ONNX GroupNormalization reverses channel/group divisibility in its documentation

Claim under test

Group Normalization partitions each instance's channels into equal groups. Let C be the positive number of channels and G the positive number of groups. If each group has k channels, then C = G * k, with positive integer k. The necessary and sufficient divisibility condition is C % G == 0.

The official ONNX operator page states the reverse condition in the summaries of versions 18 and 21. On that same page, the description of the num_groups attribute gives the correct condition, and the example uses four channels divided into two groups.

Minimal contradiction

Two integer examples distinguish the conditions without numerical tolerances:

C=4, G=2: C % G = 0, G % C = 2; valid: two groups of two channels.
C=2, G=4: C % G = 2, G % C = 0; invalid: half a channel per group.

The summary rejects the valid example and admits the invalid one. This is an internal documentation contradiction, independently of any runtime experiment.

Runtime controls

On 2026-09-07 the local reproducer exercised both schema versions with ONNX 1.22.0, ONNX Runtime 1.29.0 and NumPy 2.5.2 on macOS ARM64 / Python 3.12. Each input has shape (1, C, 2, 2) and contains consecutive float32 integers. Scale is one, bias is zero and epsilon is 1e-5.

Version 18 uses scale and bias of shape (G,); version 21 uses (C,). For each version, the following checks passed:

C=4, G=2: accepted; independent float64 calculation matched.
C=6, G=3: accepted; independent float64 calculation matched.
C=4, G=1: accepted; independent float64 calculation matched.
C=4, G=4: accepted; independent float64 calculation matched.
C=2, G=4: rejected at reshape.
C=5, G=2: rejected at reshape.

All eight valid outputs were finite and agreed with a separate calculation of the group mean, population variance and normalization at rtol=atol=1e-6. The four invalid models were rejected. The inputs require only a few dozen elements; ONNX Runtime, BLAS and OpenMP were limited to one worker, and all cases ran sequentially on the CPU.

The public reproduction repository contains the original version 21 example. The extended version 18/21 check and its captured output are currently local additions, pending publication.

Correction

Both summaries should say that the number of channels is divisible by num_groups. This leaves the attribute's existing rule intact.

A two-hunk local patch corrects the version 21 and version 18 strings in `onnx/defs/doc_strings.cc`. The patch passed git apply --check against the reviewed commit. No ONNX source build was required for these documentation and runtime checks.

Upstream status

ONNX issue #8422 was filed on 2026-09-06. The source correction and regenerated operator documentation were submitted as ONNX PR #8440 on 2026-09-09. The ONNX extension was rebuilt before running the official documentation generator; clang-format --dry-run --Werror and git diff --check passed.

Boundary

This finding concerns the direction of a documentation condition. The tested runtime enforces the expected grouping requirement. These examples do not establish a runtime numerical defect, production-model impact or a security vulnerability. Zero-sized tensors and zero or negative group counts are outside this audit's positive-integer scope.