fix(model): load NeuMF pretrain checkpoints with weights_only=False

[feiiiiii5]

Fixes #2212.

Root cause

NeuMF.load_pretrain calls torch.load(path, map_location="cpu") without weights_only. Since torch 2.6 the default flipped to weights_only=True, which rejects the full pickles this library writes for GMF/MLP pretraining (state dicts wrapped with metadata) — stage 3 fails with _pickle.UnpicklingError: Weights only load fail while stages 1–2 work.

Fix

Pass weights_only=False explicitly on both loads, with a TypeError fallback to the plain call: setup.py allows torch>=1.10.0, and torch before 1.13 has no weights_only flag (where the old default already loads fully). Diff: 1 file, +9/-2.

Test

• No existing test covers load_pretrain (searched). The failure mode is the documented torch-2.6 default flip plus the reporter's stage-3 traceback; a full 3-stage NeuMF run is heavy for a unit test — full verification left to CI.

[feiiiiii5]

Checking in on this one, since it has been quiet for a couple of weeks.

The fix is one line in the NeuMF loader: torch.load(..., weights_only=False),
because the released checkpoints pickle the embedding tensors and
weights_only=True (the PyTorch 2.6 default) refuses to load them.

One question I could not answer from the issue tracker: whether you would rather
this be handled centrally, so every checkpoint load in RecBole is covered, rather
than per call site. I kept it to the one loader because that is the only path
that fails today, and a central switch would change behaviour for models that
currently load fine under the default. If you prefer the central change, I will
redo it that way.

Also glad to add a test that loads a released checkpoint if you want the coverage
in this PR rather than a follow-up.

[feiiiiii5]

Checking in on this before it goes stale.

The change: NeuMF.load_pretrain calls torch.load(path, map_location="cpu") without weights_only. Since torch 2.6 the default flipped to True, which rejects the full pickles this library writes for GMF/MLP pretraining — stage 3 dies with _pickle.UnpicklingError: Weights only load fail while stages 1–2 pass, which is what makes it look like a training bug rather than a load bug.

The fix: weights_only=False explicitly on both loads, with a TypeError fallback to the plain call, because setup.py allows torch>=1.10.0 and torch before 1.13 has no weights_only flag.

What I would like to know: whether you would rather see the alternative — declaring torch>=2.6 and letting the default stand — or keeping the explicit flag. I went with the explicit flag because the repo's own floor is torch 1.10, but if you are planning to move that floor anyway this becomes unnecessary.

I am not asking you to re-read the diff to answer; it is a one-line policy question about where the compatibility floor should live.

Happy either way, and no rush if this is simply sitting lower in the queue — I will not ping again.