tensorizer-header-len-poc

Tensorizer Header Length Parser PoC

This repository contains a local-only, non-destructive proof of concept for a Tensorizer .tensors parser issue.

The PoC demonstrates that changing the first per-tensor header length in a valid Tensorizer file from 79 to 128 makes TensorDeserializer return the expected key while loading the serialized tensor value as [0.0] instead of [1.0]. The same result is observed with verify_hash=True.

Files

• valid_minimal.tensors: baseline Tensorizer file containing one float32 tensor named weight with value [1.0].
• malformed_header_len.tensors: same file with only the per-tensor header length mutated.
• repro_tensorizer_header_len.py: local reproduction harness.

Reproduce

python3 -m venv .venv
.venv/bin/python -m pip install --upgrade pip setuptools wheel
.venv/bin/python -m pip install 'tensorizer==2.12.1'
.venv/bin/python repro_tensorizer_header_len.py

Expected relevant output:

original_header_len=79
mutated_header_len=128
first_tensor_value=[1.0]
first_tensor_value=[0.0]
first_tensor_value=[0.0]

SHA-256

01cd57e67442b0bcabd2f69230c0fd4a876f78a357afef1a6534328123497956  malformed_header_len.tensors
9d3d8919d307bcc8c6b5ee0f1d7635e1ee47b7e772aa0f7e737614277c9ab2a0  valid_minimal.tensors