I feel there is a catch here. I forgot mixture of experts was a thing -_-. They only activate like 32B at a time.
But this can be incredibly good, lets see!
Btw “and train” people missed the very very key word.
EDIT: @_@ training qwen 3.8 on under 6 GB. Yes, this seems VERY acceptable! I can see the endless potential here.
Nah, it seems it can finetune a 128 GB model with 6 GB RAM, I don’t see evidence that it can finetune something of 1.4T or more. I was excited for nothing.
Call me when something lets me train and run my own 1.4T model locally, in about 120 years.
https://github.com/lyogavin/airllm
Those 120 years went by in a flash.
He is the world’s dumbest man afterall
I feel there is a catch here. I forgot mixture of experts was a thing -_-. They only activate like 32B at a time.
But this can be incredibly good, lets see!
Btw “and train” people missed the very very key word.
EDIT: @_@ training qwen 3.8 on under 6 GB. Yes, this seems VERY acceptable! I can see the endless potential here.
Nah, it seems it can finetune a 128 GB model with 6 GB RAM, I don’t see evidence that it can finetune something of 1.4T or more. I was excited for nothing.
Maybe just another two years, but the trajectory is promising and hardware will improve, as well. Strata is good for inference, at least.