[Feature][Relax][ONNX] Support ConvInteger operator in Relax ONNX frontend #20965

Feature Description

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

Motivation

While MatMulInteger is currently supported in the Relax ONNX frontend (computing (A - a_zero_point) * (B - b_zero_point) with int32 accumulation), its convolution counterpart ConvInteger (opset 10+) is currently missing.

ConvInteger is a foundational operator for integer-only quantized neural networks (e.g. integer CNNs, INT8 quantization pipelines). Importing ONNX models containing ConvInteger fails with:

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

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

Operator Specification (ONNX Opset 10+)

• Inputs:
  1. x: N-D input tensor (uint8 or int8), shape (N, C, D1, D2, ...)
  2. w: N-D weight tensor (uint8 or int8), shape (M, C/group, k1, k2, ...)
  3. x_zero_point (optional): zero point for x (uint8 or int8, scalar or 1D tensor of shape (1,) or (C,))
  4. w_zero_point (optional): zero point for w (uint8 or int8, scalar or 1D tensor of shape (M,) for per-channel quantization)
• Attributes:
  • auto_pad: string (NOTSET, SAME_UPPER, SAME_LOWER, VALID)
  • dilations: list of ints
  • group: int (default 1)
  • kernel_shape: list of ints
  • pads: list of ints
  • strides: list of ints
• Outputs:
  1. y: N-D output tensor of type int32

Expected Behavior

The Relax ONNX frontend should import ConvInteger nodes by subtracting zero points from input and weight (broadcasting appropriately), upcasting to int32, performing convolution via relax.op.nn.conv1d / conv2d / conv3d, and returning int32 activations.