how to enable torch.compile when use Megatron-LM as training backend?

As described, torch.compile seems not compatible with megatron-lm, especially when I train the model with PP/TP/EP etc. How Can I Enable torch.compile for kernel optimization?

No files, tests, or entry points are named. Start by reproducing the training setup with Megatron-LM and the PP/TP/EP combinations described, then investigate the interaction between torch.compile and the training backend. Done means establishing whether kernel optimization can be enabled and documenting or implementing a verified path.

Written by the indexing model from the issue text.