ONNX Runtime and the 404 response from its public GitHub repository.
AI / The story

ONNX Runtime’s GitHub repository returns 404. Its package metadata still responds

Direct checks found the source repository unavailable while documentation and PyPI metadata remained reachable. The distinction changes how developers should diagnose a failed build.

GeekRadar4 min read

Microsoft’s public ONNX Runtime repository is returning a 404 error on GitHub, while the project’s installation documentation and Python package metadata remain reachable. That split matters: losing access to a source repository can interrupt some build processes without making every installed application or package distribution unavailable.

GeekRadar checked the repository page and GitHub’s repository API at 13:45 UTC on November 18. Both returned HTTP 404. The PyPI metadata endpoint responded successfully and identified version 1.32.0, with 24 files listed for that release. Those checks establish the responses of those endpoints at that time; they do not establish that every package downloads, installs or runs correctly.

The distinction is more useful than a blanket declaration that ONNX Runtime is down. Developers need to locate the failed request in their own workflow before deciding what to change.

A repository error does not explain its cause

ONNX Runtime is an engine for running machine-learning models across different software and hardware environments. Its official installation guide describes several distribution routes, including Python packages, NuGet packages and JavaScript packages. Source access is one part of that ecosystem.

The current GitHub error does not, by itself, demonstrate that Microsoft has ended the project, deleted all copies of its source, or suffered a security compromise. GitHub’s troubleshooting documentation explains that a repository-not-found error can reflect a repository that does not exist or one the requester cannot access. The public response does not distinguish the underlying reason here.

GitHub’s service-status endpoint reported normal operation during the same checks. That is useful context, but a platform-wide status summary cannot resolve the availability of one repository or explain an individual access restriction.

The GitHub error page shown at the ONNX Runtime repository address during the November 18 checks. GitHub.
The GitHub error page shown at the ONNX Runtime repository address during the November 18 checks. GitHub.

Find the request your build actually makes

A job that fetches source directly from the affected GitHub repository has a different dependency from one that resolves a prebuilt package through PyPI. An application already running with a locally installed runtime has another relationship again. It does not necessarily contact the repository each time it performs inference; the application’s own update and startup behavior determines what it needs from the network.

The first useful evidence is therefore the failing URL and the stage at which it is requested. A source checkout, a release-asset download, package resolution and loading an installed library are separate operations. Treating them as interchangeable can send a developer toward an unnecessary version change while leaving the actual missing input untouched.

StableBuild’s account of the incident draws attention to that dependency distinction. The company sells artifact-preservation services, so its proposed solution has a commercial context. The underlying question is still practical: does a clean build need to retrieve something that a previously prepared machine already has?

A cached development machine can hide that difference. Its successful rebuild does not prove that a new continuous-integration worker, without the same stored inputs, will succeed. Conversely, a failed source download does not prove that a separately distributed package is broken. Keep the diagnosis tied to the operation observed.

Availability and trust need separate checks

A reachable mirror is not automatically an equivalent replacement for the original source. Before changing where a build downloads code or binaries, establish the provenance and expected contents of the replacement through the project’s normal dependency-review process. An availability problem is a poor reason to introduce an unexamined input.

Preserving known build inputs can reduce dependence on an upstream host remaining available. A pinned version and an existing lockfile identify what a project expects; they cannot, on their own, guarantee that an external server will still provide it. For teams maintaining open-source software years after its initial release, preserving access to the inputs is a separate responsibility from choosing their versions.

For this incident, the cause remains unestablished by the evidence above. The immediate task is narrower and actionable: identify whether a failing workflow needs the missing repository, a separately hosted package, or something already present locally. That answer determines whether the team needs to recover a build input at all—and prevents a visible 404 from becoming an unnecessary rewrite of a working deployment.