YOLOv8 ONNX Export Fails on Jetson – Assertion Error in stl_vector.h, Core Dumped

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin NX
1

Hi NVIDIA Team,

I’m working on an NVIDIA Jetson device and trying to export a YOLOv8 segmentation model (best.pt) to ONNX format using the Ultralytics CLI. However, the export process fails with an assertion error and core dump:
/opt/rh/gcc-toolset-14/root/usr/include/c++/14/bits/stl_vector.h:1130:
std::vector<_Tp, _Alloc>::reference std::vector<_Tp, _Alloc>::operator
[with _Tp = unsigned int; _Alloc = std::allocator; reference = unsigned int&; size_type = long unsigned int]:
Assertion ‘__n < this->size()’ failed.

Aborted (core dumped)

System Details:

  • Device: NVIDIA Jetson (ARM Cortex-A78AE)
  • JetPack Version: 6.0
  • Python: 3.10.12
  • Torch: 2.3.0 (CPU only build)
  • Ultralytics: 8.3.107
  • GCC Toolset: gcc-toolset-14
  • Model: YOLOv8-seg (custom trained)
  • Command Used:
yolo export model=best.pt format=onnx simplify=True dynamic=False imgsz=640

Notes:

  • The model runs fine in PyTorch inference on the Jetson.
  • ONNX export works on my x86 system but fails on Jetson.
  • Seems like an std::vector access error during export.

Questions:

  1. Is PyTorch 2.3.0 officially supported on Jetson, or could this be the cause?
  2. Is there a known issue with ONNX export and vector assertions on Jetson?
  3. Should I downgrade PyTorch or do the export on x86 and run inference-only on Jetson?

Any help or pointers would be really appreciated!
Thanks,
Karthik

3

Hi,

Does YOLO11 work for you?
We test a YOLO11 segmentation model and it can work well on our latest JetPack 6.2 release.

$ yolo export model=yolo11n-seg.pt format=onnx
Downloading https://github.com/ultralytics/assets/releases/download/v8.3.0/yolo11n-seg.pt to 'yolo11n-seg.pt'...
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5.90M/5.90M [00:02<00:00, 2.12MB/s]
Ultralytics 8.3.82 🚀 Python-3.10.12 torch-2.6.0 CPU (ARMv8 Processor rev 1 (v8l))
YOLO11n-seg summary (fused): 113 layers, 2,868,664 parameters, 0 gradients, 10.4 GFLOPs

PyTorch: starting from 'yolo11n-seg.pt' with input shape (1, 3, 640, 640) BCHW and output shape(s) ((1, 116, 8400), (1, 32, 160, 160)) (5.9 MB)

ONNX: starting export with onnx 1.17.0 opset 19...
ONNX: slimming with onnxslim 0.1.48...
ONNX: export success ✅ 2.0s, saved as 'yolo11n-seg.onnx' (11.2 MB)

Export complete (3.2s)
Results saved to /home/nvidia
Predict:         yolo predict task=segment model=yolo11n-seg.onnx imgsz=640  
Validate:        yolo val task=segment model=yolo11n-seg.onnx imgsz=640 data=/ultralytics/ultralytics/cfg/datasets/coco.yaml  
Visualize:       https://netron.app
💡 Learn more at https://docs.ultralytics.com/modes/export

Thanks.

4

Hi,

Thanks for confirming that YOLO11 segmentation works on JetPack 6.2 — that’s really encouraging!

I tested the same export command on my Jetson device, which is currently running JetPack 6.0 (versions 6.0+b106 and 6.0+b87):

yolo export model=yolo11n-seg.pt format=onnx

The model downloads and begins the export, but it crashes with this error:

/opt/rh/gcc-toolset-14/root/usr/include/c++/14/bits/stl_vector.h:1130: 
std::vector<_Tp, _Alloc>::reference std::vector<_Tp, _Alloc>::operator[](size_type) 
[with _Tp = unsigned int; _Alloc = std::allocator<unsigned int>; reference = unsigned int&; size_type = long unsigned int]: 
Assertion '__n < this->size()' failed.
Aborted (core dumped)

Environment details:

  • Ultralytics 8.3.108
  • Python 3.10.12
  • Torch 2.3.0 (CPU)
  • Cortex-A78AE architecture
  • JetPack 6.0

It seems like the export logic (possibly from ONNXSlim or export internals) is trying to access an index that’s out of bounds in a std::vector.

Any workaround for this issue that doesn’t require upgrading the JetPack version?

Thanks in advance!

5

Hi,

A possible reason is that the latest Ultralytics software is not compatible with the previous environment.

Have you checked this with the Ultralytics team?
Suppose there should be a version that can work on the JetPack 6.0.

Thanks.

6

Hi @AastaLLL
I Checked with Ultralytics team they replied saying
You could open an issue on NVIDIA Forums since it seems platform specific
Should i need to downgrade the jetson

7

You can try using the Ultralytics docker image for Jetson for export.

8

Hi,

Is upgrading to the latest JetPack 6.2 an option for you?
The segmentation model can work on the latest release.

Thanks.