WeightsUnpickler error: Unsupported global: GLOBAL numpy._core.multiarray._reconstruct was not an allowed global by default.

[sa-paul]

trafficstars

🐛 Bug

WeightsUnpickler error: Unsupported global: GLOBAL numpy._core.multiarray._reconstruct was not an allowed global by default. Please use torch.serialization.add_safe_globals([_reconstruct]) or the torch.serialization.safe.globals([_reconstruct]) context manager to allowlist this global if you trust this class/function.

The Problem:

In fairseq, at fairseq/fairseq/checkpoint_utils.py the torch.load method is throwing error message as: Weights only load failed. And there is 2 options to solve this:

1. Make the weights_only=False
2. If you to keep weights_only=True, Do Torch.serialization.add_safe_globals([argparse.Namespace])

I need weights and so used the second method, but the problem is, it throws error again telling: RuntimeError: Can not safely load weights when explicit pickle_module is specified I removed pickle, again it throws error saying there is those said 2 options with the new error that I shown above.

So, please help me to solve the error deadlock.

To Reproduce

I used fairseq interactive command after data preprocessing, training as shown below:

CUDA_VISIBLE_DEVICES=0 fairseq-interactive data-bin/wmt9_cs_en --path checkpoints/wmt19_cs_en/checkpoint_best.pt --beam 5 --batch-size 1 --remove-bpe sentencepiece --input sampleCS.txt 

Environment

• fairseq Version (e.g., 1.0 or main): v0.12.2
• PyTorch Version (e.g., 1.0) 2.6.0+cu126
• OS (e.g., Linux): ubuntu 24.04
• How you installed fairseq (pip, source):

git clone https://github.com/pytorch/fairseq
cd fairseq
pip install --editable ./

• Python version: 3.10.12
• CUDA/cuDNN version: CUDA 12.6
• GPU models and configuration: NVIDIA RTX A6000