Hello everyone,
I’m currently experiencing an issue trying to run FramePack on my system equipped with an RTX 5090. Despite installing the latest PyTorch nightly build (2.8.0.dev20250501+cu128) and CUDA Toolkit 12.8, I encounter the following error during execution:
vbnet
CopierModifier
RuntimeError: CUDA error: no kernel image is available for execution on the device
I’ve tried several solutions, including updating NVIDIA drivers and reinstalling PyTorch with the appropriate options, but the issue persists.
My setup:
- GPU: NVIDIA RTX 5090
- OS: Windows 11 Pro
- Python: 3.10.11
- CUDA Toolkit: 12.8
- PyTorch: 2.8.0.dev20250501+cu128
I’m aware that the RTX 50 series is relatively new and compatibility issues might occur. If anyone has encountered a similar problem or has suggestions to resolve this error, I’d really appreciate your help.
Thanks in advance for your support!
I don’t know what FramePAck is but our binaries built with CUDA 12.8 work on Blackwell architectures. You could create a new and empty virtual environment, install the latest stable or nightly binary with CUDA 12.8 support, and execute a simple smoke test without installing FramePack:
import torch
print(torch.version.cuda)
print(torch.cuda.is_available())
print(torch.cuda.get_arch_list())
print(torch.randn(1).cuda())
Hi [@ptrblck],
I just wanted to sincerely thank you for the valuable help you provided regarding the installation of flash-attn. Your suggestion to use the --no-build-isolation flag was exactly what I needed — it allowed pip to properly detect torch during the build process, which was the root cause of all my installation issues.
Thanks to your advice, I was finally able to compile and install flash-attn correctly in a clean virtual environment with CUDA 12.6 and Python 3.10, without dependency conflicts or broken packages.
The only remaining issue I’m now working on is saving the final output video. It currently fails due to a missing module (PyAV), but I’m now installing av to fix that.
Thanks again for your time and clear explanation. Your help really made a difference. 
Best regards,
Bernard (aka test5090 
)
1 Like