Upgrade PyTorch to 2.3.1+ to support NumPy 2.0

[tallamjr]

Summary

Currently, the project is pinned to PyTorch 2.3.0, which is incompatible with NumPy 2.0. This requires constraining NumPy to <2.0, preventing users from benefiting from NumPy 2.0's performance improvements and new features.

Current State

• PyTorch version: 2.3.0 (pinned in pyproject.toml)
• NumPy constraint: >=1.24,<2.0 (added as workaround)
• Issue: PyTorch 2.3.0 was compiled against NumPy 1.x C API and fails on Windows with NumPy 2.0+

Problem Details

Error on Windows with NumPy 2.0:

RuntimeError: Numpy is not available
UserWarning: Failed to initialize NumPy: _ARRAY_API not found

Affected operations:

• torch.from_numpy() conversions
• torchaudio.load() operations
• Sparse tensor operations
• Any torch/numpy interoperability

Root cause:
PyTorch binaries must be recompiled against NumPy 2.0's new C API to support both NumPy 1.x and 2.x. PyTorch 2.3.0 predates NumPy 2.0 (released June 2024) and lacks this compatibility.

Proposed Solution

Upgrade PyTorch to version 2.3.1 or later:

torch = [
    "torch>=2.3.1",  # NumPy 2.0 compatible
    "torchvision>=0.18.1",
    "torchaudio>=2.3.1",
]

Then remove the NumPy upper bound:

"numpy>=1.24",  # NumPy 2.0 supported with PyTorch 2.3.1+

Benefits

1. NumPy 2.0 features: Access to performance improvements and new functionality
2. Future compatibility: Aligns with ecosystem migration to NumPy 2.0
3. Simplified dependencies: Removes artificial version constraints
4. Better Windows support: Eliminates Windows-specific NumPy compatibility issues

Considerations

Python version compatibility:

• PyTorch 2.3.1: Supports Python 3.10-3.12
• PyTorch 2.4+: Adds Python 3.13 support (if we want to re-enable it in CI)
• PyTorch 2.9+: Adds Python 3.14 preview support

Testing requirements:

• Verify all tests pass on Linux, macOS, Windows
• Check for any breaking changes in PyTorch 2.3.1+ release notes
• Regenerate uv.lock after version updates

References

Upstream issues:

• PyTorch #107302: NumPy 2.0 Support
• PyTorch #131668: Release 2.4 windows wheels not compatible with numpy 2.0
• PyTorch #135013: NumPy 2.1.0 Compatibility Issue

PyTorch releases:

• PyTorch 2.3.1 Release Notes
• PyTorch 2.4.0 Release Notes

Related PRs:

• augmentation method in numpy: flip up/down #2: Current CI modernisation (adds numpy<2.0 workaround)

Acceptance Criteria

• Upgrade PyTorch to 2.3.1+ (or 2.4+ if Python 3.13 support desired)
• Remove <2.0 constraint from NumPy dependency
• Regenerate uv.lock file
• All CI tests pass on Linux, macOS, Windows
• Update documentation if there are any breaking changes
• Consider if torchvision/torchaudio versions need updating for compatibility

Priority

Medium - Current workaround (numpy<2.0) is functional, but upgrading provides better long-term compatibility and performance.