Release 2024-04 follow-up #1356
Description
release date: end august (may be delayed early september waiting pytorch and numba)
wanted:
- technical debt removal:
- no more NullSoft usage to create the launchers (only mkshim400.py + pywin32)
- no more inno setup code
- no more NullSoft installer code
- Fitness challenge:
- a slim version < 650 Mo , so it can be checked on virustotal.com
- standard version < 900 Mo
- easing Security Checking challenge:
- releasing a .7z big distro seems now a better idea than .exe, as Windows now supports unpacking .7z
- as interim, a slim build with .exe format will be provided
- numpy-2.0.1
- numpy-2+ may require torch-2.4.1 and numba-0.61 ?
focus:
- simplification of build and maintenance:
- drop packages.ini, use only package provided summary
- further simplification or removal of internal code
- less packages... less updates to check
- remove NSIS use for launchers from ALL launchers
- remove redundant "WinPython Terminal" launcher
- remove all Inno Setup code
- Python needs for a build to get down to WinPythondot: flit + packaging + Pywin32
- simplification of end-user checkings:
- slim version
- no auto-executable archives the big version
postpone (no time):
- jupyterlite more: webnn may be a thing: https://webmachinelearning.github.io/webnn-intro/
- llm and generative AI: whisper.cpp
- replacing torch per jax[cpu] would shrink WinPython by 100Mo (or not increase per 200 More) ?
- WinPythondot being enough to build a WinPython...
- free-threading effort delayed to 3.14, and eco-system is a big mountain to climb:
- essential-for-speed patches will only be in 3.14 cycle, 3.13t will have accepted leakages ... it's a Work-in-Process state
- pyzmq doesn't support multi-threads
- so... no cython, no jupyterlab (until pyzmq replaced), .... hummmm end-game is looking year 2026
- free-threading test:
- llm: AI agents and the 7b models, whispers made easy coming to town ?
Direction:
- Help to Python Sustainability improvement:
- favorise latest=more efficient and maintained python
- favorise speed = Free-threading (complex) revolution
- favorise external data motors: dataframe interchange protocol, numpy-2, pandas-3, scikit-learn integrating polars,
- favorise AI: onnx (for AI PC), llvm local small equivalent to ggerganov/whisper.cpp
- Help to Python Reachability:
- in browser (and WebApp): Jupyterlite prototype (https://stonebig.github.io/my_demo_4/lab/index.html),
- WASI: in VSCode
- inside Excel 365: asks Operational Research packages
- in Windows Store: a requirement file- remove need of build pieces: nsis, 7zip
- PEP-751:
- WinPython ideal world is to shrink to a toml file specifying package/version/hash