amznlabs/amazon-dsstneArchived
Amazon Dsstne
Amazon DSSTNE 是一个机器学习工具包和稀疏张量网络库，专为具有稀疏输入和输出的深度学习模型而设计。它提供了一个模型并行训练框架和一个 GPU 加速的稀疏引擎，以支持内存密集型网络。
该框架专门为推荐系统训练和大规模稀疏学习而设计。它实现了将大型权重矩阵和嵌入表分布在多个 GPU 设备上，以处理超过单个处理器内存容量的模型。
该项目涵盖了广泛的能力，包括分布式 GPU 计算、稀疏数据集处理以及可扩展稀疏张量网络的构建。这些实用程序允许在 GPU 集群上执行高性能机器学习操作和模型扩展。
Features
- Model Parallelism - Provides a model-parallel training framework to split large weight matrices across multiple GPUs.
- Distributed GPU Computing - Distributes training and prediction tasks across multiple GPUs to increase processing speed and memory capacity.
- Distributed GPU Training - Enables distributing neural network training and prediction workloads across multiple GPUs to increase memory and speed.
- Embedding Table Sharding - Supports partitioning massive embedding tables across distributed graphics processors to handle memory-intensive models.
- Recommendation Model Training - Designed for training deep learning recommendation models with weight matrices that exceed single-GPU memory.
- Model Training Frameworks - Provides a distributed training framework that scales large weight matrices across multiple GPUs for memory-intensive networks.
- Recommendation Models - Serves as a toolkit specifically for building deep learning recommendation models with sparse inputs and outputs.
- Sparse Computing Kernels - Implements specialized computational kernels to accelerate sparse neural network operations on GPU hardware.
- Sparse Tensor Network Libraries - Provides a library for building and training machine learning models using scalable sparse tensor networks.
- Sparse Computation Engines - Provides a GPU-accelerated sparse engine using custom kernels to process sparse datasets without zero-value expansion.
- Distributed Tensor Networks - Provides a distributed tensor network to organize large-scale model computations as a graph spread across a GPU cluster.
- Sparse Learning - Supports large-scale sparse learning by processing vast datasets with mostly zero values efficiently.
- Large-Scale Model Training - Supports building large-scale recommendation models using model-parallel scaling for weight matrices.
- Sparse Tensor Network Modeling - Enables constructing machine learning models using scalable sparse tensor networks to handle large-scale data.
- Sparse-to-Dense Embedding Tables - Provides embedding tables that map high-dimensional sparse inputs directly to dense vectors without matrix expansion.
- Sparse Data Processing - Executes computations on sparse datasets using custom GPU kernels to maintain performance without expanding zero entries.
- Deep Learning Frameworks - Library for building deep learning recommendation systems.
- General Machine Learning - Library for training and deploying deep neural networks at scale.
Star 历史
