TL;DR
When you install PyTorch on the ROCm platform for an AMD GPU, a pytorch-triton-rocm package will be installed along with it. If you also install Triton at the same time, you may encounter a version conflict, which can be tricky to resolve.
According to PyTorch, you can use the following commands to install PyTorch for the ROCm backend using pip:
1 | pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.1 |
In addition to the packages listed in the command, a pytorch-triton-rocm package will also be installed. This package is essentially a selected version of Triton, so the actual content of this package(in site-packages) is triton/, not pytorch-triton-rocm/.
In this case, if you then install a standalone Triton package, there is a potential conflict because both pytorch-triton-rocm and triton packages target the same triton/ directory.
For example, you can’t actually uninstall triton due to dependency issues. Instead, you will need to uninstall both pytorch-triton-rocm and triton to resolve this.
Hopefully this will help anyone in the same situation.