feat(fsdp2): load model on meta device for non-rank0 ranks (0 CPU RAM per worker)

[cben484]

Problem

With --fsdp fsdp2, transformers 5.x has no rank0-only zero-memory loading path. In from_pretrained, the FSDP non-rank0 branch (transformers/modeling_utils.py, _move_missing_keys) materializes every parameter with torch.zeros_like(param, device="cpu") — i.e. each rank allocates the full model in CPU RAM (a 35B bf16 model ≈ 67G per rank, ~136G peak including prepare-time copies). On hosts that budget CPU RAM per device (e.g. 125G per accelerator), N x 136G inevitably OOM-kills workers during loading/prepare, regardless of world size.

Measured on 16x Ascend 910B (2T host RAM, Qwen3.5-35B-A3B): 16 workers x ~136G anon RSS ≈ 2.2T → kernel global OOM kill during prepare (dmesg evidence captured).

Solution

Only rank 0 needs real weights; accelerate's cpu_ram_efficient_loading path (accelerate/utils/fsdp_utils.py: "If cpu_ram_efficient_loading is enabled, only rank 0 loads the weights") already distributes rank-0 weights to all ranks during prepare. This PR makes non-rank0 ranks load on the meta device and keeps the model there until accelerate takes over:

• module-level helper _fsdp2_use_meta_loading() — enabled when FSDP2 + world_size>1, disable via SWIFT_FSDP2_META_LOADING=0
• in get_model_processor(): non-rank0 loads under init_empty_weights(), with ACCELERATE_USE_FSDP temporarily masked during the loading window

Two details that matter:

1. Masking ACCELERATE_USE_FSDP inside the loading window is essential: transformers' non-rank0 branch would otherwise turn meta params back into CPU zeros (full model in host RAM per rank), silently defeating the optimization. device_map is resolved earlier in the same function, so masking at this point is safe.
2. Do NOT add a custom weight-sync hook after prepare: we prototyped a per-parameter broadcast(p.to_local(), src=0) hook on Trainer._prepare_for_training and it deadlocks HCCL collective ordering (HcclAllGather dispatch timeout), because accelerate's own sync already runs inside prepare. The official path handles everything — this PR deliberately contains no sync code.

Verification (16x Ascend 910B 64G, ms-swift 4.4.2, Qwen3.5-35B-A3B LoRA)

Metric	Before (full load per rank)	After (meta loading)
Host CPU RAM peak (loading+prepare)	~2.2T for 16 ranks → kernel OOM kill	187G total, independent of world size
Per-worker host RAM (non-rank0)	~136G	~0 (meta device)
NPU HBM per card	28.3G @ 8 cards	19.5G (DPO) / 20.9G (KTO) @ 16 cards
Correctness (DPO)	loss start 0.6914 (=ln2); logps/chosen -128.8	loss start 0.6914; logps/chosen -129.0 (matches full-load baseline)

DPO and KTO smoke tests both pass end-to-end at full 16-card world size.

Requires #9980 (ACCELERATE_USE_FSDP fix) as prerequisite — without it, device_map='npu:{rank}' forces full per-device NPU loading before this code path is even reached.