# Voxl Pytorch Support

Source: https://forum.modalai.com/topic/706/voxl-pytorch-support
Category: General Questions (https://forum.modalai.com/category/2/general-questions)
Posted: 2022-02-03 10:02:43 UTC by Philemon Benner
Replies: 4 · Views: 706

## Philemon Benner · 2022-02-03 10:02:43 UTC

Hey,
Pytorch supports by default CUDA(Nvidia Gpu) devices, but VOXL has a Qualcomm Adreno 530 wich is a AMD Gpu. In the Pytorch Docs there is  also a Support for AMD Gpu´s with Rocm. Does anybody know if it's possible to use Rocm on Voxl or another way to enable GPU for Pytorch? 
PS. I know Modal AI has a tflite Server for tensorflow Models but we already have a Model trained on Pytorch and Inference with CPU is to slow.

## Reply by Eric Katzfey (ModalAI staff) · 2022-02-03 16:58:27 UTC

@Philemon-Benner This doesn't answer your question, but Adreno is a GPU designed by Qualcomm, not AMD.

## Reply by Philemon Benner · 2022-02-03 17:20:11 UTC

Thank you for your fast response. Good to know that...

## Reply by Guest · 2022-02-04 18:22:04 UTC

@Philemon-Benner  In the tflite-server, we use the OpenCL backend to enable GPU support. However, according to a pytorch dev, "As Tensorflow already tentatively supports OpenCL, albeit through a proprietary compiler framework, it looks like Tensorflow is the way to go for anyone with an AMD card or, er, any CPU." 

That said, it may be a bit easier to convert your pytorch model over to a tflite model that can run with gpu acceleration. A good resource for this process is [this repo](https://github.com/sithu31296/PyTorch-ONNX-TFLite), and if you are interested in some other tflite models this [model zoo](https://github.com/PINTO0309/PINTO_model_zoo) has many preconverted options. 

If you were to follow this conversion process, be sure to read up on [this section](https://docs.modalai.com/voxl-tflite-server/#using-your-own-models) in our documentation!

## Reply by Philemon Benner · 2022-02-07 05:46:38 UTC

Thank you for your Answer yeah i will definitely look into using Tensorflow Lite