PyTorch .pt Legacy Format OOM PoC

Summary

Missing numel validation in _legacy_load() -> persistent_load() in torch/serialization.py. A 334-byte crafted .pt file triggers 1-4 GiB allocation on torch.load(). Bypasses weights_only=True safety mode.

Files

poc_generate.py - generates crafted files, tests OOM with memory limit
pytorch_oom_poc_1gb.pt - pre-crafted, targets 1 GiB
pytorch_oom_poc_4gb.pt - pre-crafted, targets 4 GiB

Reproduce

pip install torch
python3 poc_generate.py --test --gb 1

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support