Pass weights_only=False to torch.load for PyTorch 2.6+ compatibility

[netlinux-ai]

Problem

In PyTorch 2.6, the default value of weights_only in torch.load was changed from False to True. The published StyleTTS2 sub-module checkpoints (ASR, JDC, PLBERT) and the LibriTTS pretrained model contain pickled Python objects that the new safe loader rejects, so loading any of them under PyTorch ≥ 2.6 fails with:

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

This blocks train_finetune.py, train_finetune_accelerate.py, the ASR/JDC/PLBERT loaders, and the load_checkpoint helper.

Reproducer (any of the published artefacts triggers it):

import torch
torch.load("Utils/JDC/bst.t7", map_location="cpu")  # raises on torch ≥ 2.6

Fix

Pass weights_only=False explicitly at every torch.load call site that loads these checkpoints. They come from a trusted source (the repo's own pretrained artefacts), so the new safety guard is a false positive here.

Five call sites total — three in models.py, one in Utils/PLBERT/util.py. (train_*.py files load via these helpers, so no further changes are needed there.)

Tested with

• PyTorch 2.7.0 + CUDA 12.6
• Both the LibriTTS pretrained checkpoint and the bundled ASR/JDC/PLBERT modules load cleanly after the patch
• Full fine-tune training run completed; rendered output matches expected behaviour