Increase shared memory for local mode

[bengruher]

Currently, SageMaker local mode for model training does not allow for adjusting smh-size. Defaults to 64MB which is not enough for most use cases. Does not allow for model to be trained on custom hardware using SageMaker SDK.

[yinweisu]

Encountered this problem and modified the source code to make it work...

Basically you'll want to add 'shm_size': 'Xgb', to the host_config dict linked below
https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/local/image.py#L741-L749

Will plan to send a PR later. But if you got this problem and you have access to the source code. ^^should do the trick.

[mbahri]

Bumping this thread: would it be possible to make this configurable? It's important when testing multi-gpu training locally (PyTorch DDP)