.. _release_note_2.1:

============================
Palette SDK version 2.1
============================

This document outlines the new and regression features, and known issues in SiMa.ai’s Palette software version v2.1.0, along with their respective workarounds. It also includes information on the Palette software packages, supported systems, and compatible software. Palette v2.1.0 is released on April 16th, 2026.

New Features and Enhancements
-----------------------------
.. list-table::
   :widths: 20 20 40
   :header-rows: 1

   * - **Feature Category**
     - **Availability**
     - **Feature Description**
   * - Platform
     - GA
     - Yocto support for Modalix.
   * - MLA
     - GA
     - 10X and Non-10X models; multiple inputs; Video/Image, Audio, and Text input types; sequential and async multi-model execution.
   * - MLA
     - GA
     - GStreamer, C++ API, and Python API support; Generic Pre-Proc and GenericBoxDecoder; BF16 and multi-pipeline support.
   * - MLA (GenAI)
     - GA
     - LFM2 and Qwen VL runtime support.
   * - SoC Features
     - GA
     - Linux Kernel 6.18; Ubuntu 24.04 for SDK and PCIe host driver; 16 GB SOM support.
   * - Video Platform
     - GA
     - MIPI camera support: IMX568 (FPGA), IMX477, IMX678, AR234, OG05C10.
   * - Video Platform
     - GA
     - USB Camera and OpenPose pipelines.
   * - Video Platform (GenAI)
     - EA
     - GenAI_SoCPipeline, GenAI_VisionPipeline, YoloBasedGenAI, GenAI_Audio_Pipeline.
   * - Video Platform
     - GA
     - MJPEG 420 encoding.
   * - Platform
     - GA
     - Kontron PCIe driver; ROS2 .deb for eLxr; ConnectTech 4-MIPI pipeline; Leopard Carrier Card.
   * - MLA
     - GA
     - Automatic Graph Surgery.
   * - MLA (GenAI)
     - EA
     - 16 GB and 8 GB SoM support; per-token quantization; GGUF Q4_1/Q4_K; LLM-Compressor; Dynamic LoRA; MoLE Evaluation Framework.

Known Issues
------------
.. list-table::
   :widths: 20 30 40
   :header-rows: 1

   * - **Issue**
     - **Impact**
     - **Workaround**
   * - QAT is not supported in SDK 2.1 due to Python 3.10 → 3.12 upgrade (``PT2EQ`` accuracy regression).
     - QAT is not functional in SDK 2.1 ModelSDK.
     - Run QAT in SDK 2.0 (Python 3.10) to produce the annotated ONNX, then pass it to SDK 2.1 for compilation.