# Changelog

## [Unreleased]

### CLI
- Renamed `python -m qbcompiler mblt_compile` → `parse` and `python -m qbcompiler mxq_compile` → `quantize`. The `compile` end-to-end pipeline and all flags are unchanged. The `check` capabilities list now advertises `parse` / `quantize` in place of the old names.

## [v1.2.0] - 2026-06-26

### API
- Unified qbcompiler support for multiple hardware targets, including REGULUS and ARIES, into a single compiler.

## [v1.1.0] - 2026-03-31

### API
- The configuration of the mxq_compile function has been structurally refactored.
- For ONNX model parsing, graph optimization is now handled by the newly refactored parser based on mblt-graph.
- Improved quantization accuracy for non-linear activation functions.

## [v1.0.2] - 2026-02-12

### API
- Added support for all in inference_scheme.
- Fixed an issue where the optimize option was not being applied correctly.


## [v1.0.0] - 2026-01-30

### API
- Introduced support for Qwen3 LLMs.
- Expanded compatibility to include the YOLO26 series.
- Removed the deprecated singlecore compile and startdramoffset options.
- Added support for custom masks and dynamic RoPE.
- Added NPU-accelerated execution for parts of the pre-processing pipeline, along with new input process configuration.
  - Added support for uint8 inputs.
  - Added NPU-based normalization.
- Added support for Torch tensors and image files as calibration data.
- Removed redundant quantization configuration options and unified Percentile, MSE, and KL into a histogram-based observer.
- Added support for dynamic core and memory allocation for all models.


## [v0.12.0.0] - 2026-01-02

### API

- Refactored compilation and quantization configuration.
- Optimized CPU memory usage during compilation.
- Added support for Transformers v4.57.1 (aligned with the 0.12 Mobilint Docker release).
- Added support for the MiniCPM model.

## [v0.11.0.0] - 2025-09-10

### API

- Added support for Torch parser
- Added support for Yolo12l and Yolo12x
- Expanded support range for ViT models
- Fixed minor bugs for GRU/RNN/LSTM

## [v0.10.0.0] - 2025-07-24

### API

- Added Inference_scheme global4/global8 modes
- Added support for Yolov10 series, Yolo11 series, Yolo12n, Yolo12s, and Yolo12m
- Added LLM config options
- Added support for GRU/RNN/LSTM

## [v0.9.0.5] - 2025-06-25

### API

- Changed YOLO model decoding to be linked with the model zoo
- Added layer config options

## [v0.9.0.4] - 2025-05-22

### API

- Added support for direct compilation of HuggingFace LLM models

## [v0.9.0.3] - 2025-04-02

### API

- Added support for saving mblt using mmap
- Added support for diverse visualization types

## [v0.9.0.2] - 2025-01-16

### API

- Added API for saving model architecture
- Added support for Visualization tool
- Added support for Run min output difference QAT mode

## [v0.9.0.1] - 2024-12-13

### API

- Added API for global core (beta)
- Added API for YOLO post (beta)
- Added support for diverse hardware

## [v0.9.0.0] - 2024-11-27

### API

- Updated the high-level parsing processes

### Docker

- ONNX: 1.13.0 -> 1.16.2
- TensorFlow: 2.9.0 -> 2.17.0
- Torch: 1.13.0 -> 2.4.1

## [v0.8.5] - 2024-06-20

### API

- Added support for FastPercentile quantization method

## [v0.8.4] - 2024-05-20

### API

- Connected TF backend to ONNX backend by TF2ONNX
- Enabled compilation of models with custom input shape
- Supported more operations

### Docker

- ONNX: 1.12.0 -> 1.13.0

## [v0.8.3] - 2024-03-07

### API

- Added support for TF Lite backend

## [v0.8.2] - 2024-02-23

## [v0.8.1] - 2023-12-08

## [v0.8.0] - 2023-11-02

### API

- Deprecated TVM backend

## [v0.7.12] - 2023-09-12

## [v0.7.11] - 2023-08-31

### API

- Added support for TorchScript backend

## [v0.7.10] - 2023-08-11

## [v0.7.9] - 2023-08-11

## [v0.7.8] - 2023-08-08

## [v0.7] - 2023-03-23

- Added multi-channel quantization
- Supported more operations

### API

- Improved calibration dataset processing
- Added support for CPU offloading (beta version)

## [v0.6] - 2022-08-10

- Made minor updates

## [v0.5] - 2022-07-01

### Docker

- Switched from Conda to Virtualenv
- Python: 3.7.7 -> 3.8.10
- Torch: 1.8.1 -> 1.10.1
- TensorFlow: 1.15.0 -> 2.3.0
- ONNX:1.6.0 -> 1.11.0

### Parser

- Refactored code

### API

- Enabled saving sample inference results (inputs and outputs)

## [v0.4] - 2022-02-23

### Optimizer

- Made minor updates in fusing reshape

## [v0.3] - 2022-02-05

### Parser

- Identified preprocess and postprocess of the model
- Excluded preprocess and postprocess if they were unsupported by the NPU

### API

- Added integer inference simulation in the Python API

## [v0.2] - 2021-12-01

- First release