Hi all,
I’m trying to deploy a yolov11m instance segmentation model trained on a custom dataset on a Jetson Xavier NX 8GB running JetPack 5.1.2
I’ve been referring to the following guide: NVIDIA Jetson - Ultralytics YOLO Docs
However the PyTorch and torchvision wheels in that guide compile with CUDA 11.7 or 11.8, which is mismatched with JP5.1.2’s native CUDA build of 11.4
I read a bunch of forum posts explaining that the CUDA mismatch could be causing the CUDA errors I was facing, so I rebuilt PyTorch and torchvision using the custom wheels found here: PyTorch for Jetson (thanks Dusty you legend) however this did not change the error I was seeing.
Current error is this:
RuntimeError: CUDA error: no kernel image is available for execution on the device
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
What I’m seeing online is that is error is caused by a CUDA mismatch between PyTorch and JetPack, however I’ve ensured the two versions match. And yes the default version of Python is 3.8 and has the correct TensorRT version assigned to it, which is the same version being used to run inference. Basic test script is as follows:
from ultralytics import YOLO
model = YOLO(‘yolov11m_seg.engine’, task=‘segment’)
results = model.predict(source=“images/test1.jpg”)
Versions:
Thanks in advance! Please help me find the stupid mistake I made XD