safetensors 0.3.0

• Added support for reading and writing F8_E4M3 and F8_E5M2 tensors with torch. (#13, 1)

• Fixed reading and writing tensors with empty names. Because x[[""]] returns NULL for a named R list, access these tensors by position. (#10,

• Fixed reading and writing tensors at offsets larger than 2 GB. (#14,

• Fixed writing bfloat16 tensors. (#11, 1)

safetensors 0.2.1

• Changed maintainer to Tomasz Kalinowski.

safetensors 0.2.0

• Refactored the package so other packages can extend it. This is possible by:

  1. Implementing the safe_tensor_buffer and safe_tensor_meta methods.
  2. Registering the framework with the reflection safetensors_frameworks.

• safe_load_file no longer defines a default framework. Set framework='torch' to get the previous behavior. # safetensors 0.1.2

• Added support for BF16 data types.

safetensors 0.1.1

• Added a NEWS.md file to track changes to the package.
• Copy tensors to cpu before proceeding with serialization. (#2)