Latest in AI
ONNX Runtime GitHub 404: Check Your Build Before Changing Dependencies
The source repository is unavailable while the docs and PyPI listing remain reachable. Separate a failed source fetch from a package or runtime failure.
The public Microsoft ONNX Runtime repository returned HTTP 404 during checks on November 18 at 05:33 UTC. Its documentation and PyPI metadata endpoint returned HTTP 200. For a team with a failing build, those are different dependencies, and the distinction should guide the first response.
A missing repository does not establish that installed packages have stopped working, that the project has ended or that a security incident occurred. The checks for this article establish availability only. No package was installed or tested, and the cause of the repository's unavailability has not been verified.
Identify the failing fetch
Start with the first error in the build log, not the last summary. Is the job cloning microsoft/onnxruntime, downloading a GitHub release asset, resolving a package from an index or failing after importing the runtime? Save the URL, expected version and error before changing anything.
The official source-build instructions begin with a recursive clone of Microsoft's repository. That path depends on GitHub source access. A job using a previously approved binary package has a different path, even when both eventually run the same model.
For example, a documentation task that fetches an example from GitHub can fail while an inference service continues using an existing container image. Restarting or rebuilding that service to investigate the documentation failure could introduce a dependency it did not previously need. Establish the affected job first.
A visible package is not a safety verdict
At the time checked, the PyPI project listing showed version 1.32.0, released November 13, and its metadata endpoint listed 24 release files. That confirms the listing was available. It does not establish that a particular wheel installs on your platform, passes your tests or has been assessed for security.
Keep those conclusions separate in an incident note. “Package metadata is reachable” is an observation. “Our deployment is unaffected” needs evidence from your deployment. “Every package is safe” is a much stronger claim that an HTTP response cannot support.
I would avoid an unplanned upgrade simply because the latest version is visible. Preserve the version and artifacts your team has already approved while investigating, subject to your normal security review and any subsequent vendor advisory.
Do not turn a retrieval problem into a migration
Switching to a fork or a newly advertised mirror may restore a download while changing who supplies your executable code. A familiar project name is not enough to establish that the replacement contains the same source or build artifacts.
If you already maintain an internal artifact store, check the recorded origin, version, digest and approval history. A cached file can help reproduce a build, but caching does not prove the file is safe. If those records are missing, pause the affected release rather than silently weakening the dependency checks.
For Python environments, pip's secure-install guidance describes hash-checking with --require-hashes and excluding source distributions with --only-binary :all:. Hash-checking requires pinned requirements and hashes for dependencies too. These controls identify expected artifacts; they do not certify the behavior of the code inside them.
Use the hashes from your reviewed dependency record. Generating a new hash for an unfamiliar replacement and immediately trusting it defeats the purpose of comparing against an established artifact.
Check compatibility before considering a binary substitute
A source build may exist for a reason: a specific execution provider, a custom operator or a platform requirement. Replacing it with a generic CPU package can change functionality or performance even if installation succeeds.
The installation guide distinguishes packages by language, operating system and accelerator. Use that matrix to assess whether a published binary actually matches your workload. Do not replace a GPU configuration merely to make the dependency resolver happy.
Once the source becomes accessible, review any official explanation and relevant advisories before resuming the paused path. Re-run the affected build and its application tests with the intended artifacts. Repository availability, artifact identity and runtime correctness are three checks, not one green light.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help developer-tool teams turn technical incidents and dependency changes into clear, practical guidance.