[Python] PyTorch not recognizing RTX 5090 (sm_120) on Windows 11 – CUDA error: no kernel image...

I'm trying to use PyTorch with an NVIDIA GeForce RTX 5090 (Blackwell architecture, CUDA Compute Capability sm_120) on Windows 11, and I keep running into compatibility issues. PyTorch detects CUDA, but any GPU operation fails with a kernel image error.

System details:

OS: Windows 11
GPU: NVIDIA GeForce RTX 5090
Driver version: 581.80
​

nvidia-smi output:

NVIDIA-SMI 581.80 Driver Version: 581.80 CUDA Version: 13.0
| 0 NVIDIA GeForce RTX 5090 WDDM | 00000000:01:00.0 On |
| 0% 36C P8 14W / 600W | 1094MiB / 32607MiB |
​

Python version: 3.12

PyTorch version:

PyTorch version: 2.7.0.dev20250310+cu124
CUDA available: True
CUDA device count: 1
Device 0: NVIDIA GeForce RTX 5090
​

Warning message:

UserWarning:
NVIDIA GeForce RTX 5090 with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
​

Error when running a simple GPU operation: x = torch.randn(1000, 1000, device="cuda")

Result:

RuntimeError: CUDA error: no kernel image is available for execution on the device
​

What I want to know:

Does PyTorch currently provide any official Windows binaries that support the sm_120 architecture used in the RTX 5090?

If not, are these the only current options?

• 
  Using Linux/WSL2 and installing the latest PyTorch nightly with CUDA 12.8+
  
  
• 
  Building PyTorch from source on Windows with:
  
  set TORCH_CUDA_ARCH_LIST=8.6;9.0;12.0
  

Has anyone successfully used RTX 5090 + Windows 11 + PyTorch? If yes, please share the exact working combination of:

PyTorch version

CUDA version

Installation command

I'm trying to understand the current support status for Blackwell GPUs on Windows.

Continue reading...