[Feature Request] MatMulNBits 2-bits Weights + Float Zero Point kernel support

[thpereir]

Describe the feature request

Summary

Add native support in the MatMulNBits operator for 2-bit quantized weights paired with float16/bfloat16 Zero Points.

Motivation

2-bit quantization (e.g., from QAD/Quark-quantized models) uses non-uniform quantization levels such as [-1, -1/3, 1/3, 1]. These levels can be expressed in the MatMulNBits dequantization formula (index - zero_point) * scale only when zero_point = 1.5 — a fractional value that cannot be represented as an integer zero-point.

Currently, onnxruntime-genai works around this by packing 2-bit weights into MatMulNBits nodes and passing a float16 zero-point tensor of value 1.5 per group. However, MatMulNBits kernels does not implement support for:

• bits = 2
• float16/bfloat16 zero-points (as opposed to the packed uint8 integer zero-points used for 4-bit)

This creates a fragile dependency on undocumented runtime behavior.

Requested Changes

1. Kernel support for bits = 2 — ensure the MatMulNBits CUDA, CPU, and other EP kernels correctly handle 2-bit packed weight tensors (4 values per byte, LSB-first).

2. Float16/float32 zero-point input — formally specify and implement support for float-typed zero-points in MatMulNBits. For 2-bit non-uniform quantization the zero-point is fractional (e.g., 1.5) and must be stored as a float.

Describe scenario use case

Example Dequantization for 2-bit QAD

dequant_value = (uint2_index - 1.5f16) * (original_scale * (2/3))

uint2 index	dequant value (relative to scale s)
0	-s
1	-s/3
2	+s/3
3	+s