Qualcomm NPU (AI Engine Direct) with LiteRT

LiteRT supports Qualcomm AI Engine Direct (QNN) through the Compiled Model API for both AOT and JIT. This page consolidates the vendor guide previously under eap/.

Set up development environment

Qualcomm AI Engine Direct has the following development environment requirements:

  • Operating System: Ubuntu 22.04 LTS
  • Build System: Bazel version 7.4.1
  • Standard libraries: Libc++ and Libc++abi 16+
  • Platform Specific Requirements:
    • Android:
    • Android SDK: Android API Level 34 (Android 14).
    • Android NDK: Support for API Level 28 (Android 9 Pie).
    • IoT: coming soon.
    • Windows: coming soon.

You can use the Quick system setup with Docker to configure dependencies automatically.

Supported operations

LiteRT lowers a broad set of ops to the QNN Hexagon Tensor Processor (HTP) backend. Both AOT and on-device compilation are supported. See the supported ops list.

Supported SoCs

  • Snapdragon 8 Elite Gen 5 Mobile Platform (SM8850)
  • Snapdragon 8 Elite Mobile Platform (SM8750)
  • Snapdragon 8 Gen 3 Mobile Platform (SM8650)
  • Snapdragon 8 Gen 2 Mobile Platform (SM8550)
  • Snapdragon 8+ Gen 1 Mobile Platform (SM8475)
  • Snapdragon 8 Gen 1 Mobile Platform (SM8450)

Next steps

  1. Start with the unified NPU guide: NPU acceleration with LiteRT
  2. Follow the conversion and deployment steps there, choosing Qualcomm where applicable.
  3. For LLMs, see Execute LLMs on NPU using LiteRT-LM.