CPU QLinearConv rejects per-channel weight zero points with different values

[ALinrunrun]

Describe the issue

ONNX Runtime CPUExecutionProvider raises a runtime exception for QLinearConv when the weight zero point is a 1-D per-channel tensor with different values.

In the reproducer below, the weight tensor has Cout=2, with:

w_s = [0.02, 0.02]

w_zp = [5, 90]

Both w_s and w_zp are 1-D tensors of shape [Cout]. However, ORT raises:

QLinearConv : zero point of per-channel filter must be same

This appears related to asymmetric per-channel weight quantization support in the CPU QLinearConv kernel.

To reproduce

import numpy as np
import onnxruntime as ort
from onnx import TensorProto, helper, numpy_helper

Cout = 2

w_q = np.full((Cout, 1, 2, 2), 100, np.uint8)
w_s = np.array([0.02, 0.02], dtype=np.float32)
w_zp = np.array([5, 90], dtype=np.uint8)
x_q = np.full((1, 1, 3, 3), 130, dtype=np.uint8)

inits = [
numpy_helper.from_array(np.float32(0.05), "x_s"),
numpy_helper.from_array(np.uint8(128), "x_zp"),
numpy_helper.from_array(w_q, "W"),
numpy_helper.from_array(w_s, "w_s"),
numpy_helper.from_array(w_zp, "w_zp"),
numpy_helper.from_array(np.float32(0.1), "y_s"),
numpy_helper.from_array(np.uint8(128), "y_zp"),
]

node = helper.make_node(
"QLinearConv",
["x", "x_s", "x_zp", "W", "w_s", "w_zp", "y_s", "y_zp"],
["y"],
pads=[0, 0, 0, 0],
)

g = helper.make_graph(
[node],
"g",
[helper.make_tensor_value_info("x", TensorProto.UINT8, [1, 1, 3, 3])],
[helper.make_tensor_value_info("y", TensorProto.UINT8, [1, Cout, 2, 2])],
initializer=inits,
)

m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 13)])
m.ir_version = 8

sess = ort.InferenceSession(
m.SerializeToString(),
providers=["CPUExecutionProvider"],
)

out = sess.run(None, {"x": x_q})[0]
print(out)

Urgency

Expected output

Expected: the model runs with per-channel w_zp values, since w_s and w_zp are both 1-D tensors with shape [Cout].

Actual output

[ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Non-zero status code returned while running QLinearConv node.

QLinearConv : zero point of per-channel filter must be same.
This happens by design if the quantization is symmetric.

Platform

Linux

OS Version

Linux-6.17.0-20-generic-x86_64-with-glibc2.39

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.25.1

ONNX Runtime API

Python

Architecture

X86

Execution Provider

Default CPU

Execution Provider Library Version

No response