Add opt-in foreach path to GPA-AdamW step

[shaneraphel]

Summary

• GPAAdamW.step can now run the parameter update with multi-tensor foreach ops on a single device, instead of only the per-parameter Python loop.
• The arithmetic is the same as the loop. Sparse grads, mixed devices, or mixed dtypes fall back to the loop.
• Default behavior does not change: foreach defaults to off so existing runs are unaffected. It is opt-in because the win is shape-dependent.
• The flag is not stored in the param group, so existing checkpoints keep their keys.

Evidence

Apple M4, PyTorch 2.11.0, CPU, 10 threads, 8 steps, median of 3 runs. foreach result is bit-identical to the loop (torch.equal on params, z-buffer, exp_avg, exp_avg_sq).

shape	loop	foreach
200 x 1024	2.587 ms	1.727 ms	1.50x faster
48 x 768 x 768	41.121 ms	63.859 ms	0.64x, slower

The 768 case is why this is opt-in rather than a new default.

Test plan

• python -m unittest gpa.tests.gpa_adamw_test gpa.tests.gpa_equivalence_test (20 tests, new GPAAdamWForeachTest checks bit-identical params and state over 4 steps)
• ruff check and ruff format on the two touched files
• GPU run of the new test (make test-gpu needs CUDA hardware I do not have here)

Prepared with an AI assistant. The table was measured on this machine before the pull request was opened. The author of the commit is shanyu.

[meta-cla]

Hi @shaneraphel!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

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 cla@meta.com. Thanks!

[meta-cla]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks!