Jetson AGX Thor: official PyTorch 25.08 container works for Conv2d and ResNet18, but pip-installed PyTorch 2.12.0.dev+cu128 fails with "no kernel im

Hi,

I am reporting a specific sm_110 PyTorch compatibility issue on Jetson AGX Thor. I have isolated which environment
works and which fails.

System:

  • Device: Jetson AGX Thor
  • Jetson Linux: R38.4.0
  • Kernel: 6.8.12-tegra
  • Driver Version: 580.00
  • CUDA Version: 13.0
  • etc nv_tegra_release:

R38 (release), REVISION: 4.0, GCID: 43443517, BOARD: generic, EABI: aarch64, DATE: Wed Dec 31 00:15:19 UTC 2025

KERNEL_VARIANT: oot

TARGET_USERSPACE_LIB_DIR=nvidia
TARGET_USERSPACE_LIB_DIR_PATH=usr lib aarch64-linux-gnu nvidia


What WORKS — official NVIDIA container:

Container: nvcr.io nvidia pytorch:25.08-py3

  • torch: 2.8.0a0+34c6371d24.nv25.08
  • CUDA: 13.0
  • cuDNN: 91200
  • Conv2d forward pass: OK
  • ResNet18 inference: OK

Working ResNet18 command:
docker run --rm --runtime=nvidia --gpus all --ipc=host
–ulimit memlock=-1 --ulimit stack=67108864
nvcr.io nvidia pytorch:25.08-py3
python3 -c “import torch, torchvision; print(‘torch:’, torch.version);
print(‘device:’, torch.cuda.get_device_name(0));
model = torchvision.models.resnet18(weights=None).cuda().eval();
x = torch.randn(1,3,224,224,device=‘cuda’); y = model(x); print(‘ok’, y.shape)”
Output:
torch: 2.8.0a0+34c6371d24.nv25.08
device: NVIDIA Thor
ok torch.Size([1, 1000])


What FAILS — pip-installed PyTorch nightly:

  • Install method: pip install (to ~ .local lib python3.12 site-packages )
  • torch: 2.12.0.dev20260317+cu128
  • CUDA: 12.8
  • cuDNN: 91900
  • Python: 3.12.3
  • torchvision: not installed
  • triton: 3.6.0+git9844da95

Failing command:
import torch, torch.nn as nn
m = nn.Conv2d(3,16,3,1,1).cuda().eval()
x = torch.randn(1,3,224,224, device=‘cuda’)
y = m(x)

Full warning + error output:
lib python3.12 site-packages torch cuda init.py:375: UserWarning: Found GPU0 NVIDIA Thor
which is of compute capability (CC) 11.0.
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}
  • 9.0 which supports hardware CC >=9.0,<10.0
  • 10.0 which supports hardware CC >=10.0,<11.0 except {10.1}
  • 12.0 which supports hardware CC >=12.0,<13.0
    Please follow the instructions at to install a PyTorch release that supports
    one of these CUDA versions: 13.0

python3.12 site-packages torch cuda init.py:493: UserWarning:
NVIDIA Thor with CUDA capability sm_110 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_80 sm_90 sm_100 sm_120.
If you want to use the NVIDIA Thor GPU with PyTorch, please check the instructions at

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

Key detail: the PyTorch build supports sm_80, sm_90, sm_100, sm_120 — but sm_110 is missing from that list. Thor is CC
11.0 (sm_110), so it falls in the gap between sm_100 and sm_120.


Relevant pip packages from failing environment:
torch 2.12.0.dev20260317+cu128
triton 3.6.0+git9844da95
cuda-bindings 12.9.4
cuda-toolkit 12.8.1
nvidia-cublas-cu12 12.8.4.1
nvidia-cuda-runtime-cu12 12.8.90
nvidia-cudnn-cu12 9.19.0.56
nvidia-nvjitlink-cu12 12.8.93


Question:

The official nvcr.io nvidia pytorch:25.08-py3 container (torch 2.8.0, CUDA 13.0) works perfectly on Thor. But the
pip-installable PyTorch nightly (torch 2.12.0.dev+cu128, CUDA 12.8) explicitly lists sm_110 as unsupported and fails
with no kernel image.

  1. Is sm_110 support expected to be included in future PyTorch pip releases for CUDA 12.8+?
  2. Is the only supported path for Thor currently the NVIDIA-provided container with CUDA 13.0?
  3. Should sm_110 be added to the PyTorch build matrix, or does Thor require a JetPack-specific PyTorch build?

Other notes:

  • I captured dmesg and journalctl -k output after testing — I do not see a GPU kernel fault. I mainly see repeated
    nvethernet “PCS block lock SUCCESS” lines.
  • Setting conservative flags (cudnn.benchmark=False, allow_tf32=False) does not help in the failing environment — the
    issue is the missing sm_110 kernel image, not a runtime instability.

Thanks

Regards
Ferenc

I don’t know that this solves sm110, but try cu130

pip3 install --pre torch torchvision --index-url \
  https://download.pytorch.org/whl/nightly/cu130