nan output in fp16 mode of TensorRT 10.16 when running trtexec on GPU RTX 5080

Description

trtexec --onnx=camera.backbone.onnx --saveEngine=camera.backbone.engine
trtexec --onnx=camera.backbone.onnx --fp16 --saveEngine=camera.backbone_fp16.engine

trtexec --loadEngine=camera.backbone_fp16.engine --loadInputs=img:img.bin,depth:depth.bin --exportOutput=output_fp16.json

trtexec --loadEngine=camera.backbone.engine --loadInputs=img:img.bin,depth:depth.bin --exportOutput=output.json

output.json is correct while output_fp16.json are all nan values

Environment

RTX 5080

TRT 10.16

Driver Version: 580.142

Cuda version 13.0

Operating System: Ubuntu 22.04

Model link : https://drive.google.com/file/d/1JpS_gWXe_a96EIFh5DZmYihUS-lUZnjt/view?usp=drive_link

input bin files can be generated randomly

import numpy as np

np.random.seed(42)
img_data = np.random.rand(1, 2, 3, 256, 704).astype(np.float32)

depth_data = np.random.rand(1, 2, 1, 256, 704).astype(np.float32)

img_data.tofile('img.bin')
depth_data.tofile('depth.bin')

Start by reproducing the listed trtexec commands with TensorRT 10.16 on an RTX 5080, using the linked model and the generated img.bin and depth.bin inputs. Compare output.json with output_fp16.json and inspect the trtexec FP16 engine path. Done means the cause is isolated and FP16 inference no longer produces all-NaN output, or the issue is narrowed to a documented environment or model condition.