[Subodh-17]

Summary

Fixes #49297 by preventing FSDP2 non-rank-0 processes from materializing the entire model onto the host CPU during from_pretrained.

Root Cause

FSDP1 and FSDP2 both export identical distributed environment variables (ACCELERATE_USE_FSDP=1). When loading missing keys from meta devices in _move_missing_keys_from_meta_to_device(), transformers assumes this implies FSDP1 and iterates over all parameters to allocate torch.zeros_like(param, device="cpu"). For FSDP2—which natively broadcasts meta-tensors from rank-0—this is entirely redundant and causes immediate Host OOMs for large models.

Changes

• Localized an explicit FSDP2 lifecycle check within the CPU-materialization block in src/transformers/modeling_utils.py using accelerate.state.is_initialized().
• For FSDP2, the loop now materializes ONLY non-persistent buffers (preserving existing fallback behavior) and safely returns early to leave parameters and persistent buffers on meta.
• Added a regression test specifically mocking PartialState to ensure FSDP2 paths leave weights on the meta device.
• The global is_fsdp_enabled() helper remains strictly generic to avoid breaking upstream accelerate dispatch paths.

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 37201087984:2
Result: success | Jobs: 16 | Tests: 196,923 | Failures: 0 | Duration: 16h 11m

[Rocketknight1]

Existing PR at #45649 so I'm going to try to wake that one up instead