tomerip
tomerip
Hi @RezaYazdaniAminabadi, I wanted to give an update with additional information: I tested the new v0.6.0 version and in fact now running even with `batch_size=1`, if using fp16 the output...
Hey @RezaYazdaniAminabadi, Was curious if you've had time to look into this issue. Also, please let me know if I can help somehow (although if it's indeed from CUDA files,...
Just wanted to update, issue still exists in v0.6.5
Hi @RezaYazdaniAminabadi, Is the issue now closed due to v0.7.3 being released?
Hey @RezaYazdaniAminabadi, I just tested v0.7.4 and unfortunately I still can't properly run the input mentioned in this issue (although now due to probably a different bug introduced in v0.7.4,...
This was solved by PyTorch in this PR: https://github.com/pytorch/vision/pull/2705 A solution that follows this logic without upgrading PyTorch is replacing this line: `selected = torch.nonzero(choice_cluster == index).squeeze().to(device)` with this line:...
Hey @RezaYazdaniAminabadi, Sorry for the delay, just tested [PR](https://github.com/microsoft/DeepSpeed/pull/2433) and it still outputs junk similar to v0.6.0-v0.7.3 (actually getting "." instead of "!").
Hi @RezaYazdaniAminabadi, are there any expectations regarding a fix? Thanks!