Workaround for CUDA errors on Linux

[iwr-redmond]

There are multiple reports of image generation failing when using the CUDA version of Nexa SDK on Linux.

Unfortunately, the Nexa AI team are too busy to pay attention to this repository right now, so as a workaround I have compiled some Vulkan wheels for Linux that can be downloaded here.

To install the wheel for your version of Python, use the following command:

CMAKE_ARGS="-DGGML_VULKAN=on" pip install nexaai --prefer-binary --index-url https://huggingface.co/iwr-redmond/linux-wheels/resolve/main/index.html --extra-index-url https://pypi.org/simple --no-cache-dir

This should allow you to use the SDK as normal for both text and image inference. While I am unlikely to be able to offer much support for these custom wheels, please post in this thread if you encounter major difficulties.

[iwr-redmond]

This discussion relates to the former Python SDK and can now be closed.