[Bug]: OpenVINO silently accepts MaxPool with pad >= kernel while other backends reject it

[ALinrunrun]

OpenVINO Version

2026.1.0

Operating System

Other (Please specify in description)

Device used for inference

CPU

Framework

ONNX

Model used

Minimal ONNX model generated by the reproducer script.

Issue description

OpenVINO CPU accepts an ONNX MaxPool model where the explicit padding is greater than or equal to the kernel size.

For the same model, ONNX Runtime and onnx2torch reject the configuration with a padding/kernel validation error, while OpenVINO compiles and runs it, returning an extended output shape.

OS: Ubuntu 24.04.2 LTS / Linux x86_64

Step-by-step reproduction

#!/usr/bin/env python3
import sys
import numpy as np
import onnx
import onnxruntime as ort
import openvino as ov
from onnx import helper as oh, TensorProto as TP

x = np.random.RandomState(88).randn(1, 1, 8, 8).astype(np.float32)

node = oh.make_node(
    "MaxPool",
    ["x"],
    ["y"],
    kernel_shape=[3, 3],
    pads=[3, 3, 3, 3],
    strides=[1, 1],
    auto_pad="NOTSET",
)

graph = oh.make_graph(
    [node],
    "maxpool_bad_pad",
    [oh.make_tensor_value_info("x", TP.FLOAT, [1, 1, 8, 8])],
    [oh.make_tensor_value_info("y", TP.FLOAT, [1, 1, 12, 12])],
)

model = oh.make_model(graph, opset_imports=[oh.make_opsetid("", 13)])
model.ir_version = 8
onnx.checker.check_model(model)
mb = model.SerializeToString()
feed = {"x": x}

results = {}

try:
    so = ort.SessionOptions()
    so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
    out = ort.InferenceSession(
        mb,
        sess_options=so,
        providers=["CPUExecutionProvider"],
    ).run(None, feed)[0]
    results["ORT_ref"] = ("ok", out.shape)
except Exception as e:
    results["ORT_ref"] = ("ERR", str(e).splitlines()[0][:80])

try:
    so = ort.SessionOptions()
    so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
    out = ort.InferenceSession(
        mb,
        sess_options=so,
        providers=["CPUExecutionProvider"],
    ).run(None, feed)[0]
    results["ORT_opt"] = ("ok", out.shape)
except Exception as e:
    results["ORT_opt"] = ("ERR", str(e).splitlines()[0][:80])

try:
    core = ov.Core()
    compiled = core.compile_model(core.read_model(mb, b""), "CPU")
    out = compiled(feed)[compiled.output(0)]
    results["OpenVINO"] = ("ok", out.shape)
except Exception as e:
    results["OpenVINO"] = ("ERR", str(e).splitlines()[0][:80])

try:
    import onnx2torch
    import torch

    net = onnx2torch.convert(onnx.load_from_string(mb)).eval()
    with torch.no_grad():
        out = net(torch.from_numpy(feed["x"]))
    results["onnx2torch"] = ("ok", tuple(out.shape))
except Exception as e:
    results["onnx2torch"] = ("ERR", str(e).splitlines()[0][:80])

for name, (status, detail) in results.items():
    print(f"{name:<12} {status:<4} {detail}")

if results.get("OpenVINO", ("ERR",))[0] == "ok":
    print("BUG REPRODUCED: OpenVINO accepts MaxPool with pad >= kernel.")
    sys.exit(0)

print("NOT REPRODUCED")
sys.exit(1)

Relevant log output

Expected:
OpenVINO should reject MaxPool with pad >= kernel consistently with ORT and onnx2torch, or document this behavior clearly.

Actual:
ORT_ref      ERR  [ONNXRuntimeError] : 1 : FAIL : Exception during initialization: pads[dim] < kernel_shape[dim] ... Pad should be smaller than kernel.
ORT_opt      ERR  [ONNXRuntimeError] : 1 : FAIL : Exception during initialization: pads[dim] < kernel_shape[dim] ... Pad should be smaller than kernel.
OpenVINO     ok   (1, 1, 12, 12)
onnx2torch   ERR  pad should be at most half of effective kernel size, but got pad=3, kernel_size=3

BUG REPRODUCED: OpenVINO accepts MaxPool with pad >= kernel.

Issue submission checklist

• I'm reporting an issue. It's not a question.
• I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
• There is reproducer code and related data files such as images, videos, models, etc.