Add NPU 910B (Ascend) support for Wan2.2-S2V-14B

[0xPabloLI]

Summary

This PR adds support for running Wan2.2 on Huawei Ascend NPU 910B (32GB HBM + CPU offload), enabling free inference on platforms like AtomGit which provide NPU 910B 64GB instances at no cost.

What's included

• scripts/patch_npu.py — A non-invasive patch script that applies 10 runtime modifications to enable NPU support:

  1. torch.cuda.* → torch.npu.* mapping + import torch_npu
  2. NCCL → HCCL distributed backend
  3. Staged loading: release T5 encoder before loading diffusion model (critical for 32GB HBM)
  4. device_map='auto' for automatic HBM/CPU offload via accelerate
  5. Skip .to(device)/.cpu() for accelerate-managed models (_hf_hook check)
  6. flash_attention → SDPA fallback (NPU has no flash_attn package)
  7. VAE dtype fix (encode input .float() cast)
  8. dtype assert → .float() cast in model_s2v.py
  9. max_memory={0: '15GB', 'cpu': '60GB'} tuning
  10. empty_cache() before VAE decode

• docs/NPU_SUPPORT.md — Complete documentation with setup instructions, patch overview, and test results.

Test Results

Configuration	Value
Hardware	NPU 910B 64GB (AtomGit free tier)
Model	Wan2.2-S2V-14B
Resolution	480×832
Steps	5
Frames	40
Inference time	~550s (9.2 min)
Core-hours	~2.47
Output	480×832 H264, 40 frames, 1.8MB
Cost	Free

Design decisions

• Non-invasive: The patch is a standalone script rather than direct source modifications, making it easy to review and opt-in. Users run python scripts/patch_npu.py before inference.
• Idempotent: All patches check for existing modifications and skip if already applied.
• Single-NPU only: Multi-NPU would require HCCL distributed setup (untested).

Environment

• CANN 8.5, PyTorch 2.9, torch_npu 2.9

---

Tested by @0xPabloLI. Happy to address any feedback!