Blackwell GPU (RTX 5060 Ti) - CUDA Not Available to Applications in WSL Despite Working nvidia-smi

Accelerated Computing CUDA CUDA on Windows Subsystem for Linux
1

Subject: Blackwell GPU (RTX 5060 Ti) - CUDA Not Available to Applications in WSL Despite Working nvidia-smi

CUDA on WSL

nvidia-smi.txt (1.7 KB)

dxdiag (2).txt (31.4 KB)

log.txt (595 Bytes)

Hello NVIDIA Developer Team,

I am writing to report a critical issue I’m facing while setting up a development environment on the new Blackwell architecture within WSL. Despite nvidia-smi functioning perfectly, no application seems to be able to initialize the CUDA context.

System Details:

  • GPU: NVIDIA GeForce RTX 5060 Ti (Blackwell Architecture)

  • Host OS: Windows 11 Enterprise (Build 26100)

  • Guest OS: WSL 2 (Ubuntu 24.04)

  • NVIDIA Driver Version: 581.15 (as reported by nvidia-smi)

  • CUDA Version: 13.0 (as reported by nvidia-smi)

Problem Description:

I am attempting to use my RTX 5060 Ti for machine learning development inside WSL.

  1. The NVIDIA driver is correctly installed on the Windows host, and nvidia-smi works perfectly inside the WSL 2 terminal, correctly identifying the GPU, driver version, and CUDA version. This confirms the base-level connection between WSL and the driver is active.

  2. However, any user-space application that tries to access the CUDA device fails. The primary test case has been PyTorch. Since no pre-compiled binaries exist for the Blackwell architecture (Compute Capability 12.0), I have successfully compiled PyTorch from the latest source code, explicitly targeting my GPU with TORCH_CUDA_ARCH_LIST="12.0".

  3. Despite this custom build, the resulting PyTorch library consistently reports that CUDA is unavailable (torch.cuda.is_available() returns False).

Troubleshooting Steps Performed:

To isolate the issue, I have confirmed the following:

  • The WSL kernel has been updated (wsl --update).

  • The NVIDIA drivers on the Windows host are the latest available.

  • The PyTorch compilation process was clean and completed without errors.

  • The LD_LIBRARY_PATH within WSL was explicitly set to /usr/lib/wsl/lib, which is where a dynamic search located the libcuda.so.1 library.

The fact that a system-level tool like nvidia-smi works, but a user-space application (even one compiled from source) cannot initialize the CUDA context, strongly suggests a potential issue in the driver’s API layer as it is exposed to the WSL environment for this new hardware. It seems like the driver is visible, but not fully usable by applications.

Question for the Team:

Is this a known issue with the current driver for Blackwell-series GPUs on WSL? Are there any specific workarounds, environment variables, or driver settings required to enable full CUDA context initialization for applications on this new architecture?

I have extensive logs, including DxDiag, nvidia-smi output, and the full PyTorch compilation and verification logs that I can provide immediately.

Thank you for your hard work and for any guidance you can offer. This is a critical issue for developers looking to adopt the new Blackwell platform.

2

I have no experience with either pytorch or WSL, but just wanted to make a couple of comments.

The version of cuda shown in nvidia-smi indicates the Cuda version up to which the driver supports, not the active, installed Toolkit version. Running nvcc -V will show the Toolkit version.

Looking at the, “Install PyTorch”, section on the homepage, prebuilt versions based on Cuda 12.8 and 12.9, both of which should support CC12.0, are there.