App/Demo exception since fix: use weights_only=True (CWE-502)

[bluemoehre]

Latest commit (303b283) is causing exception on startup when using the official ROCm image docker.io/rocm/pytorch:rocm7.2.3_ubuntu24.04_py3.12_pytorch_release_2.10.0.

[vibevoice] | You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
[vibevoice] | ERROR:    Traceback (most recent call last):
[vibevoice] |   File "/opt/venv/lib/python3.12/site-packages/starlette/routing.py", line 638, in lifespan
[vibevoice] |     async with self.lifespan_context(app) as maybe_state:
[vibevoice] |   File "/opt/venv/lib/python3.12/site-packages/fastapi/routing.py", line 241, in __aenter__
[vibevoice] |     await self._router._startup()
[vibevoice] |   File "/opt/venv/lib/python3.12/site-packages/fastapi/routing.py", line 4884, in _startup
[vibevoice] |     await handler()
[vibevoice] |   File "/app/demo/web/app.py", line 353, in _startup
[vibevoice] |     service.load()
[vibevoice] |   File "/app/demo/web/app.py", line 126, in load
[vibevoice] |     self._ensure_voice_cached(self.default_voice_key)
[vibevoice] |   File "/app/demo/web/app.py", line 164, in _ensure_voice_cached
[vibevoice] |     prefilled_outputs = torch.load(
[vibevoice] |                         ^^^^^^^^^^^
[vibevoice] |   File "/opt/venv/lib/python3.12/site-packages/torch/serialization.py", line 1548, in load
[vibevoice] |     raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
[vibevoice] | _pickle.UnpicklingError: Weights only load failed. In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
[vibevoice] | Please file an issue with the following so that we can make `weights_only=True` compatible with your use case: WeightsUnpickler error: 
[vibevoice] | 
[vibevoice] | Can only SETITEMS for dict, collections.OrderedDict, collections.Counter, but got <class 'transformers.modeling_outputs.BaseModelOutputWithPast'>
[vibevoice] | 
[vibevoice] | Check the documentation of torch.load to learn more about types accepted by default with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.
[vibevoice] | 
[vibevoice] | ERROR:    Application startup failed. Exiting.

Reverting to previous commit (e73d1e1) solves the issue.

[pengzhiliang]

Thanks for reporting this. I can reproduce the issue.

The current bundled voice preset .pt files contain BaseModelOutputWithPast objects. Even with safe_globals, torch.load(..., weights_only=True) is still not compatible with this object structure on recent PyTorch versions, because the weights-only unpickler only allows SETITEMS for plain dict-like containers.

So this is a regression from the CWE-502 hardening change, not a ROCm-specific issue. We will update the voice preset format/loading path so the bundled presets can be loaded with weights_only=True safely.

For now, reverting to the previous commit or locally using weights_only=False only for trusted bundled voice preset files can unblock the demo, but that is not the final fix we want to keep.