UDA no kernel image is available for execution on the device on 5090D (WSL2) when starting sensorsim backend #6
Description
Hi there 👋,
First of all, thank you for the great work on this project.
I’m running into a CUDA runtime error when trying to start a scene on a machine with a 5090D GPU under WSL2. The sensorsim container fails to create a backend with the following error:
RuntimeError: CUDA error: no kernel image is available for execution on the device
From the logs:
sensorsim-0-1 | [2025-12-03 00:35:35,877][nre.grpc.serve][INFO] get_available_cameras
sensorsim-0-1 | [2025-12-03 00:35:35,877][nre.grpc.serve][INFO] Creating new backend for clipgt-05bb8212-63e1-40a8-b4fc-3142c0e94646 from /mnt/nre-data/all-usdzs/sample_set/25.07_release/Batch0001/05bb8212-63e1-40a8-b4fc-3142c0e94646/05bb8212-63e1-40a8-b4fc-3142c0e94646.usdz
sensorsim-0-1 | [2025-12-03 00:35:36,448][I11llIl11lllIll1I11l11ll1][INFO] Config v25.7.9 is already compatible with NuRec v25.7.9, no upgrade needed
sensorsim-0-1 | [2025-12-03 00:35:36,611][lI1IIIllII1l1Il111lII11I1][INFO] SHGaussianModel/progressive_training: increase_frequency=1000 increase_step=1
sensorsim-0-1 | [2025-12-03 00:35:36,699][nre.grpc.serve][ERROR] Failed to create backend for clipgt-05bb8212-63e1-40a8-b4fc-3142c0e94646: CUDA error: no kernel image is available for execution on the device
...
runtime-0-1 | grpc.aio._call.AioRpcError: <AioRpcError of RPC that terminated with:
runtime-0-1 | status = StatusCode.UNKNOWN
runtime-0-1 | details = "Exception calling application: CUDA error: no kernel image is available for execution on the device
runtime-0-1 | CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
runtime-0-1 | For debugging consider passing CUDA_LAUNCH_BLOCKING=1
runtime-0-1 | Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
runtime-0-1 | "
After that, the rollout fails and the controller reports that the session does not exist.
Environment
GPU: NVIDIA 5090D
Host: WSL2 on Windows
Containers involved: sensorsim-0-1, runtime-0-1, controller-0-1
Error location: when calling get_available_cameras / creating backend for scene clipgt-05bb8212-63e1-40a8-b4fc-3142c0e94646
Framework: PyTorch (from the project’s provided environment)
Command / setup: running via the provided Docker setup (default configuration)
If helpful, I can also provide the full log output and exact driver / CUDA / PyTorch versions.
Steps to Reproduce
Run the project under WSL2 on a machine with an NVIDIA 5090D GPU using the provided Docker setup.
Start a simulation / rollout that loads the scene clipgt-05bb8212-63e1-40a8-b4fc-3142c0e94646.
Wait until the sensorsim-0-1 container tries to create a backend and call get_available_cameras.
Expected Behavior
The backend for the scene should be created successfully, and get_available_cameras should return the available cameras without CUDA errors.
Actual Behavior
sensorsim-0-1 fails with RuntimeError: CUDA error: no kernel image is available for execution on the device.
The rollout fails and the runtime/controller report that the session does not exist because initialization never completed successfully.
Question
It looks like the CUDA kernels / PyTorch extensions in the image might not be compiled for the 5090D’s compute capability.
Is this GPU / architecture currently supported?
If so, is there a recommended way to rebuild or configure the project so that the CUDA kernels are compiled for this device?
If not yet supported, are there any workarounds (e.g., flags, environment variables, different build / image) that you would recommend?
Thanks a lot for your help, and please let me know if I can provide any additional details (full logs, environment info, etc.) to debug this further. 🙏