mx-driving 26.1.0rc1.post1

mx_driving v26.1.0rc1 for PyTorch 2.1.x on Atlas A2 (Ascend NPU acceleration for autonomous driving)

mx_driving

Overview

mx_driving is an operator and model acceleration library for autonomous driving, embodied intelligence VLA, and world models, developed on the Ascend NPU platform. It provides a suite of high-performance operators and model migration examples that support the PyTorch framework, enabling developers to leverage the powerful compute capabilities of Ascend AI Processors for autonomous driving workloads.

Ascend is a full-stack AI computing infrastructure for industry applications and services based on Huawei Ascend processors and software. For more information about Ascend, see Ascend Community.

Installation

From Binary

1. Install PyTorch and torch-npu

Before installing mx_driving, complete the installation of PyTorch and torch-npu. For torch-npu installation, refer to torch-npu on PyPI.

# Install PyTorch
pip install torch==2.1.0

# Install torch-npu
pip install torch-npu==2.1.0

2. Install mx_driving

mx_driving provides two PyPI packages based on hardware platform:

PyPI Package	Hardware	Install Command
mx-driving	Atlas A2 (default)	pip install mx-driving
mx-driving-a5	Atlas A5	pip install mx-driving-a5

Both packages provide the same import mx_driving interface. Choose based on your hardware:

# For Atlas A2 (default, most users)
pip install mx-driving

# For Atlas A5
pip install mx-driving-a5

If you need a specific PyTorch version:

# Atlas A2 + PyTorch 2.1.x (this version)
pip install mx-driving==26.1.0rc1.post1

# Atlas A2 + PyTorch 2.6.x/2.7.x/2.8.x
pip install mx-driving==26.1.0rc1.post2

# Atlas A5 + PyTorch 2.6.x/2.7.x/2.8.x
pip install mx-driving-a5==26.1.0rc1.post2

Note: mx-driving and mx-driving-a5 cannot be installed simultaneously since both provide the mx_driving Python package.

From Source

For source compilation, refer to Deploying Driving SDK Environment.

Getting Started

Prerequisites

Initialize the CANN environment variable:

source /usr/local/Ascend/ascend-toolkit/set_env.sh

Quick Verification

You can quickly verify the installation with a simple example using the scatter_max operator:

import torch
import torch_npu
from mx_driving import scatter_max

updates = torch.tensor([[2, 0, 1, 3, 1, 0, 0, 4],
                        [0, 2, 1, 3, 0, 3, 4, 2],
                        [1, 2, 3, 4, 4, 3, 2, 1]], dtype=torch.float32).npu()
indices = torch.tensor([0, 2, 0], dtype=torch.int32).npu()
out = updates.new_zeros((3, 8))
out, argmax = scatter_max(updates, indices, out)
print(out)
print(argmax)

One-Click Patcher

mx_driving provides a one-click Patcher mechanism that automatically replaces GPU-based implementations with NPU-optimized implementations, enabling seamless model migration from GPU to Ascend NPU:

from mx_driving import patcher
patcher.patch_all()

For more details, refer to One-Click Patcher.

High-Performance API

mx_driving provides various categories of high-performance operators optimized for Ascend NPU:

Category	Operators
General	scatter_max, scatter_mean, scatter_add, knn, furthest_point_sampling, group_points, unique_voxel, graph_softmax, ...
Sampling	bev_pool_v1, bev_pool_v2, bev_pool_v3, npu_voxel_pooling_train, roiaware_pool3d, border_align, ...
Voxelization	voxelization, dynamic_scatter
Detection	nms3d, boxes_iou_bev, box_iou_quadri, diff_iou_rotated_2d, points_in_boxes_all, ...
Sparse	SparseConv3d, SubMConv3d, SparseInverseConv3d
Fusion	multi_scale_deformable_attn, deformable_aggregation, npu_add_relu, npu_fused_bias_leaky_relu, npu_batch_matmul, ...

For the complete API list, refer to API Reference.

Supported Models

mx_driving provides migration and optimization examples for autonomous driving models on Ascend servers, including perception, planning, end-to-end, and VLA models:

• BEVFusion — Multi-task multi-sensor fusion framework
• BEVFormer — Bird's-eye-view perception from camera images
• Sparse4D — Sparse 4D perception framework
• And more...

For the full model list, refer to Model Support List.

PyTorch and Python Version Compatibility

mx-driving (Atlas A2)

Version	CPU Architecture	Python Version	PyTorch Version	torch_npu Version
26.1.0rc1.post1	x86 & aarch64	3.8, 3.9, 3.10, 3.11	2.1.0	v2.1.0-26.1.0rc1
26.1.0rc1.post2	x86 & aarch64	3.9, 3.10, 3.11	2.6.0/2.7.1/2.8.0	v2.7.1-26.1.0rc1

mx-driving-a5 (Atlas A5)

Version	CPU Architecture	Python Version	PyTorch Version	torch_npu Version
26.1.0rc1.post2	x86 & aarch64	3.9, 3.10, 3.11	2.6.0/2.7.1/2.8.0	v2.7.1-26.1.0rc1

Hardware Support

Product Series	Product Model
Atlas A2 Training Series	Atlas 800T A2, Atlas 900 A2 PoD, Atlas 200T A2 Box16, Atlas 300T A2
Atlas A3 Training Series	Atlas 800T A3, Atlas 900 A3 SuperPoD

Branch Maintenance Policies

Status	Duration	Description
Planning	1-3 months	Plan features
Development	3 months	Develop features
Maintained	6-12 months	Incorporate resolved issues and release versions. Regular versions: 6 months; Long-term support versions: 12 months
Unmaintained	0-3 months	Incorporate resolved issues, no dedicated maintainers, no releases
End Of Life (EOL)	N/A	No longer accept any modifications

Version Maintenance

Version	Maintenance Policy	Status	Release Date	Subsequent Status	EOL Date
v26.1.0rc1	Release Candidate	Maintained	2026/05/30	Expected unmaintained from 2026/11/30
v26.0.0	Regular Release	Maintained	2026/03/30	Expected unmaintained from 2026/09/30
v7.3.0	Regular Release	Maintained	2025/12/30	Expected unmaintained from 2026/06/30
v7.2.RC1	Regular Release	Maintained	2025/09/30	Unmaintained from 2026/03/30
v7.1.RC1	Regular Release	Unmaintained	2025/06/30	Unmaintained from 2025/12/30

Suggestions and Communication

If you have any questions or suggestions, you can submit Issues. We will reply as soon as possible.

License

• Driving SDK is licensed under the Apache-2.0 License.
• Documentation under the docs directory is licensed under CC-BY 4.0.

Contributing

Before contributing, please sign the Open Source Contributor License Agreement (CLA).

If you encounter a bug, please submit an issue.

If you plan to contribute bug fixes, please submit Pull Requests. See contribution guidelines for details.

If you plan to contribute new features, please create an issue for discussion first.