Fix: handle remaining meta tensors in from_single_file before dispatch

[kitsune-hash]

Summary

Fixes #12009

When loading models from single-file checkpoints (e.g., GGUF format), some parameters or buffers may not be present in the checkpoint and remain on the meta device after load_model_dict_into_meta. This causes dispatch_model to fail with Cannot copy out of meta tensor errors.

Root Cause

In from_single_file(), models are created under init_empty_weights(), which initializes all tensors on the meta device. After loading checkpoint weights via load_model_dict_into_meta, any tensors not present in the checkpoint remain on meta. This includes:

1. Non-persistent buffers like RoPE sinusoidal embeddings (freqs_cos, freqs_sin in WanTransformer3DModel) — these are computed deterministically during __init__() and intentionally excluded from checkpoints
2. Missing parameters due to incomplete key mapping in checkpoint converters

Fix

Three layers of handling before dispatch_model():

1. Non-persistent buffer re-materialization: Identifies submodules with non-persistent meta buffers and re-creates them outside init_empty_weights() context, allowing deterministic buffers to be properly recomputed
2. Persistent buffer fallback: Zero-initializes any persistent buffers still on meta with a warning
3. Parameter safety net: Zero-initializes remaining meta parameters with a warning (indicates incomplete key mapping)

Testing

Validated with:

• WanTransformer3DModel.from_single_file() with GGUF checkpoints (Q2_K, Q4_K_M, Q8_0)
• Confirmed RoPE buffers are correctly recomputed (cos/sin values in [-1, 1], 100% match with reference model)
• Full T2V inference produces valid video output
• 0 meta tensors remaining after fix

Specific models tested:

• QuantStack/Wan2.2-T2V-A14B-GGUF (1095 params, 2 non-persistent buffers recomputed)
• QuantStack/Wan2.2-I2V-A14B-GGUF (safety net correctly catches 150+ missing I2V-specific params from incomplete key mapping)

Before this fix

RuntimeError: Cannot copy out of meta tensor; no data!

After this fix

INFO: Re-creating submodule 'rope' (WanRotaryPosEmbed) to materialize non-persistent buffers.
✅ 0 meta tensors remaining, model loads and runs correctly