PyTorch Container (26.06-py3) Missing Compute Capability 8.7 Kernels for Jetson Orin Nano

Robotics & Edge Computing Jetson Systems Jetson Orin Nano
1

Hello,

I am currently running JetPack 7.2 on a Jetson Orin Nano Super.

For a recent project, I need to use the latest nvcr.io/nvidia/pytorch:26.06-py3 container, primarily to leverage the latest NumPy version (as Torch is NumPy 2-compatible in the 26.x unified containers). While the container initially appears to run smoothly, I consistently encounter the following warning:

/usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py:384: UserWarning: Found GPU0 Orin which is of compute capability (CC) 8.7.
The following list shows the CCs this version of PyTorch was built for and the hardware CCs it supports:
- 8.0 which supports hardware CC >=8.0,<9.0 except {8.7}
- 8.6 which supports hardware CC >=8.6,<9.0 except {8.7}
- 9.0 which supports hardware CC >=9.0,<10.0
- 10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
- 11.0 which supports hardware CC >=11.0,<12.0
- 12.0 which supports hardware CC >=12.0,<13.0
- 12.0 which supports hardware CC >=12.0,<13.0
_warn_unsupported_code(d, device_cc, code_ccs)

Based on this output, it appears the PyTorch build in this container specifically excludes compute capability 8.7 (Orin Nano) kernels.

My smoke tests are still passing on the Nano, which suggests PyTorch is utilizing a fallback path - likely JIT-compiling embedded PTX. However, relying on this fallback for production is concerning, as there is no guarantee that PTX won’t be dropped in future releases.

I have two questions:

  1. What is the most recent PyTorch image that natively supports the Orin Nano (CC 8.7)? Which NumPy version is it compiled with?
  2. Are there plans to restore official CC 8.7 support in upcoming PyTorch NGC container releases?

Thank you in advance for your help!

3

Hi,

This is a harmless warning.
Orin can run sm_80 SASS code so no JIT-compiling will be triggered.
Please see this document:

An example of the latter is that code compiled for the target `sm_80` will run on all other CC 8.x GPUs, such as `sm_86` or `sm_89`.*

You can turn off the warming with the setting below:

import warnings
warnings.filterwarnings("ignore", message=".*Found GPU.*compute capability.*")

Thanks.