> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gpuhub.com/llms.txt
> Use this file to discover all available pages before exploring further.

# PyTorch for Blackwell

> Using RTX PRO 6000 / RTX 5090 GPUs with PyTorch

<Note>
  When using RTX 5090 or RTX PRO 6000 (**Blackwell architecture**), please use the latest images and install PyTorch as follows.
</Note>

<img src="https://mintcdn.com/gpuhub/VFlaWwIH5B1lMPUz/images/pytorch280.png?fit=max&auto=format&n=VFlaWwIH5B1lMPUz&q=85&s=0580cc0ff46bfb0eaa4c364f7674ba01" alt="" width="2326" height="905" data-path="images/pytorch280.png" />

1. RTX 5090 and RTX PRO 6000 require CUDA ≥ 12.8 and PyTorch ≥ 2.7.1 to detect and use the GPU at all.
2. **ONLY PyTorch ≥ 2.8.0 nightly builds currently support** multi-GPU (DataParallel / DDP) on these cards.
3. Strongly recommended to install the [nightly version](https://pytorch.org/get-started/locally/#nightly-builds) yourself for now.

<img src="https://mintcdn.com/gpuhub/VFlaWwIH5B1lMPUz/images/pytorch-nightly.png?fit=max&auto=format&n=VFlaWwIH5B1lMPUz&q=85&s=97e8b83d3a8456a8cc98a13173be8851" alt="" width="2560" height="1300" data-path="images/pytorch-nightly.png" />