[Relax][ONNX] Add frontend support for QuantizeLinear, DequantizeLinear, and DynamicQuantizeLinear

[hunghsiangwang]

Summary

This PR adds Relax ONNX frontend support for:

• QuantizeLinear
• DequantizeLinear
• DynamicQuantizeLinear

The implementation follows existing TVM ONNX frontend patterns and keeps QDQ handling consistent for singleton quantization parameters and optional zero-point inputs.

Changes

• add ONNX frontend converters for QuantizeLinear,DequantizeLinear, and DynamicQuantizeLinear
• register Q/DQ-related ops in the ONNX converter map
• handle optional zero-point inputs consistently during import
• preserve singleton quantization parameter semantics in the QDQ legalization path
• improve QDQ legalization behavior for imported ONNX models
• add and update frontend tests for Q/DQ and DynamicQuantizeLinear

Tests

Added or updated tests in tests/python/relax/test_frontend_onnx.py to cover:

• singleton-qparam QuantizeLinear in opset 10
• singleton-qparam DequantizeLinear in opset 10
• optional-zero-point QuantizeLinear in opset 13
• DynamicQuantizeLinear in opset 11

Validation

Validated with:

• python -m pytest -n 1 tests/python/relax/test_frontend_onnx.py -k "quantizelinear or dequantizelinear or dynamicquantizelinear" -v

Result:

• 4 passed

[gemini-code-assist]

Code Review

This pull request implements the ONNX operators QuantizeLinear, DequantizeLinear, and DynamicQuantizeLinear in the Relax frontend. It also enhances the legalization and struct info inference for quantization operations to correctly handle singleton tensors (shape-[1]) as scalars and expands the supported data types for the zero_point parameter. The review feedback indicates that the axis attribute in the v10 implementations of QuantizeLinear and DequantizeLinear is currently hardcoded to 0, whereas it should default to 1 and be retrieved from the operator attributes to comply with the ONNX specification.

[gemini-code-assist]

The axis for quantization is hardcoded to 0. According to the ONNX specification for QuantizeLinear opset 10, there is an axis attribute that defaults to 1. This implementation should handle the axis attribute from attr, similar to how it's done in _impl_v13.

Suggested change

	x, scale = inputs[0], inputs[1]
	zp = inputs[2] if len(inputs) > 2 and inputs[2] is not None else None
	out_dtype = "uint8" if zp is None else zp.struct_info.dtype
	if zp is None:
	zp = relax.const(0, out_dtype)
	return relax.op.quantize(x, scale, zp, axis=0, out_dtype=out_dtype)
	x, scale = inputs[0], inputs[1]
	zp = inputs[2] if len(inputs) > 2 and inputs[2] is not None else None
	axis = attr.get("axis", 1)
	if hasattr(x.struct_info, "ndim") and x.struct_info.ndim <= 1 and axis == 1:
	axis = 0
	out_dtype = "uint8" if zp is None else zp.struct_info.dtype
	if zp is None:
	zp = relax.const(0, out_dtype)
	return relax.op.quantize(x, scale, zp, axis=axis, out_dtype=out_dtype)

[gemini-code-assist]

The axis for dequantization is hardcoded to 0. The ONNX specification for DequantizeLinear opset 10 includes an axis attribute that defaults to 1. This should be handled from the attr dictionary, similar to the implementation in _impl_v13.

Suggested change

	x, scale = inputs[0], inputs[1]
	zp = inputs[2] if len(inputs) > 2 and inputs[2] is not None else None
	if zp is None:
	zp = relax.const(0, x.struct_info.dtype)
	return relax.op.dequantize(x, scale, zp, axis=0, out_dtype="float32")
	x, scale = inputs[0], inputs[1]
	zp = inputs[2] if len(inputs) > 2 and inputs[2] is not None else None
	axis = attr.get("axis", 1)
	if hasattr(x.struct_info, "ndim") and x.struct_info.ndim <= 1 and axis == 1:
	axis = 0
	if zp is None:
	zp = relax.const(0, x.struct_info.dtype)
	return relax.op.dequantize(x, scale, zp, axis=axis, out_dtype="float32")

[tlopex]

LGTM