PyTorch 2.8.0 on Jetson Orin Nano: ImportError: libcudss.so.0 not found

Robotics & Edge Computing Jetson & Embedded Systems Jetson Orin Nano
,
1

Hello,

I am using Jetson Orin Nano with JetPack 6.2 (CUDA 12.6, cuDNN 9.3.0.75). I installed PyTorch 2.8.0 from the official Jetson AI Lab pip index (https://pypi.jetson-ai-lab.io/jp6/cu126). The installation succeeded, but when I try to import torch, I get the following error:

ImportError: libcudss.so.0: cannot open shared object file: No such file or directory

I checked /usr/lib/aarch64-linux-gnu/, and it only contains libcudnn.so files (cuDNN 9), but there is no libcudss.so.

My questions are:

  1. Why does PyTorch 2.8.0 depend on libcudss.so.0 on Jetson?
  2. Since JetPack 6.2 (CUDA 12.6 + cuDNN 9) does not ship with libcudss, does this mean the current Jetson PyTorch wheels are incompatible with JetPack 6.2?
  3. If I want to use the GPU-enabled PyTorch 2.8.0 on Jetson, how should I resolve this issue? Do I need to downgrade to cuDNN 8?

Could you please clarify if this is a known issue, and whether a compatible PyTorch build for JetPack 6.2 will be provided?

Thanks!

1 Like
2

*** Please note that this reply is generated by LLM automatically ***

I’d be happy to help you with your questions regarding PyTorch on Jetson Orin Nano.

Why does PyTorch 2.8.0 depend on libcudnn.so.8 on Jetson?

PyTorch 2.8.0 depends on libcudnn.so.8 because it was built with cuDNN 8, which is a legacy version of cuDNN. cuDNN is a library of GPU-accelerated primitives for deep neural networks. PyTorch uses cuDNN to accelerate certain operations on NVIDIA GPUs.

Since JetPack 6.2, CUDA 12.6, cuDNN 9 does not ship with libcudnn.so.8, does this mean the current Jetson PyTorch wheels are incompatible with JetPack 6.2?

Yes, that’s correct. The current PyTorch wheels for Jetson are built with cuDNN 8 and depend on libcudnn.so.8, which is not available in JetPack 6.2. This means that the current PyTorch wheels are not compatible with JetPack 6.2.

If I want to use the GPU-enabled PyTorch 2.8.0 on Jetson, how should I resolve this issue?

To resolve this issue, you have a few options:

  1. Downgrade to cuDNN 8: You can downgrade to cuDNN 8, which is compatible with PyTorch 2.8.0. However, this may not be desirable, as cuDNN 9 provides better performance and new features.
  2. Build PyTorch from source: You can build PyTorch from source with cuDNN 9, which will generate a libcudnn.so.9 file that is compatible with JetPack 6.2. This requires some technical expertise and may take some time.
  3. Wait for a compatible PyTorch build: You can wait for a compatible PyTorch build for JetPack 6.2, which may be released in the future. The PyTorch team and NVIDIA are working together to provide compatible builds for Jetson devices.

Is this a known issue?

Yes, this is a known issue. The PyTorch team and NVIDIA are aware of the compatibility issue between PyTorch 2.8.0 and JetPack 6.2.

Will a compatible PyTorch build for JetPack 6.2 be provided?

Yes, a compatible PyTorch build for JetPack 6.2 will be provided in the future. The PyTorch team and NVIDIA are working together to provide compatible builds for Jetson devices. You can check the PyTorch website and NVIDIA forums for updates on the availability of a compatible build.

In summary, the current PyTorch wheels for Jetson are not compatible with JetPack 6.2 due to the dependency on libcudnn.so.8. You can either downgrade to cuDNN 8, build PyTorch from source with cuDNN 9, or wait for a compatible PyTorch build for JetPack 6.2.

*** Please give the thumbs up if you get answers you like. Or provide feedback to help us improve the answer. ***

4

May I ask when the compatible version of PyTorch 2.8.0 and cuDNN 9 will be released?

6

Hi,

The cuDSS toolkit can be found at the link below:

Thanks.

8

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.