Add a context manager for activation sharding.

[luyug]

What does this PR do?

This pr adds a context manager for sharding (potentially long-living) activation saved tensors across model parallel group. It can be used to wrap any non-mp regions where the activations are duplicated across all ranks in a mp group. One specific and typical use case is to wrap a series of gradient checkpoints of transformer layers, creating a layer input sharding setup similar to the one described in the GSPMD paper.

The implementation uses torch.autograd.graph.saved_tensors_hooks and share similar spirits with torch.autograd.graph.save_on_cpu.

[facebook-github-bot]

Hi @luyug!

Thank you for your pull request.

We require contributors to sign our Contributor License Agreement, and yours needs attention.

You currently have a record in our system, but the CLA is no longer valid, and will need to be resubmitted.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at [email protected]. Thanks!