Skip to main content
blobhub 0.2 is a rewrite, and it has no compatibility layer, because there is nothing left to stay compatible with: 0.1.x cannot talk to the current API. Every ONNX command it builds carries an engine key that the API now refuses with 400 invalid_request_body, and its revisions.latest() reads a revision listing that is now paged. Upgrade, and pin the new line:
0.2 needs Python 3.10 or newer.

​
Call by call

​
What behaves differently

  • The default revision, not the newest. latest() took the newest revision; revision() takes the blob’s default. They are usually the same, because a new draft becomes the default. For the newest, next(blob.revisions()).
  • Credentials come from the usual places. A key passed to a constructor becomes connect(api_key=…), BLOBHUB_API_KEY, or a profile stored by blobhub login. Public content needs no key at all: connect(anonymous=True). See Credentials and profiles.
  • An upload waits for its processing. upload() returns once the platform has processed the model, and raises OperationFailed if it could not.
  • Nothing is unpacked locally. 0.1.x downloaded the archive and extracted it. 0.2 downloads model.onnx, which the platform already extracted, and the archive only when you ask for model.tar.gz.
  • No size switches. Config(force_multipart_upload=…, force_multipart_download=…) has no equivalent. A transfer above 4 MiB always moves in parts.
  • Writes are not repeated blindly. A write that fails in a way that may have reached the platform is raised, not retried. See Errors and retries.

​
Old downloads

0.1.x saved models under ~/.blobhub/<blob id>/<revision id>/. 0.2 caches them under ~/.cache/blobhub/revisions/<revision id>/, and never reads or removes the old folders. To get the space back, delete the per-blob folders by hand. Do not delete ~/.blobhub itself: it also holds credentials.yaml, with the profiles blobhub-cli and blobhub-worker use, and blobhub-cli’s cache/.

​
Notebooks

The example notebooks now ship inside the package, rewritten for 0.2, and python -m blobhub.notebooks copy <id> copies one out. See Notebooks.