Thorsten Kurth

cuda::cuda_driver not linked to aluminum

4

When I am trying to build lbann via Spack, it does not link the app with -lcuda, just with the runtime. Checking the SetupCUDAToolkit file shows that it does not...

REQ conditioned on context values

1

Hi all, I am wondering if it is possible to create checks which are conditional on context values. For example for the HPC deepcam benchmark, the optimizers which can be...

This PR refactors the RL code for more flexibility. This was necessary for adding on-policy routines, which is also done with PPO along with all C and Fortran wrapper routines....

### Is this a new feature, an improvement, or a change to existing functionality? Improvement ### How would you describe the priority of this feature request Critical (currently preventing usage)...

🚀[FEA]: Print communicator layout when using verbose=True in DistributedManager.create_groups_from_config

2

### Is this a new feature, an improvement, or a change to existing functionality? New Feature ### How would you describe the priority of this feature request Critical (currently preventing...

🐛[BUG]: Order of communicators in comm tree is not in the order as they were added

1

### Version 0.5.0 ### On which installation method(s) does this occur? Pip ### Describe the issue I have the following problem: although I am adding process groups in a specific...

fixing member reference

2

# PhysicsNeMo Pull Request ## Description This fixes a wire error with new torch, because the wrong member variable is referenced in init process group, so the except branch is...

This MR fixes a bug in the distributed spherical resampling routine. It also adds more thorough tests for distributed convolution as well as resampling.

Add support for gradient clipping.

6

This MR adds gradient clipping to all training procedures.

This adds a small update to distributed SHT to reproduce the exact same behavior as serial SHT