Python 3.12 Compatibility

Last updated: 2026-06-23 Status: Python 3.12 is not yet tested or supported. Upgrade dependencies pending FFCV verification.

Current Status

DynVision currently targets Python 3.11 (python_requires = ">=3.11,<3.14"). Python 3.12 support is planned but blocked on FFCV compatibility testing.

Dependency Compatibility Matrix

Dependency	Current Spec	3.12 Compatible?	Minimum 3.12-Compatible Version
scipy	==1.12	✅ Yes	No bump needed
snakemake	~=9.1.5	✅ Yes	No bump needed
numba	~=0.61.0	✅ Yes	No bump needed
ruamel.yaml	~=0.18.3	✅ Yes	No bump needed
python-dotenv	~=1.0.1	✅ Yes	No bump needed
scikit-image	~=0.20.0	✅ Yes	No bump needed
matplotlib	~=3.5.1	✅ Yes	No bump needed
seaborn	~=0.13.0	✅ Yes	No bump needed
pillow	==10.0.0	✅ Yes	No bump needed
tqdm	~=4.66.0	✅ Yes	No bump needed
torchsummary	~=1.5.0	✅ Yes	No bump needed
jupyterlab	~=4.0.0	✅ Yes	No bump needed
ipympl	~=0.7.0	✅ Yes	No bump needed
wandb	(unpinned)	✅ Yes	No bump needed
tensorboard	(unpinned)	✅ Yes	No bump needed
torch	>=2.2.0	❌ No	>=2.4.0
torchvision	>=0.16.0	❌ No	>=0.19.0
pytorch-lightning	>=2.0.0	❌ No	>=2.4.0
scikit-learn	>=1.2.0,<2	❌ No (needs 1.4+)	>=1.4.0
ffcv	~=1.0.2	⚠️ Unclear	Testing required

Key Risk: FFCV

FFCV 1.0.2 was last released on 2023-03-05, before Python 3.12 was released (2023-10-02). Key concerns:

• No official Python 3.12 wheels or CI for FFCV
• The Ubuntu/Debian package was removed for Python 3.12 due to reported incompatibility
• FFCV depends on numba, which historically had Python 3.12 issues (resolved in numba 0.59.0+)
• The upstream GitHub repository (libffcv/ffcv) has not released a Python 3.12-compatible version

Mitigation Options

If FFCV proves incompatible with Python 3.12:

1. Stay on Python 3.11 for the 0.1 release — safest option, no code changes needed
2. Make FFCV optional — allow fallback to standard PyTorch DataLoader when FFCV is unavailable
3. Replace FFCV — alternatives include:
   • WebDataset — streaming I/O for large datasets
   • NVIDIA DALI — GPU-accelerated data loading
   • Standard PyTorch DataLoader with appropriate tuning

Version Bumps Required for 3.12

To update pyproject.toml for Python 3.12 support, the following changes are needed:

 dependencies = [
     ...

-    "torch >=2.2.0",
-    "torchvision >=0.16.0",
+    "torch >=2.4.0",
+    "torchvision >=0.19.0",
     ...

-    "pytorch-lightning >=2.0.0",
+    "pytorch-lightning >=2.4.0",
     ...

-    "scikit-learn >=1.2.0,<2",
+    "scikit-learn >=1.4.0,<2",
     ...
     "ffcv ~= 1.0.2",    # ⚠️ Test Python 3.12 compatibility
 ]

And update the Python version constraint:

- python_requires = ">=3.11,<3.14"
+ python_requires = ">=3.12,<3.14"
+ classifiers = [
+     "Programming Language :: Python :: 3.12",
+ ]

Testing Plan

1. Create a Python 3.12 conda environment
2. Install DynVision with updated dependencies
3. Run the test suite: pytest tests/ -v
4. Test FFCV data loading specifically:

   from ffcv.loader import Loader
   # Verify basic loading works on 3.12
   

5. If FFCV fails, implement fallback option or document 3.11-only for 0.1

Recommendation for 0.1 Release

Keep Python 3.11 as the minimum requirement for 0.1. The risk from FFCV incompatibility with 3.12 is too high to resolve without thorough testing. Add explicit 3.12 support in a point release (0.1.1 or 0.2.0) after FFCV testing is complete.

Set python_requires = ">=3.11,<3.14" to allow forward compatibility while clearly documenting the 3.12 testing status.