Changelog

Changelog#

Vitis AI 6.2 Release Notes#

The following changes and enhancements have been introduced in Vitis AI 6.2 relative to the 6.1 release.

Board Setup and Quick Start#

  • QSPI Board Setup: Added EOU (Ease of Use) scripts to automate and simplify board bring-up, replacing manual setup steps.

  • Quick Start Demo: ResNet-50 and YOLOx models are now included directly in the root filesystem (Rootfs), enabling out-of-the-box inference demos without additional model downloads.

Quantization#

  • Quark Options: The following options are now mandatory when configuring Quark for INT8 quantization:

    quant_config.extra_options["DedicatedQDQPair"] = True
quant_config.extra_options["QuantizeAllOpTypes"] = True

    These options ensure that Quantize/Dequantize (Q/DQ) nodes are generated for all INT8 operations, which is required for correct quantization graph representation and inference execution.

API and Configuration Changes#

  • dp_size (replaces device_batch_size): Specifies the Data Parallelism size. Users must update their configurations to use dp_size in place of the deprecated device_batch_size parameter.

  • tp_size (replaces optimize-target=latency): Specifies the Tensor Parallelism size. Users must update their configurations to use tp_size in place of the deprecated optimize-target=latency parameter.

    Important

    In vitisai_config.json, the placeholder device name ve2 must be replaced with the actual target device name. The ve2 placeholder will continue to be supported on an interim basis.

Mixed Precision#

  • FP16 Support Added: Mixed Precision Islands now support FP16 data type throughout the model, not only at the head or tail layers. This enables more flexible precision assignment across the full model graph.

Device Support#

  • New Devices Supported: Additional target devices are now supported in Vitis AI 6.2. Refer to Model Compilation for the full list of supported device names.

Contents