• English
  • 3.10.0

    • Release Notes

    Notes Before Use

    Attention

    This version is applicable to J6E/M/P/H/B and S6P, it is the official release of the algorithm toolchain, the version properties based on different chip platforms are as follows:

    • For J6E/M Linux platform: Commercial-ready version.

    • For J6E/M QNX platform: Restricted commercial version.

    • For J6P/H Linux platform: Commercial-ready version.

    • For J6B QNX platform: Commercial-ready version.

    • For S6P Linux platform: Integration version.

    Please make sure to read the “J6 Algorithm Toolchain License Agreement” carefully before using the toolchain. If you choose to continue using the toolchain, it means that you have fully understood, comprehended and accepted the “J6 Algorithm Toolchain License Agreement”.

    Release information

    Horizon OpenExplorer 3.10.0(release_date:202609)

    ModuleVersion
    HMCTv2.9.3
    HORIZON_TC_UIv3.5.20
    HORIZON_Plugin_Pytorchv3.4.6
    HBDK_Compilerv4.13.3
    UCP_Tutorialv3.16.3
    J6EM-Linux-BSPv4.0.0, v3.1.0, v3.0.0, v2.1.0
    J6EM-QNX-BSPv4.0.0, v3.0.0
    J6PH-Linux-Matrix-BSPv4.0.0, v3.0.0, v2.0.0, v1.0.0
    J6B-QNX-BSPv4.0.0, v3.0.0, v2.0.0, v1.0.0
    S6P-Linux-BSP00.02.22

    Main Functional Updates

    1. On the J6P platform, Horizon_plugin_pytorch supports configuring Softmax and LayerNorm to use high-precision output, for details, see the description of PropagateTemplate in the Default Optimization Template Pass Description section.

    2. The Horizon_plugin_pytorch accuracy debugging tool adds four core features: custom metric registration, badcase visualization, layer comparison recommendations, and sensitivity optimization, improving the efficiency of accuracy debugging. For details, see the PyTorch Model Accuracy Tuning Tool section.

    3. Horizon_plugin_pytorch adds optimization for the input range of activation functions. It is recommended to enable this option by default. For details, see OptimActFuncInputRangeTemplate.

    4. Horizon_plugin_pytorch adds optimization for the output clipping logic of activation operators. It is recommended to enable this option by default. For details, see Handling Operators with Unbounded Output Range.

    5. Horizon_plugin_pytorch adds the get_calib_qconfig_setter API, which supports obtaining a dedicated QconfigSetter during the Calibration stage. For details, see Reusing Calibration Results when Adjusting Quantization dtype.

    6. The Toolchain official website now provides Model Lab, which offers reference models and deployment metrics for Horizon BPU. Typical models for scenarios such as intelligent driving and large language models are ready to use out of the box, with performance at a glance.

    7. HBIR model simulation on the X86 platform now supports acceleration using AVX-512 CPU instructions. For details, see the quantized.bc Simulation Acceleration and Result Consistency section.

    8. HBDK adds the hbm_strip_debug API, which supports removing debug information from HBM models. For details, see hbm_strip_debug.

    Compatibility Statement

    None.

    Use Restrictions

    The inference preemption function for models containing the TopK operator depends on the BSP version, which needs to meet: J6EM-Linux >= v3.1.0, J6PH-Linux-Matrix >= v3.0.0, J6B-QNX >= v2.0.0.

    Deprecated Functions

    The OE sample package no longer includes model samples. For model deployment, refer to the Model Lab page for details.

    Fix Issues

    Fix the issue where the simulation inference speed is slow when using HBRuntime to run HBIR models.

    Known Issues

    On the S6P platform, the hrt_ucp_monitor tool currently does not support DDR bandwidth monitoring or querying the number of DSPs.