1

github code: https://github.com/bellowman/Deep-Learning-Practice/blob/main/BioBert%20for%20Multi%20Label%20AMD.ipynb

Hello everyone,

I am a beginner with pytorch, tensorflow, and BERT. I have a machine at home with an AMD Ryzen 7 1800x and a Radeon RX 6600 video card.

I am trying to run a bioBERT model at home. I have trouble leveraging my model to use my AMD card. I posted my github notebook. I have troubles in cell 3 and 9.

  1. First Question: In cell 3,I am trying to convert the bioBERT weight to PyTorch with transformmer-cli. I get the warning of "Could not load dynamic library 'cudart64_110.dll'". Does this affect performance later?
  2. Second Question: In cell 9, My model load is really slow because it is using just the CPU. How can I get the model to run on my AMD GPU
Dung Tran
  • 357
  • 1
  • 2
  • 13
  • 2
    Well you can't load the cudart dll because you don't have the right graphics card for that. To use PyTorch with AMD you need to follow [this](https://pytorch.org/blog/pytorch-for-amd-rocm-platform-now-available-as-python-package/). Another option is just using google colab and loading that ipynb and then you won't have those issues. – Chrispresso Dec 08 '21 at 18:16
  • Yes, I am familiar with Google Colab, but I do want to make this work with AMD. Was wondering if anyone knows the exact code change or steps to make it work with AMD – Dung Tran Dec 08 '21 at 18:36
  • 1
    Did you install PyTorch with the ROC-m package? The code in the cell might not be the problem, it is most likely that you have the wrong PyTorch installed. – Chrispresso Dec 08 '21 at 19:04
  • Ahh, it seems like I have to do this on linux to make this work. Thank you for the advices. – Dung Tran Dec 09 '21 at 07:57

1 Answers1

1

Thank you to chrispresso

AMD ROCm seems to be the way to go, but it requires one to run under linux

Dung Tran
  • 357
  • 1
  • 2
  • 13