[guijacquemet]

Summary

1. On Windows computers with an NVIDIA driver, TensorFlow apps fail to install.

How it happens:

• post_install.bat installs requirements_gpu.lock.txt whenever it finds nvidia-smi.
• merge_requirements.py rewrote tensorflow to tensorflow[and-cuda] for every system.
• That extra lists nvidia-* packages with no platform marker, and some of them only have Linux wheels (nvidia-nccl-cu12 has never had a Windows wheel).
• The universal uv lock therefore pins them for Windows too, pip reports "No matching distribution", and the installer stops with "Application requirements installation failed".

The fix: in the GPU file, tensorflow now becomes two lines:

• tensorflow[and-cuda]==X ; sys_platform == "linux"
• tensorflow==X ; sys_platform != "linux"

If the line already has a marker, it is kept and combined with and. TensorFlow has had no native Windows GPU build since 2.10, so Windows loses nothing.

2. The version-cell snippet (notebook_upload.md) always printed "not listed" on Colab. It looked the notebook up at the top level of notebook_latest_versions.yaml, but sync_notebook_versions.py writes that file grouped by folder. The snippet now looks inside the folder groups.

Migration 0.1.21 → 0.1.22 applies both changes to existing apps:

• It only touches known blocks, and running it twice changes nothing.
• It keeps the files' line endings.
• It skips customised files and never creates missing ones.
• Notebooks that already contain the old snippet are not edited. The design note explains the one-line manual fix.

Found while planning a StarDist 2D (TensorFlow 2.20) LabConstrictor app.

Evidence

GPU lock for a StarDist stack (TF 2.20, Python 3.12), checked with uv pip install --dry-run --python-platform ...:

Lock	Windows x86_64	macOS arm64	Linux x86_64
0.1.21	No solution	n/a (CPU file is used)	OK
0.1.22	OK, 106 packages, no nvidia-*	OK	OK, 119 packages including nvidia-* (every one marked sys_platform == 'linux')

Tests

• python -m pytest tests -q: 235 passed, 246 subtests passed.
• New file tests/test_template_sync_0_1_22.py covers:
  • the GPU lines;
  • the version-cell snippet lookup;
  • the migration on 0.1.21 files, with LF and CRLF line endings;
  • running the migration twice (nothing changes the second time);
  • customised files and missing files;
  • the sync.py chain.
• test_template_sync_0_1_21.py: the manifest-version check now accepts any later version, and its chain test pins its copy of the manifest to 0.1.21. This is the same pattern the 0.1.20 test uses.
• ruff check and mypy are clean.
• Not verified: a real installer run on a Windows computer with an NVIDIA GPU. The lock results come from uv's resolver for each platform.

Checklist

• Docs that name the changed behaviour are updated: .tools/docs/notebook_upload.md (snippet), .tools/docs/notebook_requirements.md (what the GPU file does for TensorFlow), the new .tools/docs/migration_0_1_22.md, and a link in .tools/docs/README.md.
• Bumped template_version to 0.1.22. Users must regenerate the GPU lock and release a new installer.
• Screenshots still valid. No UI changed.

---

Generated by Claude Code