No-op: onnxruntime-genai#2188 is C++ runtime only, nothing to propagate to the Python builder

[Copilot]

The task was to propagate onnxruntime-genai#2188 ("Fix Gemma4 multimodal processing") into mbext. After analyzing the upstream diff, no changes are required: the PR modifies only the C++ runtime, which mbext does not mirror.

Findings

• PR #2188 touches 7 C++ files only — src/config.{cpp,h}, src/models/embeddings.{cpp,h}, src/models/multi_modal.{cpp,h}, src/models/gemma4_multimodal_processor.cpp. No builder.py / Python changes.
• mbext mirrors only the Python model builder (src/python/py/models/ → modelbuilder/), per the README. The three upstream fixes are all runtime-level:
  • pixel_values float32-stride trim (NaN fix) — C++ memcpy.
  • pixel_position_ids extraction — C++ ToBeAssigned() temporary-lifetime bug.
  • per_layer_inputs embedding→decoder forwarding — C++ pipeline wiring + genai_config schema reads.
• per_layer_inputs exists nowhere in the upstream Python builder (GitHub code search returns only the C++ files). The runtime support landed ahead of any builder, so there is no Python layout that emits it to propagate.
• Gemma4 PLE is already implemented in modelbuilder/builders/gemma.py (_make_ple_pre_computation / _make_ple_layer_block), computed internally in a monolithic decoder and covered by existing tests/fast/test_random_gemma4.py PLE tests. It does not depend on the new runtime per_layer_inputs boundary.

Recommendation

Close the tracking issue as not applicable to the Python builder.

A separate, larger effort could add decoder-side per_layer_inputs as a model input under --exclude_embeds (with decoder.inputs.per_layer_inputs / embedding.outputs.per_layer_inputs in genai_config), but that would put mbext ahead of upstream rather than propagating #2188, and there is no Gemma4 embedding-model producer or test harness here to validate it. Happy to scope that as a follow-up if desired.