Change main weights precision to bfloat16
#5
by
Molbap
HF staff
- opened
This changes the base weights to be bfloat16 by default instead of float32.
Molbap
changed pull request title from
Upload MistralForCausalLM
to Change main weights precision to bfloat16
you dont need to as you can use the unsloth in ntraining to do this with the trainer !
FP16 is the correct format : when loading in transformers you can specify to chage it there while loading also using bits and bytes !