Skip to content

fix(install): complete macOS setup, plus a double-clickable macOS app bundle - #6

Open
fabio-krauss wants to merge 2 commits into
estebanstifli:mainfrom
fabio-krauss:fix/macos-install-and-app-bundle
Open

fabio-krauss wants to merge 2 commits into
estebanstifli:mainfrom
fabio-krauss:fix/macos-install-and-app-bundle

Conversation

@fabio-krauss

Copy link
Copy Markdown

What changed

Two commits, both aimed at making a fresh macOS install actually finish. Neither touches src/, so behaviour on Windows/Linux is unchanged.

1. fix(install): complete macOS setup for PyTorch and llama.cpp

  • PyTorch on Darwin: use the standard PyPI wheel (torch==2.13.0, no --index-url), mirroring the branch scripts/install_torch.py already has. The +cpu local-version builds only exist for Windows/Linux, so torch==2.13.0+cpu against the cpu index never resolved on macOS and set -e aborted the script right after venv creation — before any dependency was installed. The CUDA path and the Windows install.ps1 behaviour are deliberately untouched.
  • llama.cpp: the published macOS arm64 wheels for 0.3.31–0.3.36 are corrupt (full Content-Length, members failing CRC-32). The previous fallback only fired for AI_BACKEND=auto && LLAMA_BACKEND=cuda, so on macOS a corrupt wheel hit exit 1. Now the step tries the newest compatible release, then the newest release known to have intact wheels, then the portable CPU index, and only fails if all of them fail. Once fix(ci): prevent corrupted Metal wheel releases abetlen/llama-cpp-python#2379 lands the first attempt succeeds again and the pinned fallback is never reached.

Fixes #5.

2. feat(macos): a double-clickable app bundle and app icon

macOS had no launcher at all — no Dock icon, nothing in Finder, no /Applications entry; Windows has start.bat. scripts/create_macos_app.sh builds Meet2Notes.app from the existing source install using only tools that ship with macOS (sips, iconutil), with no new downloads:

  • .icns icon in every size macOS asks for, generated from assets/macos/Meet2Notes-1024.png
  • an Info.plist and a launcher that resolves the checkout and execs .venv/bin/meet2notes, so the browser opens on 127.0.0.1:8765 as usual
  • a Finder-appropriate PATH (so Homebrew FFmpeg is found) and ~/Library/Logs/Meet2Notes.log for launches with no terminal attached
  • unsigned but quarantine-free by construction; --install also copies it to /Applications

The icon is a placeholder I drew, not a branding decision — replace assets/macos/Meet2Notes-1024.png with whatever you prefer and the script regenerates everything from it. The .app is a build artifact and stays gitignored.

Validation

  • bash -n install.sh and bash -n scripts/create_macos_app.sh
  • real macOS install (26.7, arm64 M3, Homebrew Python 3.13.9, fresh clone): ./install.sh --ai-backend auto --no-models → Meet2Notes is ready.
  • pip check → no broken requirements; check_environment.py reports Python, platform and FFmpeg paths
  • server started and answered HTTP 200 on http://127.0.0.1:8765
  • ./scripts/create_macos_app.sh → bundle built, launcher resolves the venv and runs the CLI (--help routed to the log file)
  • corrupt-wheel claim verified with curl -L + zipfile.testzip() against 0.3.25/0.3.30 (intact) and 0.3.31–0.3.36 (corrupt), and reproduced with pip itself

ruff/mypy/pytest are unaffected (no Python source changed). If you'd rather review these separately, say so and I'll split commit 2 into its own PR — and if you'd prefer the launcher to be a Pinokio/installer concern instead of a script in the repo, happy to move it.

Use the standard PyPI PyTorch wheel on Darwin, matching scripts/install_torch.py.
The download.pytorch.org '+cpu' local-version builds are Windows/Linux only, so
'torch==2.13.0+cpu' from the cpu index never resolves on macOS and the script
exited before installing any dependency.

Also treat the llama.cpp wheel step as best-effort: the published arm64 wheels
for 0.3.31-0.3.36 fail their CRC-32 checks, and the existing fallback was
CUDA-only, so a corrupt wheel aborted the whole install on macOS. Try the newest
compatible release, then the newest release known to have intact wheels, then the
portable CPU index.
Build Meet2Notes.app from the existing source install with
scripts/create_macos_app.sh: an .icns generated with sips/iconutil, an Info.plist
and a launcher that resolves the checkout and runs .venv/bin/meet2notes. Finder
launches get a usable PATH and a log file, and the bundle is unsigned and
quarantine-free by construction. The bundle is a build artifact and stays
gitignored. Windows already has start.bat; macOS previously had no launcher at
all.

Copy link
Copy Markdown
Owner

Thanks, Fabio, for the detailed report, the fixes, and the testing on your Mac — especially for tracing the wheel corruption upstream.

The Darwin PyTorch fix looks good. Before merging, I’d like to address two points: preserve the project’s dependency constraints during the llama.cpp retries (particularly numpy<2.4), and safely escape the checkout path in the generated app launcher. The current substitution breaks paths containing & and can misinterpret shell characters.

The Pinokio installation path mentioned in #5 also remains affected. Could we cover that too, or explicitly track it separately? Splitting the optional app bundle into a separate PR would help us land the installation fix independently.

Thanks again for contributing and offering to test on macOS!

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

install.sh cannot complete on macOS: PyTorch pin never resolves, llama.cpp step aborts on corrupt upstream wheels

2 participants