torch.load Truncated Storage + Forged Byteorder Corruption
Defensive validation / triage only.
This PoC shows that a crafted ZIP-format checkpoint can truncate a backing
storage record while preserving a larger logical storage length and forging the
checkpoint byteorder record. torch.load() still succeeds and returns
corrupted tensor data instead of failing closed.
Files:
model/clean.pt: clean control checkpointmodel/trunc_big.pt: crafted checkpoint with truncated storage and forged byteordermodel/build-meta.json: hashes and generation metadatareproduce/replay_truncated_storage.py: replay script for clean vs crafted load
Quick repro:
python reproduce/replay_truncated_storage.py \
--clean model/clean.pt \
--crafted model/trunc_big.pt
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support