Charlelie Laurent comments

Results 19 comments of


                                            Charlelie Laurent

[CorrDiff]: refactor corrdiff configs to dataclasses

@mnabian this overlaps with [824](https://github.com/NVIDIA/modulus-roadmap/issues/824). I already started to work on that so I'll reassign myself.

🐛[BUG] CorrDiff: noise level inconsistent between regression model training and inference

Hi @nbren12 , I am working on that. The line in the inference [inference](https://github.com/NVIDIA/modulus/blob/1f89d93da8cfb0e093bb9ad83f0eadde6c1cd5c9/examples/generative/corrdiff/generate.py#L590) script that you provided is for the **U-Net regression**, which is purely deterministic, so I don't...

🐛[BUG]: Running `corrdiff/generate.py` raises a shape exception

@gideonite @windsoryin The checkpoints distributed through the inference package are not expected to be compatible with the latest physicsnemo training and generation scripts. The right way to use the checkpoints...

🐛[BUG]: CorrDiff - Double nested `model_kwargs` passed to U-Net constructor

I think this issue is solved. I haven't been able to reproduce it.

Enable saving/loading Module compiled with torch dynamo and dynamically set AMP in CorrDiff Module wrappers

/blossom-ci

Enable saving/loading Module compiled with torch dynamo and dynamically set AMP in CorrDiff Module wrappers

/blossom-ci

Enable saving/loading Module compiled with torch dynamo and dynamically set AMP in CorrDiff Module wrappers

/blossom-ci

Enable saving/loading Module compiled with torch dynamo and dynamically set AMP in CorrDiff Module wrappers

/blossom-ci

Added backend context manager to select SDPA implementation

/blossom-ci

🚀[FEA]: Allow passing models as to EDMPrecond

Thanks for reporting @simonbyrne. We will consider your proposed enhancement in the multiple refactorings of the `EDMPrecond` and the underlying models that are underway.