PipelineModule inflated checkpoints when using FP16 param flattening

Hi,

When using the pipeline module/engine, checkpoints are saved per layer within each module/stage. But when using one of the FP16 optimizers, each layers checkpoint size is the full size of the whole pipeline stage (all layers in the stage), which causes checkpoint sizes to be quadratic in the number of layers per stage.

The issue seems to be related to the FP16 optimizers using one large flattened param buffer for the whole param group, but pytorch saves the whole tensor even when only saving a view of it.

Currently, I can overcome this by passing per-layer param-groups to the optimizer:

params = [{"params": [p for p in layer.parameters() if p.requires_grad]} for layer in net.forward_funcs]

But this doesn't seem super clean. Is there a cleaner way I missed to handle this? And if not maybe worth adding a method to PipelineModule to retrieve the parameters in groups per layer?

Start by reproducing the pipeline module checkpoint behavior with an FP16 optimizer and inspect how per-layer parameters and flattened parameter buffers are serialized. Done means checkpoints no longer duplicate the full pipeline-stage buffer for each layer, while preserving correct checkpoint loading.