Change Log

DX-Compiler v2.4.0 (2026-06-16)

DX-Compiler Version

• DX-COM: v2.4.0
• DX-TRON: v2.0.1 (Deprecated)

Deprecation Notice — DX-TRON

Starting with DX-Compiler v2.4.0, DX-TRON is deprecated and will be removed in a future release. No further feature updates or bug fixes are planned for DX-TRON. We recommend migrating to the new HTML graph viewer bundled with DX-COM (see Interactive HTML Graph Viewer below) for model inspection and visualization.

Deprecation Notice — PPU Type 2

Starting with DX-Compiler v2.4.0, the legacy PPU type 2 post-processing mode (used for YOLOv8-family DFL post-processing) is deprecated and will be removed in a future release. Compiling a model that uses PPU type 2 now emits a deprecation warning. Please migrate to the new dx_com.pre_optimize() API with the built-in YOLO post-processing passes (see Added below).

Changed

• PPU Type 2 Deprecated: The legacy PPU type 2 post-processing mode is now deprecated and emits a deprecation warning when used. See the Deprecation Notice above for the migration path.
• Faster, Lighter Compilation: Reduced memory usage and improved compile time, especially on large models.
• Expanded Python Version Support: In addition to Python 3.8–3.12, DX-COM now supports Python 3.13 and 3.14.

Fixed

• Fixed a Python API issue where models expecting integer inputs were sometimes fed float data, causing accuracy degradation.
• Fixed several Q-PRO / DXQ quantization crashes and stability issues observed on real models.
• Fixed multiple compilation errors and runtime issues caused by tiling, partitioning, and memory allocation in models containing Split, Concat, Reshape, Bilinear Resize, Clip, or odd spatial dimensions.
• Fixed compatibility issues with NumPy 2.4+ and onnxruntime ≥ 1.25.0.
• Reduce ExpandDim Preprocessing In Config For GrayScale: When converting to GrayScale color, one channel was automatically squeezed, which caused a dimensional difference compared to the RGB image. So additional ExpandDim preprocessing is needed before. Therefore, we resolve this problem and preprocessing config to be identical to that of RGB images.

Added

• Automated Q-PRO Configuration: Q-PRO quantization (formerly available only by hand-picking DXQ combinations) is now much easier to use. DX-COM can now automatically generate DXQ combinations for you and run Q-PRO under the hood, removing the need to manually tune the many DXQ knobs to get higher-accuracy quantization.
• Quantization-Aware Training (QAT): Added end-to-end QAT support directly through dx_com.compile(). When the supplied config JSON includes a qmaster block, dx_com.compile() automatically switches to QAT mode and runs the training pipeline using the same dataset settings as PTQ calibration. Available from both the dxcom CLI and the Python API. A new fast_run flag is also available for quick QAT smoke tests. See Quantization-Aware Training (QAT).
• QXNN Resume (Re-quantization without Recompile): Added a checkpoint-based QXNN resume flow available from both the dxcom CLI and the Python API. Once a model has been compiled, users can re-run quantization with different settings (e.g., a different calibration method) without repeating the earlier compile phases, dramatically shortening the iteration loop when tuning quantization quality.
• Quantization Diagnosis Report (HTML): Added an HTML report that visualizes per-layer quantization quality, highlights problematic layers, and includes ready-to-paste compile snippets to retry compilation with recommended settings. Enabled via the new quant_diagnosis option, available from both the dxcom CLI and dx_com.compile().
• Interactive HTML Graph Viewer (replaces DX-TRON): DX-COM now produces a standalone HTML viewer for inspecting compiled models, including parameter shapes, CPU/NPU partition reasons, and cross-subgraph connections. This replaces the DX-TRON workflow (see Deprecation Notice above).
• dx_com.pre_optimize() API: Added a new top-level dx_com.pre_optimize() API for applying ONNX-level pre-processing transforms before compilation, with built-in support for YOLO post-processing integration (detection and segmentation modes). See the Pre-Optimize API chapter.
• Ubuntu 26.04 Validation: DX-COM is now validated on Ubuntu 26.04, in addition to the previously supported Linux distributions.
• DEEPX Agent-Driven Development (dx-agent-dev) — Beta: Compile models with a natural-language prompt. An AI coding agent drives DX-COM to turn ONNX/.pt (including Ultralytics format=deepx) into a deployable .dxnn, with ONNX-vs-DXNN numerical verification. (Beta — behavior may change.)

Known Issues

• Significant FPS degradation has been observed in models using PReLU as an activation function.

---

DX-Compiler v2.3.1 (May 2026)

DX-Compiler Version

• DX-COM: v2.3.0
• DX-TRON: v2.0.1

Fixed

• Fixed uninstall.sh not removing installed packages and extracted module directories.
  • dx_com Python package is now properly uninstalled via pip3 uninstall before the virtual environment is removed.
  • dxtron Debian package is now properly removed via apt-get remove.
  • dx_com/ and dx_tron/ directories are now deleted on uninstall.

Added

• Added --target=<dx_com|dx_tron|all> option to uninstall.sh (default: all), consistent with install.sh.

---

DX-Compiler v2.3.0 (March 2026)

DX-Compiler Version

• dx_com : 2.3.0
• dx_tron : 2.0.1

Changed

• Python Packaging: Relaxed Python package dependency constraints to improve installation flexibility.

• Distribution: The standalone DX-COM executable distribution is deprecated. The user manual now documents the wheel-based workflow only.

• Performance:

  • Reduced NPU inference latency for most models.
  • Improved compiler performance to reduce compilation time.

Fixed

• Compiler Stability: Fixed various DX-COM compiler stability issues.

Added

• TopK-Optimized Post-Processing Pipeline: Added an optimized post-processing pipeline for supported DFL-based YOLO models that applies TopK filtering before bounding box decoding, reducing CPU post-processing workload and improving runtime efficiency.
• Batched Convolution Support: Added support for batched convolution.
• Linux Distribution Validation: In addition to Ubuntu 20.04, 22.04, and 24.04, DX-COM v2.3.0 was also validated on Fedora 42-45, Red Hat 9-10, and CentOS Stream 9-10.

---

DX-Compiler v2.2.1 (February 2026)

DX-Compiler Version

• dx_com : 2.2.1
• dx_tron : 2.0.1

Changed

• None

Fixed

• DXQ Quantization: Fixed DXQ enhanced quantization option bugs.
• Python Wheel Package: Fixed PPU compilation bug for Python 3.8, 3.9, and 3.10.
• Input Validation: Fixed an issue where compilation proceeded without error when invalid model input names were specified.

Added

• GPU Quantization (JSON Config): Added quantization_device support in JSON configuration file, enabling GPU-accelerated quantization via CLI (dxcom) in addition to the dx_com Python module. Available only with the Python wheel package installation.
• GPU Auto-Detection: When quantization_device is not specified, DX-COM now automatically uses GPU if a CUDA-compatible GPU is available, otherwise falls back to CPU.

---

DX-Compiler v2.2.0 (December 2025)

DX-Compiler Version

• dx_com : 2.2.0
• dx_tron : 2.0.1

Changed

• None

Fixed

• Model Accuracy: Resolved accuracy degradation issue in the DeepLabV3PlusMobilenet-1 model from DX ModelZoo.

Added

• PPU Support: Extended PPU support to YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12.
• Python Wheel Package: DX-COM is now available as a Python wheel package (in addition to the existing executable file), supporting Python 3.8, 3.9, 3.10, 3.11, and 3.12, enabling programmatic compilation using torch DataLoader without configuration files.
• Multi-Input Model Support: Added support for multi-input models through torch DataLoader in the Python wheel package.
• DX-TRON:
  • Debian package (.deb) installation support
  • Local web browser hosting (dxtron command)
  • Ubuntu 24.04 support

---

DX-Compiler v2.1.0 (November 2025)

DX-Compiler Version

• dx_com : 2.1.0
• dx_tron : 2.0.0

Changed

• Command-Line Interface: Removed deprecated command-line options: --jobs, --shrink, --info (or -i).

• ONNX Support:

  • Removed restrictions on Split, Transpose, Reshape, Flatten, and Slice operators.
  • Clarified ONNX opset version support (versions 11-21 are supported; version 22 and above are not supported).

Added

• Command-Line Interface:

  • --aggressive_partitioning: Enables aggressive partitioning to maximize operations executed on NPU.
  • --opt_level {0,1}: Controls optimization level (default: 1).
  • --compile_input_nodes and --compile_output_nodes: Support for partial compilation.

• ONNX Support: Added support for Gather operator.

• Quantization: Reintroduced the DXQ enhanced quantization option (enhanced_scheme, DXQ-P0 to DXQ-P5), previously removed in dx_com v2.0.0.

• PPU (Post-Processing Unit): Reinstated PPU support.

  • Supported models: YOLOv3, YOLOv4, YOLOv5, YOLOv7 (anchor-based), YOLOX (anchor-free).

---

DX-Compiler v2.0.0 (September 2025)

ONNX Support

• Re-enabled support for the following operators:

  • Softmax
  • Slice

• Newly added support for the following operator:

  • ConvTranspose

Model Support

• Partial support for Vision Transformer (ViT) models

• Verified with the following OpenCLIP models:

  • ViT-L-14, ViT-L-14-336, ViT-L-14-quickgelu
  • RN50x64, RN50x16
  • ViT-B-16, ViT-B-32-256, ViT-B-16-quickgelu

Compatibility and Deprecations

• Compatibility with DX-RT versions earlier than v3.0.0 is not guaranteed
• The DXQ enhanced quantization option (enhanced_scheme) was removed in dx_com v2.0.0 and reintroduced in dx_com v2.1.0
• PPU (Post-Processing Unit) is no longer supported, and there are no current plans to reinstate it

---

DX-Compiler v1.60.1 (June 2025)

Bug Fixes

• Internal bug fixes

Command-Line Interface Updates

• Added support for:
  • -v option: Displays DX-COM module version
  • -i option: Displays internal module information
    → For usage, see: CLI Execution

ONNX Support

• The following operators were deprecated and are scheduled to be re-supported in a future release:
  • Softmax
  • Slice

---