Skip to content

v0.3.0

Choose a tag to compare

View all tags
@v-dziuba v-dziuba released this 21 Aug 20:48
· 188 commits to main since this release
v0.3.0
451ac5c

New Features & Improvements

  • Added support for int8 and int16 data types for the following operations: PACK, UNPACK, DIV, BROADCAST_TO, SQRT, and MAXIMUM.
  • Added support for SQUARED_DIFFERENCE (int8) and GATHER_ND, RESIZE_BILINEAR (a8w8, a16w8) operations.
  • Introduced a new CalibrationQsvAlignmentUtils class for aligning quantization parameters.
  • Improved the quantizer.save() process.
  • Added recipe-building utilities.
  • Added support for SELECT operations for int8 and int16.

Bug Fixes

  • Fixed an issue where a single consumer using the same tensor multiple times was only recorded once.
  • Fixed a bug in create_random_bool for requests with more than two elements.
  • Adjusted the clipping value for blockwise quantization.

Internal Changes

  • Refactored the is_conditionally_unquantized function.
  • Handled np.ndarray and bytes as constant buffers.
  • Deprecated emulated subchannel transformation.
  • Disabled narrow range for low bits quantization.
Assets 2
Loading