CANN Commercial Edition

Provides the procedure for quickly installing CANN. For details, see the quick installation FAQ.

Provides a comprehensive guide to installing, upgrading, and uninstalling CANN across different operating systems and use cases.

Develops operators based on the Ascend C operator programming language. For details about the APIs, see Ascend C APIs.

Describes how to develop AI CPU and AIV communication operators based on communication operator APIs.

Provides C, C++, and Python APIs for building AI applications. These APIs help you use the AI processor's compute for tasks like neural network inference, image processing, and scientific computing.

Provides GE APIs for graph construction, modification, compilation, execution, and other tasks. For details about the APIs, see GE APIs and Basic Data Structures and APIs.

Provides basic and advanced Ascend C APIs.

Provides unified APIs for applications and frameworks to efficiently utilize AI processor compute.

Constructs graphs that run directly on the Ascend platform through GE APIs.

Provides a wide array of high-performance operators with deep optimization and hardware affinity.

A high-performance collective communication library based on Ascend AI Processors, enabling data parallelism and model parallelism in single-server multi-device and multi-server multi-device modes.

Guides developers how to use unilateral communication library APIs for inter-cluster data transmission and build a disaggregated LLM inference framework.

Describes how to use the Ascend Transformer Boost acceleration library to improve the efficiency of Transformer model training and inference development.

Describes how to use high-performance signal processing operators.

DVPP is an image processing unit in the AI Processor that speeds up media processing with APIs. It handles VPC image processing, JPEG encoding and decoding, and video encoding and decoding.

Provides tool guides for PyTorch training development, model inference development, and operator development.

Describes how to use operator development tools (such as msKPP, msOpGen, msOpST, msSanitizer, msDebug, and msProf).

Builds operators to generate operator binary files.

Converts a network model into an offline model (.om) supported by Ascend AI processors.

Automates optimization to fully use hardware resources and improve network performance.

Migrates the PyTorch training script to Ascend NPUs in one click.

Compares precision and pinpoints model issues.

Collects and analyzes profile data in training and inference.

Tests HCCL function correctness and performance.

Compresses a model, including quantization and tensor decomposition.

Locates memory problems during model training and inference.

Describes how to use the BiSheng compiler to build operator code into binary executable files and dynamic libraries.

Describes the intermediate representation (IR) built based on MLIR, designed for building Ascend-optimized operators.

Describes how to locate and rectify faults.

Describes the log content format, how to view logs, and how to set log levels.

Describes the environment variables for building AI applications and services based on CANN.

Provides built-in graph fusion and UB fusion patterns of Ascend AI processors. Graph fusion and UB fusion are key methods for improving the performance of the entire network.

Describes the basic data structures and APIs on which operator development and graph development depend.

Provides the CANN communication matrix, open-source software notice, and vulnerability fixes list.

Offers hardware-accelerated APIs for short feature retrieval. These APIs let you add and remove feature libraries, change and delete feature vectors, and perform searches.

Describes the RDMA-based network detection technology RPing for sending detection packets, recording network latency, and collecting statistics on packet sending and receiving.

Describes how to develop custom TBE and AI CPU operators using corresponding APIs.

Disaggregates model inference using LLM DataDist APIs to improve inference throughput.

Constructs, modifies, compiles, and executes computational graphs through DataFlow C++ and Python APIs, and provides UDF APIs for you to write custom processing functions through FuncProcessPoint and GraphProcessPoint.

Describes how to tune algorithms and functions related to the image signal processing (ISP).