fix(pytorch): load saved models with weights_only=False (#5365)

[Bharath-970]

Fixes loading of models saved with the bentoml.pytorch framework on PyTorch >= 2.6.

Problem

bentoml.pytorch.save_model serializes the whole model object (not a state dict) via torch.save(model, file, pickle_module=cloudpickle). Loading such a file requires torch.load(..., weights_only=False).

PyTorch 2.6 changed the default of torch.load from weights_only=False to weights_only=True. weights_only=True cannot unpickle arbitrary classes, so loading any model saved with bentoml.pytorch now fails:

_pickle.UnpicklingError: Weights only load failed. ...
WeightsUnpickler error: Unsupported global: GLOBAL <YourModel> was not an allowed global by default.

This breaks both bentoml.pytorch.load_model(...) and the legacy runner serving path (_internal/frameworks/common/pytorch.py), which calls load_model with no way for the user to pass torch_load_args. Reported in #5365.

Fix

Default weights_only to False in load_model via setdefault, since the model store is a trusted, BentoML-produced artifact (consistent with torch's guidance to only disable weights_only for trusted sources). Callers can still override it through **torch_load_args.

torch_load_args.setdefault("weights_only", False)
model = torch.load(file, map_location=device_id, **torch_load_args)

Only pytorch.py is affected; pytorch_lightning and torchscript use torch.jit.load, which is unaffected.

Test

Adds test_load_model_defaults_to_weights_only_false (isolated temp model store):

• saving + loading a custom nn.Module succeeds by default
• an explicit weights_only=True override is still honored

Verified as a true regression guard: the test fails without the fix (UnpicklingError) and passes with it. Full file:

tests/integration/frameworks/test_pytorch_unit.py ...  3 passed

Closes #5365.

[Bharath-970]

gentle bump on this one. it bites anyone on torch 2.6+ since the weights_only default flipped — every model saved via bentoml.pytorch fails to load with the UnpicklingError. fix is a one-liner (setdefault so callers can still override) plus a regression test, CI's green.

@frostming @aarnphm mind taking a look when you get a sec? happy to tweak if you'd rather gate it differently.