Getting Started
First steps with DeepSpeed
deepspeed.aiCrawled Nov 18, 2026
First steps with DeepSpeed
The page guides users through training their first model with DeepSpeed.
1. What is DS4Sci_EvoformerAttention DS4Sci_EvoformerAttention is a collection of kernels built to scale the Evoformer computation to larger number of sequences and residuals by reducing the memory footprint and increasing the training speed.
The page explains how to measure model parameters, latency, and floating-point operations.
DeepSpeed is a deep learning optimization library that makes distributed training easy, efficient, and effective.
DeepSpeed introduces **DeepSpeed-FastGen**, a high-throughput text generation system for large language models (LLMs). It integrates **Dynamic SplitFuse**, **DeepSpeed-MII**, and **DeepSpeed-Inference** to optimize performance. The system targets improved efficiency in serving LLM-based applications.
DeepSpeed-VisualChat introduces an open-source framework enabling multi-round conversations with multi-image inputs. The system combines a **2B-parameter visual encoder** with a **70B-parameter LLaMA-2 decoder** to process visual and textual interactions dynamically. Researchers at Microsoft open-source the project to enhance AI-driven visual dialogue applications.
Microsoft’s AI infrastructure team launches the **DeepSpeed4Science** initiative to adapt DeepSpeed’s AI training, inference, and compression tools for scientific research models. The project aims to accelerate large-scale scientific discovery by optimizing AI systems for domain-specific applications. DeepSpeed’s existing technologies—originally developed for scaling deep learning—will now target challenges in fields like biology, chemistry, and physics.
DeepSpeed introduces ZeRO++, a technique that cuts communication overhead by fourfold during large language model (LLM) and chatbot training. The optimization enables faster distributed training workflows while maintaining performance. ZeRO++ targets memory and bandwidth bottlenecks in scaling AI model development.
DeepSpeed introduces ZeRO, a system optimization that allows training of models exceeding 100 billion parameters. The technology reduces memory usage and communication overhead during distributed deep learning. Researchers can now scale model training to unprecedented sizes while maintaining efficiency.