Incorrect Min NaN handling of TensorRT 10.16.1.11 when running ONNX Min on GPU

Description

TensorRT appears to handle NaN values incorrectly for ONNX Min.

For inputs containing NaN, ONNX Runtime propagates NaN in the output. TensorRT instead returns the non-NaN operand for mixed NaN/non-NaN inputs, which changes the result.

This appears to be a TensorRT execution issue for ONNX Min NaN semantics.

Environment

TensorRT Version: 10.16.1.11

NVIDIA GPU: N/A / not detected by nvidia-smi

NVIDIA Driver Version: N/A / nvidia-smi failed

CUDA Version: N/A / nvcc not found

CUDNN Version: N/A / torch.backends.cudnn.version() returned None

Operating System: Linux 6.17.0-20-generic x86_64, glibc 2.39

Python Version (if applicable): Python 3.11.15

Tensorflow Version (if applicable): N/A

PyTorch Version (if applicable): N/A

Baremetal or Container (if so, version): Baremetal / non-Docker environment (/proc/1/cgroup: 0::/init.scope)

Additional package versions:

ONNX Version: 1.21.0
ONNX Runtime Version: 1.25.1

Relevant Files

Model link: N/A

The ONNX model is generated inline by the minimal reproducible script below.

Steps To Reproduce

Commands or scripts:

import numpy as np
import onnx
import onnxruntime as ort
from onnx import helper, TensorProto
from _trt_helper import build_engine_from_onnx, run_engine

n = helper.make_node("Min", ["a", "b"], ["y"])
g = helper.make_graph(
    [n],
    "g",
    [
        helper.make_tensor_value_info("a", TensorProto.FLOAT, [4]),
        helper.make_tensor_value_info("b", TensorProto.FLOAT, [4]),
    ],
    [helper.make_tensor_value_info("y", TensorProto.FLOAT, [4])],
)

m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 18)])
m.ir_version = 10
ob = m.SerializeToString()

a = np.array([np.nan, 12.0, np.nan, 7.5], dtype=np.float32)
b = np.array([6.0, np.nan, np.nan, 4.5], dtype=np.float32)

ort_y = ort.InferenceSession(
    ob,
    providers=["CPUExecutionProvider"],
).run(["y"], {"a": a, "b": b})[0]

eng, _ = build_engine_from_onnx(ob)
trt_y = run_engine(
    eng,
    {"a": a, "b": b},
    ["y"],
    [(4,)],
    [np.float32],
)["y"]

print("ORT:", ort_y.tolist())
print("TRT:", trt_y.tolist())

assert (np.isnan(ort_y) != np.isnan(trt_y)).any()

Have you tried the latest release?: Yes, reproduced with TensorRT 10.16.1.11.

Attach the captured .json and .bin files from TensorRT's API Capture tool if you're on an x86_64 Unix system Not attached. The issue is reproducible from the self-contained Python script above.

Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt): Yes. ONNX Runtime runs the same model and propagates NaN values.

Actual output:

ORT: [nan, nan, nan, 4.5]
TRT: [6.0, 12.0, nan, 4.5]

TensorRT returns 6.0 for min(nan, 6.0) and 12.0 for min(12.0, nan), while ONNX Runtime returns nan for both cases.

Start with the inline Python reproduction, especially build_engine_from_onnx and run_engine, and compare its TensorRT output with the ONNX Runtime result. Trace the ONNX Min conversion and execution path for NaN inputs; done means mixed NaN/non-NaN inputs produce the NaN-propagating result shown by ONNX Runtime.