Exposes DDP training capabilities using the Sagemaker Distributed Data Parallel Library.
Data Ingestion
Three ways to get files into Sagemaker:
| Approaches | Description |
|---|---|
| File Mode | Simplest option; read file from S3. Downloads dataset into instance memory |
| Fast File Mode | Streamed into instance in real-time |
| EFS | |
| FsX Lustre |
When using Fast File Mode, specifying a distribution allows Sagemaker to know which partition of the training file to use for each node.