undefined | Better HN

0 pointshislaziness1y ago0 comments

isn't it 2 bytes (fp16) per param. so 7b = 14 GB+some for inference?

0 comments

This was trained to be run at FP8 with no quality loss.

The model description on huggingface says - Model size - 12.2B params, Tensor type - BF16. Is the Tensor type different from the training param size?

fzzzy1y ago

it's very common to run local models in 8 bit int.

qwertox1y ago

Yes, but it's not common for the original model to be 8 bit int. The community can downgrade any model to 8 bit int, but it's always linked to quality loss.

j / k navigate · click thread line to collapse