Guard torch.distributed submodule imports for ROCm builds

[0xDELUXA]

Problem

On Windows ROCm via TheRock, the torch._C._distributed_c10d C-extension is not shipped. torch.distributed itself is present as a stub, but importing any submodule that touches c10d bindings - torch.distributed.fsdp, torch.distributed._tensor, torch.distributed.tensor.parallel - raises at import time:

ModuleNotFoundError: No module named 'torch._C._distributed_c10d'; 'torch._C' is not a package

Wan2GP hits this at startup because models/wan/any2video.py unconditionally imports shard_model from models/wan/distributed/fsdp.py, which pulls in torch.distributed.fsdp at module top. Full traceback ends at:

File "Wan2GP/models/wan/distributed/fsdp.py", line 5, in <module>
    from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
...
ModuleNotFoundError: No module named 'torch._C._distributed_c10d'

This blocks Wan2GP from starting at all on AMD + TheRock, even for single-GPU inference where distributed is never used.

Fix

Wrap the affected top-level imports in try/except (ImportError, ModuleNotFoundError), falling back to None when the submodule cannot be loaded. Behavior on CUDA and on ROCm builds that ship c10d is unchanged - the imports succeed and the bound names are identical.

Changed files:

• models/wan/any2video.py - guard import torch.distributed as dist and from .distributed.fsdp import shard_model.
• models/wan/multitalk/multitalk.py - guard import torch.distributed as dist. Imported lazily from any2video.py, so this path only triggers when multitalk features are used, but fails for the same reason.
• models/kandinsky5/kandinsky/models/parallelize.py - guard torch.distributed._tensor and torch.distributed.tensor.parallel imports.

Plain import torch.distributed as dist statements elsewhere in the repo (hyvideo, nanovllm, mmaudio, TTS, magi_human, longcat, radial_attention) were not changed - the bare namespace imports succeed on TheRock; only submodule imports that touch c10d fail.

Verification

Reproduced on torch 2.12.0a0+rocm7.13.0a (TheRock), Windows 11, Python 3.12:

• Before: python wgp.py crashes at import with the traceback above.
• After: Wan2GP starts successfully. shard_model, dist, and the kandinsky tensor-parallel symbols resolve to None; no code path in single-GPU inference references them.
• On CUDA PyTorch (verified separately): no behavior change - try-blocks succeed and bindings are identical.

Resolves

• Issue with new pytorch wheels #1210
• outdated and missing steps manual for AMD installation (currently status: NOT working) #1362
• unclear and not working installation tutorials #1357

cc @deepbeepmeep @Tophness

[deepbeepmeep]

thx, but I have removed all distributed code in the wan model, please update and let me know if there is some left that crashed on AMD

[0xDELUXA]

Great! Closing this as resolved by ce961de