Changelog

All notable changes to this package will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[2.5.0] - 2026-01-23

Added

  • PyTorch model import
  • LRN (Local Response Normalization) operator implemented on all backends
  • 3D MaxPool and AveragePool operators implemented on all backends
  • Sentis Importer: Allow users to specify dynamic dimensions as static on Sentis model import, same as we do for ONNX
  • Tokenizer Additions
    • Hugging Face parser
    • Sequence decoder
    • Regex replace decoder
    • String split pre-tokenizer
    • Unigram Mapper
    • Byte-based substring feature to SubString
    • Padding: support "pad multiple of" option
    • Split pre-tokenizers: support "invert"
    • StripAccents normalizer
    • Rune split pre-tokenizer
    • Strip normalizer
    • WordLevel model
    • WhitespaceSplit pre-tokenizer
    • Metaspace pre-tokenizer and decoder
    • Whitespace pre-tokenizer
    • NMT normalizer
    • Punctuation pre-tokenizer
    • Digits pre-tokenizer
    • CharDelimiterSplit pre-tokenizer
    • BPE decoder

Changed

  • Model Visualizer: Async loading of model
  • Model Visualizer: updating com.unity.dt.app-ui to 1.3.3
  • Resize operator on CPU no longer uses main (mono) thread path
  • All model converters use switch-case instead of if-else cascade
  • Migrate Mono APIs to CoreCLR-compatible APIs

Fixed

  • Editor crash when quitting in play mode
  • Memory Leak in FuseConstantPass
  • Clip operator improvement: no longer need CPU fallback for min/max parameters
  • Mod operator fix: on some platform with float operands, could return incorrect value when one of them was 0
  • Faulty optimization pass
  • Fix in existing burst code for 2D pooling vectorization calculations
  • TopK issue on GPUCompute when dimension is specified
  • Fix source generator empty array
  • Tokenizer Fixes
    • Special added token decoding condition
    • Fix added token whole word handling
    • Gpt2Splitter subtring length computation
    • Added vocabulary pre-tokenization.
    • ByteLevelDecoder empty-byte guard in string generation
    • DefaultDecoder: joining tokens with whitespace
    • BPE: fix merging, applying on each word instead of the whole string
    • DefaultPostProcessor: apply the proper type id
    • RobertaPostProcessor: fix attention and type id assignment
    • TemplatePostProcessor: fix type id assignment
    • Assign default type id to sequences
    • Better surrogate characters support
    • Fix ByteFallback: inserting the right amount of \ufffd char
    • Fix BertPreTokenizer
    • Default model determination based of chain of responsibility

[2.4.1] - 2025-10-31

Fixed

  • Small error in documentation preventing user manual publication

[2.4.0] - 2025-10-22

Added

  • LiteRT model import
  • Tokenization API
  • STFT and DFT ONNX operators
  • BlackmanWindow, HammingWindow, HannWindow and MelWeightMatrix ONNX operators
  • BitwiseAnd, BitwiseOr, BitwiseXor, BitwiseNot ONNX operators and functional methods
  • AsStrided, Atan2, Expm1, Log10, Log1p, Log2, Rsqrt, Trunc, ReduceVariance, Diagonal layers, functional methods and optimizer passes
  • NotEqual, FloorDiv, TrueDiv layers and LiteRT operators

Changed

  • Renamed Inference Engine to Sentis in package name and documentation
  • Improved model import time for ONNX models
  • ONNX model import operator order now consistent with the original model
  • Improved optimization passes to reduce operator count in imported models
  • Improved visualizer loading times and consistency in displaying attributes
  • ScatterND operator can now run on much larger tensors, enabling new models
  • ScatterND operator now allows negative indices
  • ONNX model outputs that are not connected to any inputs are no longer incorrectly pruned
  • Improve model import warning and error display in the inspector

Fixed

  • Small errors in documentation
  • Faulty optimization passes that could lead to inference issues
  • Memory leaks on model constants
  • Non-matching ProfilerMarker calls
  • Issues in CPU callback which could lead to incorrect inference on some models
  • Enable missing modes for GridSample and Upsample operators

[2.3.0] - 2025-07-15

Added

  • Model Visualizer for inspecting models as node-based graphs inside the Unity Editor
  • Support for Tensor<int> input for GatherND operator on GPUPixel backend
  • Support for Tensor<int> input for the base of the Pow operator on all backends
  • Support for the group and dilations arguments for the ConvTranspose operator on all backends
  • Support for value_float, value_floats, value_int and value_ints values in ONNX Constant operators

Changed

  • Optimized single-argument operators on CPU backend
  • Optimized deserialization of models to avoid reflection at runtime

Fixed

  • Einsum operator now works correctly on fallback path

[2.2.1] - 2025-05-28

Fixed

  • Issue with incorrect TensorShape in Conv layer when dilations are greater than 1 and auto-padding is used
  • Incorrect Third Party Notices

[2.2.0] - 2025-05-15

Added

  • First version of Inference Engine