[Feature][Relax][ONNX] Support LpNormalization operator in Relax ONNX frontend #20969

Feature Description

Support the ONNX LpNormalization operator in the Apache TVM Relax ONNX frontend (python/tvm/relax/frontend/onnx/onnx_frontend.py).

Motivation

LpNormalization computes $ vector normalization (most commonly $ or $ norm) along a specified axis. It is universally used as the final projection layer in text embedding, multi-modal embedding, and vector retrieval models (such as BGE, E5, OpenAI CLIP, and sentence-transformers) to normalize feature vectors onto the unit sphere for cosine similarity and vector search.

Importing embedding models containing LpNormalization currently fails in the Relax ONNX frontend with:

NotImplementedError: The following operators are not supported: ['LpNormalization']

This operator was also highlighted in the ONNX operator coverage RFC tracking issue #20124 under Activations / Math operators.

Operator Specification (ONNX Opset 1+)

• Inputs:
  1. input: N-D input tensor of floating-point type (float32, float16, bfloat16).
• Attributes:
  • axis: int (default -1) - the axis along which the norm is computed.
  • p: int (default 2) - the order of the norm (=1$ for $, =2$ for $).
• Outputs:
  1. output: Normalized tensor with identical shape and dtype as input.

Expected Behavior

The Relax ONNX frontend should implement LpNormalization using Relax arithmetic operations:

1. For =2$: compute norm = relax.op.sqrt(relax.op.sum(relax.op.square(input), axis=axis, keepdims=True)) and return relax.op.divide(input, norm) (with safe numerical epsilon).
2. For =1$: compute norm = relax.op.sum(relax.op.abs(input), axis=axis, keepdims=True) and return relax.op.divide(input, norm).