Added support for reading and writing F8_E4M3 and F8_E5M2 tensors with
torch. (#13, @TroyHernandez)
Fixed reading and writing tensors with empty names. Because x[[""]] returns
NULL for a named R list, access these tensors by position. (#10,
@TroyHernandez)
Fixed reading and writing tensors at offsets larger than 2 GB. (#14, @TroyHernandez)
Fixed writing bfloat16 tensors. (#11, @TroyHernandez)
Refactored the package so other packages can extend it. This is possible by:
safe_tensor_buffer and safe_tensor_meta methods.safetensors_frameworks.safe_load_file no longer defines a default framework. Set framework='torch' to get the previous behavior.
NEWS.md file to track changes to the package.