deepbom

Deployment-artifact inspection for on-device neural network models.

Identifies model artifact formats from their container signature and reads the contracts that can be decoded without loading tensor payload values.

$ pip install deepbom
$ deepbom inspect model.safetensors

model.safetensors
  sha256   a35fd03f52c12f4e78a246bec1927e9a169377fbb8905dc13165d285010e7a44
  format   safetensors   size 2.6 MB
  evidence u64 header length followed by a JSON header

tensors
  count              38
  parameters         1,377,408
  stored payload     2.6 MB
  dtypes             F16 x38

Container-level facts only. No tensor payload values were read.

$ deepbom inspect model.gguf

model.gguf
  sha256   cb95a6e10f28b76a1dd71c15560dec5a5eee8943f591ef45d11c129786b22cff
  format   gguf   size 509.0 KB
  evidence magic "GGUF" at offset 0

container
  gguf version       3
  tensors            39
  metadata           26 / 26 (complete)
  architecture       llama
  file type          2
  quant version      2

What it reads

Format is decided from container evidence — FlatBuffer identifier, magic bytes, header structure — never from the filename extension. ONNX and Core ML are separated by their protobuf field layout rather than guessed.

Usage

deepbom inspect <file>
deepbom inspect <file> --json
deepbom --version

As a library:

from deepbom import inspect

artifact = inspect("model.gguf")
print(artifact.format, artifact.sha256)
print(artifact.detail["architecture"])
print(artifact.to_dict())

Scope

This package is pure Python with no dependencies. It reports container-level facts only: what the header and directory structures determine. Tensor payload values are never read, and nothing is inferred that the container does not state.

Graph structure, quantization contracts, predicted delegate placement, target-profile cost projections and CycloneDX ML-BOM export are not part of this package. For TFLite graph analysis:

npx deepbom audit model.tflite

ONNX, Core ML and runtime evidence are available in the browser version at https://deepbom.org.

Privacy

No network access. Model bytes, filenames and results are never uploaded.

License

ISC — see LICENSE.