Support glob patterns in OnnxKQuantQuantization nodes_to_exclude

[justinchuby]

Motivation

OnnxKQuantQuantization.nodes_to_exclude only matched exact node names. For models that are split into multiple ONNX components (e.g. multimodal decoder / vision / audio), keeping a specific layer out of INT4 required hardcoding build-specific node names such as vision_encoder/projector/MatMul_node_38 and audio_encoder/projector/MatMul_node_3. Those numeric suffixes are assigned at graph-build time and shift across builds and architectures, so the exclusion list is brittle.

Concrete case

For an encoder-free multimodal model (Gemma 4 gemma4_unified), the vision and audio "encoders" are a single projection MatMul each — i.e. the entire image/audio pathway. Measured INT4-vs-FP16 error on those projectors is non-trivial (rel-L2 ≈ 3.7% vision, ≈ 9.2% audio), while the components are tiny (76 MB / 1.4 MB), so keeping them in higher precision costs almost nothing. Doing that cleanly needs a robust way to target */projector/*.

Change

Allow each nodes_to_exclude entry to be a Unix shell-style glob pattern (matched with fnmatch.fnmatchcase) in addition to an exact node name. A node is excluded if its name equals or matches any entry, so existing exact-name configs are unaffected.

{
  "type": "OnnxKQuantQuantization",
  "bits": 4,
  "nodes_to_exclude": ["*/projector/*"]
}

Why glob and not regex

ONNX node names are slash-separated paths that are full of regex
metacharacters (e.g. decoder/model/layers.0/..., model.norm). Glob is
the safer and backward-compatible choice:

• Backward compatibility. Existing nodes_to_exclude entries are exact
  names containing .. Under glob, . is a literal, so every existing
  config matches exactly the same node. Treating entries as regex would
  reinterpret . as "any char", silently changing the meaning of existing
  configs and risking over-exclusion (layers.0 would also match
  layers00). Glob is a clean superset: exact names behave identically and
  */? are the only added powers.
• No escaping burden. Regex would force users to escape . in every
  dotted path name; glob needs none for the common case.
• Intuitive anchoring. fnmatch matches the whole string, so
  */projector/* unambiguously means "whole name matches this shape",
  avoiding the re.match/re.search/re.fullmatch ambiguity.
• Path-like names. Node names are paths, and glob is the familiar tool
  for path matching (shell, .gitignore).

Regex is strictly more expressive (alternation, character classes, numeric
ranges), but that is rarely needed here and can be added later without
breaking glob (e.g. a re: prefix or a separate option) if a real need
appears.

Testing

• Added test_kquant_with_nodes_to_exclude_glob (pattern *_1 excludes one of two MatMuls); the existing exact-name and uniform tests still pass — 4 passed.
• ruff check clean on the changed files.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[Copilot]

Pull request overview

This PR extends the OnnxKQuantQuantization pass so nodes_to_exclude can match node names using Unix shell-style glob patterns (via fnmatch.fnmatchcase) in addition to exact-name matching, making exclusions more robust for graphs whose node names vary across builds.

Changes:

• Add glob-pattern support to OnnxKQuantQuantization.nodes_to_exclude.
• Expand nodes_to_exclude configuration help text with glob usage guidance.
• Add a unit test verifying a glob pattern excludes only the intended node.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
olive/passes/onnx/kquant_quantization.py	Implements glob-based matching for nodes_to_exclude and updates config documentation.
test/passes/onnx/test_kquant_quantization.py	Adds coverage to ensure glob patterns in nodes_to_exclude behave as expected.