fix cpu offloading

[samsja]

fix cpu offloading

---

Note

Improves correctness of model init/loading when fsdp_cpu_offload is enabled.

• Use model.to_empty(device="cpu") when CPU offload is on; otherwise use CUDA (in load_dcp_from_hf and setup_model)
• Add _init_buffers_post_meta() helper and call it consistently after meta-to-empty transitions
• Add _move_buffers_to_cuda(model, config) to move buffers to CUDA (FSDP offload only manages parameters) and invoke after random init, DCP load, non-meta load, and delayed checkpoint init
• No functional changes to FSDP/EP/LoRA logic aside from ensuring correct device/buffer placement

^{Written by Cursor Bugbot for commit cc3c3a6. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}