I'm running Windows 11 on my desktop, which has an NVIDIA RTX 3070 GPU. I'm working on an NLP task using Hugging Face's AutoModelForSequenceClassification and I want to utilize my GPU for training. I've successfully installed PyTorch 1.9.0 with CUDA 11.1 and confirmed that CUDA is available on my system.
However, when I try to run my script, I encounter an ImportError suggesting that I need to install the accelerate library. When I attempt to do so, it not only fails but also replaces my existing PyTorch 1.9.0 installation with version 2.1.0. I've tried various commands like pip install transformers[torch] and pip install accelerate -U, but they all result in the same issue.
The error message also indicates that accelerate requires at least PyTorch 1.10, but I can't find a compatible CUDA version for my RTX 3070.
Does anyone have a solution for running a proper installation of transformers + torch + accelerate?
I have installed PyTorch on multiple combinations (OS+Hardware).
I have installed PyTorch successfully using those commands (in a virtual environment):
%pip install --upgrade transformers
%pip install --upgrade torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
%pip install accelerate (will take latest at the time of writing @0.23.0)
%pip install evaluate datasets
These helped me kick-start any of the projects which require HuggingFace. I hope it helps you.