Unable to Load Checkpoints Due to Weights-Only Restriction / add_safe_globals Has No Effect

[termimi]

Hi team,

I'm facing an issue with loading Bark model checkpoints under PyTorch 2.6+ due to the new weights_only=True default in torch.load. Below is a summary of my setup and what I've tried:

Context and what I tried

• I use Bark directly from this repository, inside a FastAPI server (example below).
• On startup, Bark internally calls torch.load(...) through its preload_models() logic.
• Since PyTorch 2.6, I get an error about weights_only and unsafe globals (numpy.core.multiarray.scalar etc.).

To solve this, I tried using both:

from torch.serialization import add_safe_globals, safe_globals
import numpy as np

add_safe_globals([np.core.multiarray.scalar])

# During model preloading:
with safe_globals([np.core.multiarray.scalar]):
    preload_models()

The problem

• Despite using both add_safe_globals and the safe_globals context manager, I still get the same UnpicklingError/weights_only error*.
• I have no way of passing weights_only=False to torch.load from user code, and I don't want to patch Bark's source code in the repo.
• The model weights are directly downloaded from the official repo (not third-party or modified).

What I would like to know

• Is there a workaround to allow loading checkpoints with Bark (without editing the repo code) under these new PyTorch restrictions?
• Is there any officially supported way to override or bypass weights_only, or am I missing a recommended approach here?
• Any advice for users who just want to run Bark as intended without risking model/package integrity?

BTW: If the issue sounds like AI generated it is because it is, but the issue is real please help

Thanks for your help!