Maximilian Stadler
Maximilian Stadler
As far as I know,`DataEdgeAttr` is part of the more recent idea of connecting the `Data` class to the `GraphStorage` class [(see PR)](https://github.com/pyg-team/pytorch_geometric/pull/4816/files#diff-46e4e6f655447088ed5da6e3517b0417723c19498cf31d0c28c781bea7e7d551R328). This definitely shouldn't be in the tagged...
@mnabian I added a first version of the README. Could you have a look at that and some of the remarks Akshay brought up?
https://github.com/NVIDIA/modulus/pull/171 forgot to update a few multi-gpu tests. I forgot that I ran into the same issue when working on #249 and fixed it there. Since #243 could be merged,...
@mnabian Since you are revisiting GraphCast now, adding a few comments - Can we add the option to use `transformer_engine.LayerNorm`? In AIFS benchmarks, we just could get a 1.3x end-to-end...
> @stadlmax as far as I remember, we were using fused layernorm and that gave us nice speedup (although I can't find it in the most recent code)... Did you...
> This is great comparison, thanks! I'll switch to te then. Do we have any reason to still keep fused layernorm from apex, or we should just remove it? I...