Hello Nvidia Team,
I am working on an onboard AI project using a Jetson Orin device, and I am facing issues with PyTorch GPU detection and TensorRT model export.
System Information:
-
Device: NVIDIA Jetson Orin
-
JetPack Version: 6.2.2
-
OS: Ubuntu (JetPack default)
-
Python Version: 3.10
Project Context:
I developed a YOLO-based pipeline (classification + segmentation) for oil spill detection. I am trying to optimize inference using TensorRT and FP16 to utilize Tensor Cores.
Issue Description:
- PyTorch CUDA Not Detected:
- After installing PyTorch using pip (including cu121 and NVIDIA index methods), CUDA is not detected.
Command:
python3 -c “import torch; print(torch.cuda.is_available())”
Output:
False
- TensorRT Export Fails:
- When running:
yolo export model=… format=engine half=True
I receive:
ValueError: Invalid CUDA ‘device=0’ requested
torch.cuda.is_available(): False
- NumPy Conflict (resolved partially):
-
Initially had NumPy 2.x incompatibility
-
Downgraded to numpy 1.26.x
- Installation Issues:
-
Standard pip torch installs CPU-only version
-
Attempted NVIDIA wheel installation, but some links returned 404 errors
-
Using virtual environment to avoid system conflicts
What I Need Help With:
-
Correct PyTorch installation method for JetPack 6.2.2 (with CUDA support)
-
Verified compatible versions of:
-
PyTorch
-
Torchvision
-
CUDA
-
-
Steps to ensure torch.cuda.is_available() returns True
-
Guidance for enabling TensorRT export on this setup
Additional Notes:
-
GPU is physically present and Jetson system is functioning normally
-
Issue seems specific to PyTorch/CUDA environment setup
I would appreciate detailed guidance or official installation instructions for this configuration.
Thank you.
