epic: llama.cpp is installed by default

[dan-menlo]

Goal

Cortex.cpp should have a super easy UX to on par with market alternatives

• User should have a 1-click installer, that prioritizes simple UX over size-complexity
  • Installer packages (or downloads at install time) llama.cpp binaries (e.g. up to 1gb)
  • Installer optimizes for "universal" installers (i.e. download all options, and then subsequently deletes all unnecessary files)
  • e.g. Mac Universal includes llama.cpp Mac for Intel and Apple Silicon
  • e.g. Windows and Nvidia Universal includes llama.cpp for both CUDA versions
• For this epic, I am open to either:
  • Pre-packaging (preferred)
  • Install-time download of dependencies: https://github.com/janhq/cortex.cpp/pull/1219/files

Idea

I wonder whether the solution to this is a way to have an optional local lookup, as part of cortex engines install:

• Installer can look in its installer folder to see if dependencies are available, and only pull from remote if needed
• We do not need to make any changes to the installer (it still just runs cortex engines install)
• This approach is elegant, and allows us flexibility in packaging

Out-of-scope (future)

• We should offer a "cortex-alpine" installer which has minimal file size
  • Targeted for embedded use cases, and if people want to use ONNX or TensorRT-LLM without llama.cpp
  • User will have to download engines as a post-install step
• We should offer "universal" installers that pre-package all potential dependencies
  • e.g. Large installer size, but packages all dependencies for offline install

Outcomes

• Cortex.cpp installer should install llama.cpp by default
• Cortex.cpp installer should install the correct version of llama.cpp (based on hardware)

Key Questions

• Should we align with llama.cpp's versions? (e.g. with Vulkan, sycl)

Appendix

Why?

Our current cortex.cpp v0.1 onboarding UX is not user friendly:

• llama.cpp only seems to be downloaded on first run of Cortex (at least on Windows)
  • Download UX is not very good (no progress indicator)
  • Download is often slow, or drops

• Very often, the llama.cpp engine download does not work, resulting in bad UX
  • "Engine not loaded yet"

[dan-menlo]

@hiento09 has worked on a PR that moves the llama.cpp Engine Install to Installer, but I am concerned that it's still not great UX

https://github.com/janhq/cortex.cpp/pull/1219/files

[namchuai]

Just saw this today. "Engine not loaded yet" does not mean the engine is not yet downloaded. It might have problem with the engine loading logic.

[freelerobot]

QA Updates (v75)

• ✅ works on windows
• ✅ works on mac