Closed WailordHe closed 2 months ago
Till now, we only tested our code on A100. Sorry for it.
Hi, thanks for letting us know this. Please let me clarify a little more for this issue. Here there are actually two issues.
I'm sorry for this incovenience. We implemented our custom PyTorch operators based on the computation resources of our machines, so it is a little hard to let it support all of the GPUs.
Thank you @xiaomengy and @violet-zct, I've successfully installed on A100 GPUs. Do you have any simple example training codes? The pseudo code seems to be missing some necessary elements.
Hi @WailordHe,
For the training code, the main missing component is data loader, which we are not allowed to share due to license. My suggestion is to merge your own data loader to this codebase. Thanks.
Failed to install on V100 with PyTorch 2.0.1 and cuda 11.7, is it because V100 doesn't support bf16, or is it a problem with ninja?