RTX 5080 on Ubuntu 22.04 --> NVRTC Error / Incorrect Capability (12,0)

[moribots]

Problem Summary

Hi, I have a fresh build with a 5080 GPU on ubuntu 22. I used the official guide (conda --> pip) for installation, and when running the simple agents (or anything that renders)

https://isaac-sim.github.io/IsaacLab/main/source/overview/simple_agents.html

I get:

Traceback (most recent call last):
  File "/home/mrahme/IsaacLab/scripts/environments/zero_agent.py", line 70, in <module>
    main()
  File "/home/mrahme/IsaacLab/scripts/environments/zero_agent.py", line 62, in main
    env.step(actions)
  File "/home/mrahme/miniconda3/envs/env_isaaclab/lib/python3.10/site-packages/gymnasium/wrappers/common.py", line 393, in step
    return super().step(action)
  File "/home/mrahme/miniconda3/envs/env_isaaclab/lib/python3.10/site-packages/gymnasium/core.py", line 327, in step
    return self.env.step(action)
  File "/home/mrahme/IsaacLab/source/isaaclab/isaaclab/envs/manager_based_rl_env.py", line 207, in step
    self.reward_buf = self.reward_manager.compute(dt=self.step_dt)
  File "/home/mrahme/IsaacLab/source/isaaclab/isaaclab/managers/reward_manager.py", line 148, in compute
    value = term_cfg.func(self._env, **term_cfg.params) * term_cfg.weight * dt
  File "/home/mrahme/IsaacLab/source/isaaclab_tasks/isaaclab_tasks/manager_based/classic/cartpole/mdp/rewards.py", line 24, in joint_pos_target_l2
    joint_pos = wrap_to_pi(asset.data.joint_pos[:, asset_cfg.joint_ids])
RuntimeError: nvrtc: error: invalid value for --gpu-architecture (-arch)

Notably, this does not happen when I run things with --headless. In addition, if there are no issues if I run Isaac Sim on its own.

Debugging

Not sure if related, but when I run

python -c "import torch; print(f'PyTorch: {torch.__version__}'); print(f'CUDA available: {torch.cuda.is_available()}'); print(f'CUDA version: {torch.version.cuda}'); print(f'GPU: {torch.cuda.get_device_name(0)}'); print(f'Capability: {torch.cuda.get_device_capability(0)}')"

I get:

PyTorch: 2.7.0+cu128			--> 			(note: I tried this with the latest nightly pytorch as well)
CUDA available: True
CUDA version: 12.8
GPU: NVIDIA GeForce RTX 5080
Capability: (12, 0)

I think Capability being 12 could be why Isaac Lab is struggling here. I tried setting export TORCH_CUDA_ARCH_LIST="9.0" to no avail.

Other Details

I am using NVIDIA Driver Version: 570.144, I was previously on 570.86.16 with identical results.

[moribots]

Potentially relevant: lllyasviel/Fooocus#3862 (comment)