- Torch 2.2.
- Automatically detected cuda availability for some installation options.
- Released OCaml lock during backward passes, potentially improving async performance.
- Informed OCaml GC of tensor memory allocations, greatly reducing memory usage.
- Various compilation fixes in bin and examples.
- Fixed bug that returned OCaml floats (doubles) in single precision.
- Fixed bug that performed incorrect reductions in some cases.
- Added various new functionality, e.g. optional string arguments for generated functions, Scalar.to_int.