mtmd
mtmd
@WilliamTambellini Have you tried calling `FP16` `cublasgemm()` with `n=1`? That should address the issue.
Thank you @xuqiantong. > Can you please include only the changes to Conv2D and DynamicScaler in this first version of PR? Done. This decreases the performance of training resenet-34 from...
Thank you @xuqiantong! Sounds good. > Without changing the FL dataset pipeline, I think it still worths to keep your changes to the DistributedDataset, where transformations are performed after prefetch....
> @mtmd — we'll get this merged in pretty soon - there are some broader changes to abstractions that will be helpful here to clean this up. > > Would...
@joazoa I might be able to help. However, I need to reproduce it first. Can you please provide a detailed instruction (+ the corresponding recipe) for reproducing the issue?
@joazoa Thank you for sharing all these details. I am interested in reproducing this bug, and that's the first step for fixing it anyways. > Can I PM you with...