Regression of VRAM consumption and performance on RIFE 4.25 model starting from TensorRT 10.14

Description

With FP16 dynamic engines generated from same onnx model file, the GPU memory usage and performance while inferencing are severely regressed in new TensorRT versions:

TensorRT version	10.13.3.9	10.15.1.29
VRAM usage (monitored from task manager)	~2.7G	~10G
Mean GPU Compute Time (trtexec inference)	3.53ms	9.59ms

After more testing, this issue firstly appeared in 10.14.

Environment

TensorRT Version: 10.15.1.29

NVIDIA GPU: GeForce RTX 5090 D

NVIDIA Driver Version: 595.79

CUDA Version: 13.2

CUDNN Version: 9.20

Operating System: Windows 11 25H2 26000.8037

Python Version (if applicable): N/A

Tensorflow Version (if applicable): N/A

PyTorch Version (if applicable): N/A

Baremetal or Container (if so, version): Baremetal Windows, no containers involved

Relevant Files

Model link:
rife_v4.25, the one in rife_v2

Steps To Reproduce

Commands or scripts:
Generate engine:

trtexec --onnx=C:/Users/datasone/AppData/Roaming/VapourSynth/plugins64\models\rife_v2\rife_v4.25.onnx --saveEngine=C:/Users/datasone/work/temp/10.15.engine --minShapes=input:1x7x256x320 --optShapes=input:1x7x1080x1920 --maxShapes=input:1x7x2160x4096 --fp16 --tacticSources=+CUBLAS,+CUBLAS_LT,+CUDNN,+EDGE_MASK_CONVOLUTIONS,+JIT_CONVOLUTIONS --useCudaGraph --noDataTransfers --noTF32 --inputIOFormats=fp16:chw --outputIOFormats=fp16:chw --builderOptimizationLevel=3 --precisionConstraints=obey --layerPrecisions=/Cast_2:fp32,/Cast_3:fp32,/Cast_5:fp32,/Cast_7:fp32,/Reciprocal:fp32,/Reciprocal_1:fp32,/Mul:fp32,/Mul_1:fp32,/Mul_8:fp32,/Mul_10:fp32,/Sub_5:fp32,/Sub_6:fp32,ONNXTRT_Broadcast_236:fp32,ONNXTRT_Broadcast_238:fp32,ONNXTRT_Broadcast_273:fp32,ONNXTRT_Broadcast_275:fp32,ONNXTRT_Broadcast_*:fp32 --verbose --useSpinWait --exportLayerInfo=C:/Users/datasone/work/temp/10.15.graph.json --exportProfile=C:/Users/datasone/work/temp/10.15.profile.json --profilingVerbosity=detailed >trtexec.10.15.log

Inference test:

trtexec --loadEngine=10.15.engine --shapes=input:1x7x1080x1920 --noDataTransfers --useSpinWait --warmUp=500 --duration=5

The following files are generated from the command above

trtexec.10.13.3.log
10.13.3.graph.json
10.13.3.profile.json

trtexec.10.15.log
10.15.graph.json
10.15.profile.json

Have you tried the latest release?: Yes

Attach the captured .json and .bin files from TensorRT's API Capture tool if you're on an x86_64 Unix system

Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt): N/A

Reproduce the reported comparison with the RIFE v4.25 ONNX model using the provided trtexec engine-generation and inference commands. Compare the attached 10.13 and 10.15 graph, profile, and log files to identify what changed starting in TensorRT 10.14; done means explaining or resolving the VRAM and GPU-time regression.