[hiyufan]

The Relax ONNX frontend rejects checker-valid LpNormalization models because no converter is registered.

Compose existing Relax operators for L1 and L2 normalization. Honor the default axis=-1 and p=2, use absolute values for L1, and return zero for zero-norm slices. Float16/bfloat16 inputs accumulate and divide in float32 before casting back, avoiding overflow when squaring ordinary float16 values. Other input dtypes are preserved. Reject unsupported values of p.

Related to #20124 (LpNormalization only).

Validation on main e345029c, Python 3.11, LLVM 21.1.8, ONNX 1.23.2 and ONNX Runtime 1.31.0:

• Before the converter: all 144 valid matrix cases fail to import.
• After the converter: 150 passed across LpNormalization, InstanceNormalization and MeanVarianceNormalization.
• The LpNormalization matrix covers opsets 13/22, explicit/default p and axes, float16/float32/float64, static/symbolic shapes, negative values and zero-norm slices. Values and output dtypes are checked against both NumPy and ONNX Runtime. Two additional bfloat16 VM cases use an independent NumPy reference; two cases check invalid p.
• File-scoped Ruff, ASF header/file-type and whitespace checks passed.

The NumPy L1 reference uses absolute values; ONNX 1.23.2's Python ReferenceEvaluator uses a signed sum for p=1, so it is not used as the oracle.

Prepared with OpenAI Codex assistance. The LLVM build and validation above were executed locally using the pinned TIRx/FFI sources.

CI status: Windows fails during Conda environment setup, before compilation or tests, with the same conda init/Miniforge installation errors on unchanged main. This remains an outstanding validation gap; Jenkins checks are still pending.