Training job runs but loss doesn't decrease, or you rented expensive GPUs but only use 30% capacity—you need to look inside training to see what's happening. SageMaker Debugger and Profiler collect and analyze tensors and system resources in real time, telling you "why it's not working" and "where it's slow".
In the MLA-C01 exam, this topic appears in scenarios like "auto-detect vanishing/exploding gradients", "find why GPU util is low", "auto-stop when training diverges". Today we distinguish two axes: Debugger (model internals) and Profiler (system resources).
First clarify their concerns: