# LiteRT 架构 LiteRT 框架会针对延迟、模型大小和功耗对模型进行优化。LiteRT 可帮助您在具有低功耗要求的设备上运行模型，例如移动、嵌入式和边缘平台。LiteRT 框架在 delegate 的帮助下运行模型。Delegate 是在特定硬件上使用库以高效运行神经网络模型的软件层。 下图展示了 LiteRT 框架用于运行模型的 delegate： **图：LiteRT 架构** ## LiteRT 设备端推理概述 LiteRT 设备推理将模型加载到解析器中，解析器解析模型并使用 delegate 来运行它。 该过程包括以下内容： 1. 推理将 LiteRT 模型加载到 LiteRT 解释器接口中，该接口解析模型以识别其中存在的神经网络运算符。 2. 对解析器接口进一步配置，以使用 delegate 来运行模型。 3. 解析器根据提供的输入调用模型推理，并将相应输出保存到为解析器接口提供的缓存中。 Qualcomm 支持使用 delegate 在以下加速器上运行 LiteRT 模型： - CPU - Adreno GPU - Hexagon 张量处理器 下表列出了 delegate 及其加速器： 表：支持的 delegate 和加速器 | Delegate | 加速 | | --- | --- | | XNNPACK delegate | CPU | | GPU delegate | GPU | | Qualcomm^®^ AI Engine direct delegate (Qualcomm^®^ Neural Network (QNN) delegate) | CPU、GPU 和 Hexagon 张量处理器 | ## 使用 delegate 加速 LiteRT 模型 您可以使用 delegate 将 LiteRT 图执行分流到硬件加速器，例如 CPU、GPU 和 Hexagon Tensor Processor。LiteRT 目前支持以下 delegate： - [XNNPACK delegate 用于 CPU](https://docs.qualcomm.com/doc/80-70022-54SC/topic/arch.html#xnnpack-delegate) - [GPU delegate](https://docs.qualcomm.com/doc/80-70022-54SC/topic/arch.html#gpu-delegate) - [QNN delegate](https://docs.qualcomm.com/doc/80-70022-54SC/topic/arch.html#qnn-delegate) ### 使用 XNNPACK delegate 加速 CPU 上的模型 XNNPACK delegate 使用 XNNPACK 库在 CPU 上高效加速 LiteRT 模型。XNNPACK 是 Google 的开源库，它的作用如下： - 提供可在 Arm CPU 上运行且优化的神经网络算子 - 使用底层 CPU 指令（如 Arm^®^ Neon^™^ 指令集）来优化算子，以实现高效执行 XNNPACK delegate 可以运行 32 位浮点和 int8 格式的模型。有关详细信息，请参见 [TensorFlow Lite 的 XNNPACK 后端](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/delegates/xnnpack/README.md)。 要使用 XNNPACK delegate 运行 LiteRT 模型，请参阅[使用可用 delegate 运行 LiteRT 模型](https://docs.qualcomm.com/doc/80-70022-54SC/topic/sample-applications.html#label-image-tool)。 ### 使用 GPU delegate 加速 GPU 上的模型 GPU 开源 delegate 可在各种特定于供应商的 GPU（包括 Adreno GPU）上为 LiteRT 模型提供加速。 LiteRT 可以使用 GPU delegate 来提高 GPU 的并行处理能力，从而加快推理速度。GPU delegate 使用 OpenCL 内核在 GPU 上的 LiteRT 模型执行图中运行神经网络操作。 GPU delegate 的默认交叉编译包含 LiteRT 库，可优化以下 LiteRT 模型在 Adreno GPU 上的执行： - 16 位浮点 - 32 位浮点 关于更多信息，请参阅 [GPU delegates for LiteRT](https://www.tensorflow.org/lite/performance/gpu)。 要使用 GPU delegate 运行 LiteRT 模型，请参阅[使用可用 delegate 运行 LiteRT 模型](https://docs.qualcomm.com/doc/80-70022-54SC/topic/sample-applications.html#label-image-tool)。 ### 使用 QNN delegate 加速特定硬件上的模型 QNN delegate 是专为特定于供应商的硬件加速而设计的专有 delegate，以加速 LiteRT 模型。它基于 LiteRT 的[外部 delegate 接口](https://ai.google.dev/edge/litert/performance/implementing_delegate#option_2_leverage_external_delegate)。可以使用 QNN delegate 将部分或整个 LiteRT 模型分配到专用的 Qualcomm 硬件，例如 Adreno GPU 和 Hexagon Tensor Processor。 delegat 通过减少 CPU 工作负载来提高模型执行性能和功效。它还使用现有的 Qualcomm AI Engine Direct API 和可用的后端来加速模型。有关这些 API 的更多信息，请参见 [Qualcomm AI Engine direct SDK](bundle/publicresource/topics/80-63442-50)。QNN delegate 在可用的硬件上以 32 位浮点精度和 int8 精度运行模型。 可使用以下接口编译应用程序： - Qualcomm AI Engine Direct delegate 接口 - LiteRT 外部 delegate 接口 使用独立的 LiteRT 应用程序时，您可以访问这两个界面。但是，如果您使用 Qualcomm IM SDK 部署 LiteRT 模型，则 Qualcomm TensorFlow Lite 的 qtimltflite GStreamer 插件将使用 QNN delegate。有关详细信息，请参见[利用外部 delegate](https://ai.google.dev/edge/litert/performance/implementing_delegate#option_2_leverage_external_delegate)。 下图显示了 Qualcomm AI Engine direct SDK 中的 QNN delegate 库的目录结构： ![../../_images/qnn-delegate-directory-structure.png](data:image/png;base64,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) **图：QNN delegate 目录结构** 要使用 QNN delegate 运行模型，请参阅[使用外部 delegate 运行 QNN delegate](https://docs.qualcomm.com/doc/80-70022-54SC/topic/sample-applications.html#run-qnn-delegate-using-the-external-delegate-interface)。 ## Qualcomm AI Engine Direct delegate 接口 Qualcomm AI Engine direct delegate 接口（也称为 QNN delegate）提供了 `QnnTFLiteDelegate.h` 头文件作为接口。 Tab Qualcomm Linux Tab Ubuntu - *class* tabincludedirective - 您可以在链接到 QNN delegate 库之前将此标头包含在您的应用程序中。兼容的 QNN delegate 库和 QNN 库位于 `aarch64-oe-linux-gcc11.2 cross-compiler` 工具链三元组目录中。 下表列出了 delegate 的后端及其各自的库： 表：QNN delegate 加速支持 | 后端名称 | 后端说明 | 目标和库名称 | 库描述 | | --- | --- | --- | --- | | CPU | Arm CPU 加速的后端 | | `libQnnCpu.so`：CPU 后端库 | | GPU | Adreno GPU 硬件加速器的后端 | | `libQnnGpu.so`：GPU 后端库 | | Hexagon 张量处理器 | Hexagon 张量处理器的硬件加速器的后端 | | | 确保根据 Qualcomm 开发套件的 Hexagon Tensor Processor 版本选择适当的库。版本如下： - Qualcomm Dragonwing™ RB3 Gen 2 开发套件：Hexagon Tensor Processor v68 - Dragonwing IQ-9075：Hexagon Tensor Processor v73 - IQ-8 Beta 评估套件：Hexagon Tensor Processor v75 - Dragonwing IQ-615：Hexagon Tensor Processor v66 - *class* tabincludedirective - 您可以在链接到 QNN delegate 库之前将此标头包含在您的应用程序中。兼容的 QNN delegate 库和 QNN 库位于 `aarch64-ubuntu-gcc9.4` 交叉编译器工具链三元组目录中。 下表列出了 delegate 的后端及其各自的库： 表：QNN delegate 加速支持 | 后端名称 | 后端说明 | 目标和库名称 | 库描述 | | --- | --- | --- | --- | | CPU | Arm CPU 加速的后端 | | `libQnnCpu.so`：CPU 后端库 | | GPU | Adreno GPU 硬件加速器的后端 | | `libQnnGpu.so`：GPU 后端库 | | Hexagon 张量处理器 | Hexagon 张量处理器的硬件加速器的后端 | | | 确保根据 Qualcomm 开发套件的 Hexagon Tensor Processor 版本选择适当的库。版本如下： - Qualcomm Dragonwing™ RB3 Gen 2 开发套件：Hexagon Tensor Processor v68 - Dragonwing IQ-9075 评估套件：Hexagon Tensor Processor v73 ### 使用外部 delegate 接口运行 LiteRT 模型 要使用外部 delegate 接口运行 LiteRT 模型，应用程序必须加载 `libQnnTFLiteDelegate.so` QNN delegate 库。C/C++ 应用程序和 `libQnnTFLiteDelegate.so` delegate 库彼此之间没有依赖项关系。因此，如果 delegate 发生变化，不必重新编译应用。 要使用 C++ API 并在 Qualcomm 开发套件上使用 LiteRT 运行推理，请参阅 [Run inference using C++](https://ai.google.dev/edge/litert/inference#run-c)。 QNN delegate 在 Hexagon 张量处理器、GPU 和 CPU 上提供加速。要在使用外部 delegate 接口时通过 QNN delegate 定制模型运行的场景和方式，必须提供外部 delegate 选项。相关说明，请参阅 [QNN delegate 的外部 delegate 选项](https://docs.qualcomm.com/doc/80-70022-54SC/topic/sample-applications.html#external-delegate-options-for-qnn-delegate)。 ## 后续步骤 - [开始运行 LiteRT 模型](https://docs.qualcomm.com/doc/80-70022-54SC/topic/getting-started.html#getting-started) - [部署 LiteRT 模型](https://docs.qualcomm.com/doc/80-70022-54SC/topic/tensorflow-lite-developer-workflow.html#tensorflow-lite-developer-workflow) - [运行 LiteRT 示例应用程序](https://docs.qualcomm.com/doc/80-70022-54SC/topic/sample-applications.html#run-litert-sample-apps) Last Published: Nov 05, 2025 [Previous Topic 开始运行 LiteRT 模型](https://docs.qualcomm.com/bundle/publicresource/80-70022-54SC/topics/getting-started.md) [Next Topic 部署 LiteRT 模型](https://docs.qualcomm.com/bundle/publicresource/80-70022-54SC/topics/tensorflow-lite-developer-workflow.md)