TensorRT EP Resize with +inf input returns NaN while CPU/CUDA EP return +inf

[lackjava-creator]

Describe the issue

TensorrtExecutionProvider produces NaN values for a minimal ONNX Resize model with mode="linear" when the input contains +inf, while CPUExecutionProvider and CUDAExecutionProvider both return +inf for the same model and input.

This is reproducible with a very small model:

• input shape: [1, 1, 1, 2]
• output shape: [1, 1, 1, 4]
• op: Resize(mode="linear", coordinate_transformation_mode="half_pixel")
• input values: [+inf, 1.0]

Observed outputs:

CPUExecutionProvider:       [inf, inf, inf, 1.0]
CUDAExecutionProvider:      [inf, inf, inf, 1.0]
TensorrtExecutionProvider:  [nan, nan, nan, 1.0]

The TensorRT EP profile indicates that the node was executed by TensorRT, not by CUDA/CPU fallback:

trt_count=4
cuda_count=0
cpu_count=0

I could not find an existing issue for this specific TensorRT EP Resize + infinity mismatch.

To reproduce

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


def const(name, arr):
    return helper.make_node(
        "Constant",
        [],
        [name],
        value=numpy_helper.from_array(np.asarray(arr), name),
    )


x = helper.make_tensor_value_info("x", TensorProto.FLOAT, [1, 1, 1, 2])
y = helper.make_tensor_value_info("y", TensorProto.FLOAT, [1, 1, 1, 4])

nodes = [
    const("roi", np.asarray([], dtype=np.float32)),
    const("scales", np.asarray([], dtype=np.float32)),
    const("sizes", np.asarray([1, 1, 1, 4], dtype=np.int64)),
    helper.make_node(
        "Resize",
        ["x", "roi", "scales", "sizes"],
        ["y"],
        mode="linear",
        coordinate_transformation_mode="half_pixel",
        nearest_mode="round_prefer_floor",
    ),
]

model = helper.make_model(
    helper.make_graph(nodes, "resize_inf", [x], [y]),
    opset_imports=[helper.make_opsetid("", 17)],
)
model.ir_version = 9
onnx.checker.check_model(model)

feed = {"x": np.asarray([np.inf, 1.0], dtype=np.float32).reshape(1, 1, 1, 2)}

for name, providers in [
    ("cpu", ["CPUExecutionProvider"]),
    ("cuda", ["CUDAExecutionProvider", "CPUExecutionProvider"]),
    ("trt", ["TensorrtExecutionProvider", "CUDAExecutionProvider", "CPUExecutionProvider"]),
]:
    so = ort.SessionOptions()
    so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
    so.log_severity_level = 4
    so.enable_profiling = True

    sess = ort.InferenceSession(model.SerializeToString(), so, providers=providers)
    out = sess.run(None, feed)[0]
    profile = sess.end_profiling()
    profile_text = Path(profile).read_text(errors="replace") if profile else ""

    print(name, "providers=", sess.get_providers())
    print(name, "output=", out.reshape(-1))
    print(
        name,
        "profile counts:",
        "trt=", profile_text.count("TensorrtExecutionProvider"),
        "cuda=", profile_text.count("CUDAExecutionProvider"),
        "cpu=", profile_text.count("CPUExecutionProvider"),
    )

Output on my machine:

cpu providers= ['CPUExecutionProvider']
cpu output= [inf inf inf  1.]
cpu profile counts: trt= 0 cuda= 0 cpu= 1

cuda providers= ['CUDAExecutionProvider', 'CPUExecutionProvider']
cuda output= [inf inf inf  1.]
cuda profile counts: trt= 0 cuda= 1 cpu= 0

trt providers= ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider']
trt output= [nan nan nan  1.]
trt profile counts: trt= 4 cuda= 0 cpu= 0

I also checked that ordinary finite inputs for the same model do not show this mismatch; the mismatch appears when the interpolation path contains +inf.

Note: I also reproduced that CPU EP and CUDA EP agree on the expected [inf, inf, inf, 1.0] result in onnxruntime-gpu 1.18.1. In my local TensorRT 8 setup, the 1.18.1 TensorRT provider library expected a newer TensorRT library and fell back to CUDA, so I could not independently validate the TensorRT EP path for 1.18.1 in this environment.

Urgency

Not urgent.

Platform

Linux

OS Version

Ubuntu 20.04.5 LTS, Linux 5.4.0-100-generic

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

onnxruntime-gpu 1.17.1

ONNX Runtime API

Python

Architecture

X64

Execution Provider

TensorRT

Execution Provider Library Version

TensorRT 8.6.1.6, CUDA 11.8, NVIDIA driver 580.105.08, NVIDIA GeForce RTX 3080 Ti; Python 3.10.20; onnx 1.17.0