# Changing bnb\_4bit\_compute\_dtype

**URL:** <https://discuss.huggingface.co/t/changing-bnb-4bit-compute-dtype/98121>\
**Category:** Beginners\
**Created:** [July 18, 2024, 10:29am UTC](https://discuss.huggingface.co/t/changing-bnb-4bit-compute-dtype/98121 "2024-07-18T10:29:58Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![ChawkiHjaiji](https://avatars.discourse-cdn.com/v4/letter/c/b3f665/32.png) [@ChawkiHjaiji](https://discuss.huggingface.co/u/ChawkiHjaiji)\
**Post date:** [July 18, 2024, 10:29am UTC](https://discuss.huggingface.co/t/changing-bnb-4bit-compute-dtype/98121/1 "2024-07-18T10:29:58Z")

</div>

I have a limited computation ressources and I want to use the LLaMA-3 model. So to reduce memory usage and increase computation speed, I want to indicate that computations will use the minimum bit floating point precision.I want to know is it possible to use torch.float8 instead of torch.float16 in  
`compute_dtype = getattr(torch, "float16")`