ascend-robotfeat: Add Dockerfile for 950 and update documents

0e6cce0b创建于 17 天前历史提交

Driving SDK Docker Images Overview

Quick Reference

Maintained by: Driving SDK Team
Where to file issues: Driving SDK Issue Tracker
Supported architectures: x86_64 (AMD64), aarch64 (ARM64)
Supported OS: Ubuntu 22.04, openEuler 24.03
Base image: CANN 9.0.0, CANN 8.5.1

Image Tag Keywords Description

The image tags follow this naming convention:

<DrivingSDK_VERSION>-<CANN_VERSION>-<NPU_TYPE>-<OS_TYPE>

Field Descriptions

Field	Description	Available Values
DrivingSDK_VERSION	Driving SDK version	`26.0.0`, `master`
CANN_VERSION	CANN version	`8.5.1`, `9.0.0`
NPU_TYPE	NPU type	`A2`, `A3`, `950`
OS_TYPE	Operating System type	`ubuntu22.04`, `openeuler24.03`

Examples

26.0.0-cann8.5.1-910b-ubuntu22.04: Driving SDK 26.0.0, CANN 8.5.1, A2, Ubuntu 22.04
26.0.0-cann9.0.0-a3-openeuler24.03: Driving SDK 26.0.0, CANN 9.0.0, A3, openEuler 24.03
master-cann9.0.0-950-ubuntu22.04: Driving SDK master, CANN 9.0.0, 950, Ubuntu 22.04

Dockerfile Archive Path

All Dockerfiles are archived under the docker/ directory with the following structure:

docker/
├── 8.5.1-910b-openeuler24.03/
│   └── Dockerfile
├── 8.5.1-910b-ubuntu22.04/
│   └── Dockerfile
├── 8.5.1-a3-openeuler24.03/
│   └── Dockerfile
├── 8.5.1-a3-ubuntu22.04/
│   └── Dockerfile
├── 9.0.0-910b-openeuler24.03/
│   └── Dockerfile
├── 9.0.0-910b-ubuntu22.04/
│   └── Dockerfile
├── 9.0.0-a3-openeuler24.03/
│   └── Dockerfile
├── 9.0.0-a3-ubuntu22.04/
│   └── Dockerfile
├── 9.0.0-950-ubuntu22.04/
│   └── Dockerfile
├── 9.0.0-950-openeuler24.03/
│   └── Dockerfile
├── install_bevformer.sh
├── install_bevfusion.sh
├── install_drivingsdk.sh
├── install_sparse4d.sh
├── OVERVIEW.md
└── OVERVIEW.zh.md

Quick Start

1. Building Images Locally

Build the Docker image from source:

# Clone the repository
git clone https://gitcode.com/Ascend/DrivingSDK.git
cd DrivingSDK

# Build the image
docker build -f docker/8.5.1-910b-ubuntu22.04/Dockerfile -t drivingsdk:26.0.0-cann8.5.1-910b-ubuntu22.04 .

# Run with mounted volumes for development
docker run -it --rm \
  --device=/dev/davinci0 \
  --device=/dev/davinci_manager \
  --device=/dev/devmm_svm \
  --device=/dev/hisi_hdc \
  -v /usr/local/Ascend:/usr/local/Ascend \
  -v $(pwd):/workspace \
  drivingsdk:26.0.0-cann8.5.1-910b-ubuntu22.04 \
  /bin/bash

2. Secondary Development

For custom development and modifications:

# Clone the repository
git clone https://gitcode.com/Ascend/DrivingSDK.git
cd DrivingSDK

# Modify Dockerfile as needed
vim docker/8.5.1-910b-ubuntu22.04/Dockerfile

# Build with your changes
docker build -f docker/8.5.1-910b-ubuntu22.04/Dockerfile -t drivingsdk:dev .

# Run with mounted volumes for development
docker run -it --rm \
  --device=/dev/davinci0 \
  --device=/dev/davinci_manager \
  --device=/dev/devmm_svm \
  --device=/dev/hisi_hdc \
  -v /usr/local/Ascend:/usr/local/Ascend \
  -v $(pwd):/workspace \
  drivingsdk:dev \
  /bin/bash

Hardware Support Information

Supported NPU Types

NPU Type	Architecture	Description	Status
A2	x86_64, aarch64	A2	Production Ready
A3	x86_64, aarch64	A3	Production Ready
950	x86_64, aarch64	950	Preview

Supported Operating Systems

OS	Version	Architecture	Package Manager
Ubuntu	22.04 LTS	x86_64, aarch64	apt
openEuler	24.03 LTS	x86_64, aarch64	yum/dnf

Python Environment Support

The Docker images provide multiple Python environments via Miniconda:

Environment	Python Version	PyTorch Version	Use Case	Status
`torch2.1`	3.8	2.1.0	General PyTorch development	Production Ready
`torch2.7.1`	3.10	2.7.1	Latest PyTorch features	Production Ready
`bevformer`	3.10	2.7.1	BEVFormer model training	Not ready for 950
`bevfusion`	3.10	2.7.1	BEVFusion model training	Not ready for 950
`sparse4d`	3.10	2.7.1	Sparse4D model training	Not ready for 950

Hardware Requirements

Minimum: 1 NPU device (A2 or A3 or 950)
Recommended: 2+ NPU devices for distributed training
Memory: Minimum 32GB RAM, recommended 64GB+
Storage: Minimum 100GB for Docker images and datasets

Included Components

Core Components

DrivingSDK: Driving SDK version 26.0.0 or master
CANN: Compute Architecture for Neural Networks (8.5.1 / 9.0.0)
PyTorch: Deep learning framework (2.1.0 / 2.7.1)
torch-npu: PyTorch NPU backend
Miniconda: Python environment management

Model Examples

BEVFormer: Bird's Eye View Transformer for 3D object detection
BEVFusion: Multi-modal 3D detection with BEV fusion
Sparse4D: Sparse 4D detection framework

System Dependencies

GCC/G++ compiler
CMake build system
Git, wget, curl utilities
Protocol Buffers
Network tools

Usage Examples

Activate a Specific Environment

# Inside the container
source /opt/conda/etc/profile.d/conda.sh
conda activate torch2.7.1

Install Additional Packages

# Activate environment
conda activate torch2.7.1

# Install package
pip install your-package

Troubleshooting

Common Issues

NPU Device Not Found

# Check NPU devices
npu-smi info

# Ensure devices are mounted
docker run --device=/dev/davinci0 ...

CANN Environment Not Set

# Source CANN environment
source /usr/local/Ascend/ascend-toolkit/set_env.sh

Conda Environment Not Found

# Initialize conda
source /opt/conda/etc/profile.d/conda.sh
conda env list

License

This project is licensed under the terms specified in the LICENSE file at the root of the repository.

Third-Party Licenses

CANN: CANN Software License
PyTorch: BSD 3-Clause License
Miniconda: Anaconda Terms of Service

Disclaimer

IMPORTANT: This software is provided "as is" without warranty of any kind, express or implied.

Use at your own risk
The maintainers are not responsible for any damages arising from the use of this software
Always test in a non-production environment before deployment
Ensure compliance with Atlas licensing terms
NPU hardware must be properly configured and accessible

For production deployments, please consult the official Atlas documentation and follow best practices for security and performance optimization.

Support and Resources

CANN Documentation: Ascend Documentation
Issue Tracker: Driving SDK Issues

Last Updated: 2026-05-19 Maintainer: Driving SDK Team