Incompatible with PyTorch 2.6+: torch.load() default weights_only=True breaks model loading

[Xyc2016]

Description

PyTorch 2.6+ changed the default value of weights_only in torch.load() from False to True. This breaks loading the FastSAM checkpoint because the .pt file contains custom classes (e.g. ultralytics.nn.tasks.SegmentationModel) that are not in the default allowlist.

Error

_pickle.UnpicklingError: Weights only load failed.
WeightsUnpickler error: Unsupported global: GLOBAL ultralytics.nn.tasks.SegmentationModel was not an allowed global by default.

Location

ultralytics/nn/tasks.py lines 518 and 533 (two occurrences in torch_safe_load()):

return torch.load(file, map_location='cpu'), file  # load

Suggested Fix

Add weights_only=False to the torch.load() calls:

return torch.load(file, map_location='cpu', weights_only=False), file  # load

Since the checkpoint comes from the official model release (a trusted source), this is safe.

Related

• Existing attempt: PR Fix: model load for newer python and torch versions #260 (uses safe_globals approach instead)
• Existing issue: cannot load model anymore as weights_only #271 (reports same problem)

[Xyc2016]

Fixed in PR #283. Uses try-first approach: weights_only=True first, falls back to False with warning. Safer than raw weights_only=False.