# Bf16 or fp16 with bitsandbytes int8?

**URL:** <https://discuss.huggingface.co/t/bf16-or-fp16-with-bitsandbytes-int8/48261>\
**Category:** 🤗Transformers\
**Created:** [July 26, 2023, 3:51pm UTC](https://discuss.huggingface.co/t/bf16-or-fp16-with-bitsandbytes-int8/48261 "2023-07-26T15:51:30Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Penn](https://avatars.discourse-cdn.com/v4/letter/p/ba9def/32.png) [@Penn](https://discuss.huggingface.co/u/Penn)\
**Post date:** [July 26, 2023, 3:51pm UTC](https://discuss.huggingface.co/t/bf16-or-fp16-with-bitsandbytes-int8/48261/1 "2023-07-26T15:51:30Z")

</div>

I’m using bitsandbytes int8 and PEFT with transformers. Is there any advantage to using bf16 with int8 even though it gets cast to fp16 during quantization? Or should this be strictly avoided?