Preserve requires_grad in SimpleFSDP parameter registration

[sylvesterkaczmarek]

Fixes #3885.

Summary

SimpleFSDP re-registers each distributed parameter by wrapping it in nn.Parameter(...). Because nn.Parameter defaults requires_grad=True, a parameter that was intentionally frozen before SimpleFSDP wrapping becomes trainable again.

This change passes the original parameter's requires_grad value when re-registering the distributed parameter, preserving both frozen and trainable parameter state.

Changes

• preserve p.requires_grad when constructing the replacement nn.Parameter in data_parallel()
• extend the existing single-rank SimpleFSDP test to cover a model with a frozen weight and trainable bias
• verify both the underlying registered parameters and parametrized attribute access preserve their original requires_grad values

The behavior of already-trainable parameters is unchanged.

Validation

Added test_preserves_parameter_requires_grad to torchtitan/experiments/graph_trainer/tests/test_simple_fsdp.py. It exercises the existing public SimpleFSDP application path at world size 1 with the CPU/gloo test setup already used by this file.

The test suite and pre-commit were not run locally in this environment because the repository is not available as a local checkout here. No local test result is claimed; CI should run the updated test file.

[tianyu-l]

not sure if the graph trainer error is introduced by this change

[sylvesterkaczmarek]

not sure if the graph trainer error is introduced by this change

Thanks. I checked the current run. Lint passed, while the GraphTrainer integration job is the only failing signal I found. I don't have Actions write permission on the upstream repo to rerun it.

Could you rerun the failed job so we can confirm whether this is transient before changing the approved one-line fix?