memory_efficient_attention fails with half precision tensors, with pre-compiled xformers 0.0.30 on RTX 5090

Bug

I am encountering an error with xformers 0.0.30 memory_efficient_attention on RTX 5090, with half-precision tensors. I am using torch 2.7 with cuda 12.8.

Calling the function returns the following error:

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

Note that it works in full precision.

Also, I tried installing the latest version from source (commit 8fc8ec5a4d6498ff81c0c418b89bbaf133ae3a44), and I cannot reproduce the issue, it works as expected. I installed it with poetry run pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers

Command to reproduce

from xformers.ops import memory_efficient_attention
import torch
q, k, v = torch.randn(5, 10, 128).half().cuda(), torch.randn(5, 10, 128).half().cuda(), torch.randn(5, 10, 128).half().cuda()
memory_efficient_attention(q, k, v)

Returns

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

Environment

You can run the script with:

• PyTorch Version (e.g., 1.0): 2.7
• OS (e.g., Linux): ubuntu 24.04.2
• How you installed PyTorch (conda, pip, source): poetry/pip
• Build command you used (if compiling from source):
• Python version: 3.12
• CUDA/cuDNN version: 12.8 / Nvidia driver version: 570.133.20
• GPU models and configuration: RTX 5090
• Any other relevant information: