mindspore-ai/mindspore
Mindspore
MindSpore es un framework de aprendizaje profundo diseñado para construir y entrenar redes neuronales en entornos de nube, borde y móviles. Funciona como un sistema de entrenamiento distribuido y un kit de herramientas de IA acelerado por hardware, capaz de ejecutar cargas de trabajo en CPUs, GPUs y procesadores de IA especializados.
El proyecto incluye un motor de diferenciación automática que calcula gradientes mediante la transformación de código fuente y la compilación estática. Permite el entrenamiento de modelos distribuidos dividiendo las cargas de trabajo entre el hardware utilizando paralelismo de datos y de modelos.
El framework cubre el despliegue de IA multiplataforma y la inferencia de modelos, utilizando un kit de herramientas de alto rendimiento para acelerar la ejecución y el servicio. Proporciona aceleración especializada para hardware Ascend y admite el mapeo de operadores agnóstico al hardware para backends de dispositivos heterogéneos.
El entorno se puede instalar mediante un gestor de paquetes o compilar desde el código fuente en sistemas Linux para entornos de procesadores de IA estándar o especializados.
Features
- Deep Learning Frameworks - Provides a complete deep learning framework for building and training neural networks across cloud, edge, and mobile environments.
- AI Model Inference - Provides capabilities for deploying trained neural networks to production environments with optimized execution and serving.
- Automatic Differentiation Engines - Implements an engine that computes gradients of neural networks through source transformation and static compilation.
- Data-Parallel Training - Provides built-in support for distributing training workloads across hardware by splitting data batches and synchronizing parameters.
- Gradient Computation - Automatically calculates gradients for neural networks using source transformation to enable training and optimization.
- Source-to-Source Differentiation - Computes neural network derivatives by transforming source code to support complex control flows and static compilation.
- Hardware Acceleration Toolkits - Ships a suite of tools and kernels to optimize deep learning model execution across CPUs, GPUs, and specialized AI processors.
- Large-Scale Model Training - Distributes large-scale training workloads across multiple devices using data and model parallelism.
- Distributed Training - Provides a platform to scale the training of deep learning models across multiple compute nodes using data and model parallelism.
- GPU-Accelerated Training - Leverages GPU hardware acceleration to perform deep learning training and inference tasks.
- Static Graph Compilations - Converts dynamic model definitions into static execution graphs to optimize performance during training and inference.
- Ascend Processor Optimizations - Provides specialized acceleration and optimizations specifically for Ascend AI processors.
- Model Inference Accelerators - Ships a high-performance toolkit to accelerate the execution and serving of deep learning models.
- Cross-Platform Deployments - Enables development of deep learning models that can be ported between CPU, GPU, and edge-computing hardware.
- High-Performance Inference Modes - Includes a toolkit that optimizes model execution via graph fusion and quantization to reduce inference latency.
- Operation-to-Accelerator Mapping - Implements logic to map high-level mathematical operators to specific CPU, GPU, or AI processor backends for optimized execution.
- Heterogeneous Device Abstractions - Uses a unified interface to manage memory and execution across heterogeneous hardware including Ascend, GPU, and CPU.
- Deep Learning Frameworks - Huawei-developed framework for distributed AI development.
- Deep Learning Implementations - Deep learning framework developed by Huawei.
