Agent SDK

English | 中文

1.Quick Reference

• Where to get help

  • Issue Feedback
  • Aura Code
  • AgentSDK Quick-start Documentation
  • Community

2.Aura Overview

Aura (Agentic Ultra-fast Reinforcement Architecture) is an integrated training-inference-tuning framework for foundation models. It continuously improves foundation models based on task trajectories and reward signals, and through reinforcement learning and other optimization methods, enables the model—via post-training—to progressively acquire agent-like capabilities such as planning, tool use, and long-horizon decision-making.

Through a unified abstraction interface, Aura is compatible with multiple training engines, inference engines, and agent frameworks. It supports flexible integration of custom models and toolchains, helping developers quickly build, train, and deploy their own agents.

3.Supported Tags and Dockerfile Links

3.1 Tag Naming Convention

Tags follow this pattern:

<AgentSDK_version>-<chip_series>-<os>-<python_version>

Field	Example Values	Description
AgentSDK_version	26.1.0	AgentSDK version
chip_series	910, a3, atlas 800	Target chip family
os	ubuntu22.04, openeuler24.03	Base operating system
python_version	py3.11	Python version

3.2 CANN 9.0.0 + 26.1.0 Agent SDK Image

Tag	Dockerfile	镜像内容
26.1.0-910b-ubuntu22.04-py3.11	Dockerfile	toolkit + Agent SDK
26.1.0-a3-ubuntu22.04-py3.11	Dockerfile	toolkit + Agent SDK
26.1.0-910b-openeuler24.03-py3.11	Dockerfile	toolkit + Agent SDK
26.1.0-a3-openeuler24.03-py3.11	Dockerfile	toolkit + Agent SDK

4.Quick Start

4.1 Prerequisites

4.1.1 Install Driver

• An Atlas NPU driver compatible with the container's CANN version must be installed on the host. See the CANN Compatibility Matrix for driver ↔ CANN version mapping.
• Docker version requirement: Docker version should not be lower than 24.0.x.

4.2 Running a Aura Container

Mount Devices Manually

• Device Mounting: Map host device files to the container using the --device parameter to ensure the container can access specified hardware resources. /dev/davinci is the NPU accelerator card (mount as needed), while /dev/davinci_manager, /dev/devmm_svm, and /dev/hisi_hdc are NPU management devices (mount all).

• Driver and Toolchain Mounting: Mount driver files and toolchain directories (such as /usr/local/Ascend/driver and /usr/local/bin/npu-smi) from the host to the container in read-only mode to ensure consistent runtime environment. In the example code below, /dev/davinci1 represents mounting device 1.

docker run --name your_container_name \
    --hostname agent \
    --network host \
    -it -d --shm-size=500g \
    --device=/dev/davinci0 --device=/dev/davinci1 \
    --device=/dev/davinci2 --device=/dev/davinci3 \
    --device=/dev/davinci4 --device=/dev/davinci5 \
    --device=/dev/davinci6 --device=/dev/davinci7 \
    --device=/dev/davinci8 --device=/dev/davinci9 \
    --device=/dev/davinci10 --device=/dev/davinci11 \
    --device=/dev/davinci12 --device=/dev/davinci13 \
    --device=/dev/davinci14 --device=/dev/davinci15 \
    --device=/dev/davinci_manager \
    --device=/dev/hisi_hdc \
    --device=/dev/devmm_svm \
    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver \
    -v /usr/local/dcmi:/usr/local/dcmi \
    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /usr/share/zoneinfo/Asia/Shanghai:/etc/localtime \
    -v /usr/local/sbin:/usr/local/sbin \
    your_image_name:your_image_tag  \
    sleep infinity

4.3 Quick-start Demo

Quick start reference: Qwen3-4B hybrid mode quick start

---

5. Supported Hardware

Chip Series	Product Examples	Architecture
Atlas 910	Atlas 800I A2	ARM64 / x86_64
Atlas A3	Atlas 800I A3	ARM64 / x86_64

---

6. License

View the license information for CANN and MindSeries software included in these images.

As with all container images, the pre-installed packages (Python, system libraries, etc.) may be subject to their own licenses.