Mixture of Experts Explained (opens in new tab)

(huggingface.co)

29 pointsosanseviero2y ago2 comments

2 comments

The thing I’m confused about is if MoE is supported by Metal? More specifically, can I run the 8x7B on a MacBook (with enough RAM, which I believe is 48GB required?)?

leetharris2y ago

There's nothing about MoE that is incompatible with Metal.

There is a prototype GGUF implementation (https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF) that is out this morning, but it requires a branch that hasn't been merged into the main llamacpp release.

In fact, the #1 challenge with MoE is increased RAM usage, which is something that the unified memory of Macs specializes in.

j / k navigate · click thread line to collapse

2 comments

syntaxing2y ago

The thing I’m confused about is if MoE is supported by Metal? More specifically, can I run the 8x7B on a MacBook (with enough RAM, which I believe is 48GB required?)?

leetharris2y ago

There's nothing about MoE that is incompatible with Metal.

In fact, the #1 challenge with MoE is increased RAM usage, which is something that the unified memory of Macs specializes in.

j / k navigate · click thread line to collapse