swap, coupled unused experts be offloaded to swap?

correct but it should be some ratio of model size like if model size is x GB, max context would occupy x * some constant of RAM. For quantized version assuming its 18GB for Q4 it should be able to support 64-128k with this mac