kherud/java-llama.cpp

3.4.0 packaging regression: CUDA libs packed with the main JAR

Closed this issue · 2 comments

shuttie commented

So the no-classifier JAR still seems to bundle the CUDA libs for some reason:

image

So it means that the official 3.4.0 JAR might not work on non-CUDA environments :(

kherud commented

I think the reason is this part of the earlier release workflow code:

      - uses: actions/download-artifact@v3
        with:
          name: artifacts
          path: ${{ github.workspace }}/src/main/resources/de/kherud/llama/
      - uses: actions/download-artifact@v3
        with:
          name: artifacts
          path: ${{ github.workspace }}/src/main/resources_linux_cuda/de/kherud/llama/

Both CPU and GPU artifacts used the same name, and the latter probably overwrote the former silently. Coincidentally, I just upgraded the workflow to use newer versions for the upload and download actions, which should fix the problem, I think. I'll create a second release in a second.

kherud commented

Version 3.4.1 seems to be correct now:

image

Good catch of the issue, though!

  • 👍1