Hi,
I am encountering a CUDA runtime error when running the demo:
python demo.py -p "A beautiful landscape with a river and mountains"
The error is:
CUDA error (/__w/xformers/xformers/third_party/flash-attention/hopper/flash_fwd_launch_template.h:175):
no kernel image is available for execution on the device
My environment:
GPU:
NVIDIA RTX 5090
Compute capability: sm_120
Software:
PyTorch: 2.7.1+cu128
xformers: 0.0.30
CUDA runtime: 12.8
Python: 3.11
OS: Ubuntu (WSL2)
From the traceback, the error appears to originate from xformers' flash-attention (hopper) kernel. It seems that the prebuilt xformers wheel does not contain a kernel image compiled for sm_120.
Interestingly, ‘python demo.py -p "A well-designed cozy bedroom" --use_sharp’ that do not trigger flash-attention run successfully in the same environment.
Could you please share the exact environment configuration used to test the demo? Specifically:
PyTorch version
xformers version
CUDA version
GPU model (compute capability)
Has this project been tested on RTX 5090 (sm_120)?
If flash-attention currently does not support sm_120, is there a recommended workaround (for example disabling flash-attention or using an alternative attention backend)?
Thank you very much for your help.
Hi,
I am encountering a CUDA runtime error when running the demo:
python demo.py -p "A beautiful landscape with a river and mountains"
The error is:
CUDA error (/__w/xformers/xformers/third_party/flash-attention/hopper/flash_fwd_launch_template.h:175):
no kernel image is available for execution on the device
My environment:
GPU:
NVIDIA RTX 5090
Compute capability: sm_120
Software:
PyTorch: 2.7.1+cu128
xformers: 0.0.30
CUDA runtime: 12.8
Python: 3.11
OS: Ubuntu (WSL2)
From the traceback, the error appears to originate from xformers' flash-attention (hopper) kernel. It seems that the prebuilt xformers wheel does not contain a kernel image compiled for sm_120.
Interestingly, ‘python demo.py -p "A well-designed cozy bedroom" --use_sharp’ that do not trigger flash-attention run successfully in the same environment.
Could you please share the exact environment configuration used to test the demo? Specifically:
PyTorch version
xformers version
CUDA version
GPU model (compute capability)
Has this project been tested on RTX 5090 (sm_120)?
If flash-attention currently does not support sm_120, is there a recommended workaround (for example disabling flash-attention or using an alternative attention backend)?
Thank you very much for your help.