custom_overrides flag in sdk #2794
Description
Describe the feature you'd like
Add a flag that can be passed to the sagemaker-training-toolkit from an estimator that allows a user to provide: custom_override=dict[list,list] (default can be None)
How would this feature be used? Please describe.
This would enable the sagemaker-training-toolkit to have additional functionality where users could override commands and provide their own custom launch scripts. For instance, rather than having to use MPI, I could use torch.distributed.launch, deepspeed, or my fork of any of the above.
Describe alternatives you've considered
You could embed these into hyperparameters but this would be a poor user experience. You could use environment variables, but those don't transfer to local sessions (previously listed as a bug).