CUDA error: no kernel image is available for execution on the device

[1041909892]

Hi,

When I deployed VMamba on jetson Orin nano, I ran into the following issue.The github address of VMamba is MzeroMiko/VMamba: VMamba: Visual State Space Models，code is based on mamba (github.com)

Here is my system environment. Jetpack6.0 rev.1 installed via SDK Manager 2.1.0.11682 x86_64

When I execute PYTHONPATH=. python tools/test.py configs/vssm1/mask_rcnn_vssm_fpn_coco_tiny_ms_3x.py models/mask_rcnn_vssm_fpn_coco_tiny_ms_3x_s_epoch_31.pth, I get the following error, what could be the reason?

When I set CUDA_LAUNCH_BLOCKING=1, it still doesn’t work.

[AastaLLL]

Hi,

How do you install PyTorch and TorchVision?

It looks like your package doesn’t build with CUDA support.
Please try our prebuilt with the following steps:

Installing PyTorch for Jetson Platform - NVIDIA Docs

Thanks.

[1041909892]

Hi,

I use prebuilt wheels from PyTorch for Jetson - Jetson & Embedded Systems / Announcements - NVIDIA Developer Forums to install torch=2.3 and torchvision=0.18。

And I was able to use this environment to train and inference yolov5 model.

[1041909892]

Thank you for your reply.

I’ve fixed the issue. The reason is that when installing the code packages that depend on this repository, the corresponding Compute Capability of jetson is not added.
When I set arch=compute_87, code=sm_87 and reinstall it, the model works fine.

[AastaLLL]

Thanks for the update!