Cuda Runtime (out of memory) failure of TensorRT 10.3.0 when running trtexec on GPU RTX4060/jetson/etc

[zargooshifar]

Description

I'm trying to convert yoloV8-seg model to TensorRT engine, I'm using DeepStream-Yolo-Seg for converting the model to onnx.
after running trtexec with the converted onnx file I'm getting this errors:

[11/22/2024-13:40:08] [I] Finished parsing network model. Parse time: 0.0704936
[11/22/2024-13:40:08] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[11/22/2024-13:40:47] [E] Error[1]: [defaultAllocator.cpp::allocate::31] Error Code 1: Cuda Runtime (out of memory)
[11/22/2024-13:40:47] [W] [TRT] Requested amount of GPU memory (15485030400 bytes) could not be allocated. There may not be enough free memory for allocation to succeed.
[11/22/2024-13:40:47] [E] Error[9]: Error Code: 9: Skipping tactic 0x0000000000000000 due to exception [tunable_graph.cpp:create:117] autotuning: User allocator error allocating 15485030400-byte buffer
[11/22/2024-13:40:47] [E] Error[10]: IBuilder::buildSerializedNetwork: Error Code 10: Internal Error (Could not find any implementation for node {ForeignNode[/1/Constant_36_output_0.../1/Slice_4]}.)
[11/22/2024-13:40:47] [E] Engine could not be created from network
[11/22/2024-13:40:47] [E] Building engine failed
[11/22/2024-13:40:47] [E] Failed to create engine from model or file.
[11/22/2024-13:40:47] [E] Engine set up failed

with TensorRT 10.0.0.6-1+cuda11.8 the engine can be created, but anything newer it fails.

Environment

TensorRT Version: 10.3.0
NVIDIA GPU: RTX4060
NVIDIA Driver Version: 565.57.01
CUDA Version: 12.6
CUDNN Version: 9.5.1.17-1

Operating System: Ubuntu 22.04.5 LTS
Python Version (if applicable): 3.10.12-1~22.04.7
PyTorch Version (if applicable): 2.5.1

Steps To Reproduce

 python export_yoloV8_seg.py --weights yolov8s-seg.pt
/usr/src/tensorrt/bin/trtexec --onnx=yolov8s-seg.onnx

[lix19937]

Maybe is a bug, you can try to modify the maxWorkspaceSize or memoryPoolLimit, either larger or smaller will not solve the problem.

[poweiw]

Other than trying the maxWorkspaceSize, can you also check if quantization works? e.g add an --fp16 to your trtexec command.

Would also be helpful if you can directly provide the link to the onnx file and your command to build the engine👍

[ethanlee928]

Hello, is there any updates regarding this issue? I'm facing the same issue currently in DeepStream 7.1.

Environment

TensorRT Version: 10.7.0
Hardware: Jetson Orin NX 8GB (JetPack 6.2)
CUDA Version: 12.6

Thank you!

[sufalroy]

please try building the engine with batch size 2.

'nvinfer', 'primary-infer', {
'config-file-path': 'primary-inference.txt',
'batch-size': 2,
'gpu-id': 0,
}