deepseek-ai/3FS
3FS
3FS is a distributed file system and RDMA storage cluster designed for high-performance AI training and inference workloads. It functions as a strongly consistent storage layer that utilizes a disaggregated architecture to pool SSDs and memory resources across multiple nodes.
The system provides specialized storage implementations including an AI training checkpoint store for parallel state preservation and a distributed key-value cache store for decoder layer vectors to optimize inference processing. It ensures data integrity through chain replication and apportioned query distribution.
The project covers broader capabilities for training dataset streaming with random access, high-throughput parallel checkpointing, and the organization of large-scale data pipelines for analytics and preparation. Data movement is handled via remote direct memory access to minimize CPU involvement.
Features
- High-Performance File Systems - Provides a high-performance distributed file system utilizing SSDs and RDMA for AI training and inference.
- Distributed Storage Clusters - Implements a disaggregated distributed storage cluster using RDMA and SSDs to provide high-throughput data transfers for AI workloads.
- Distributed Model Checkpointing - Provides high-throughput parallel checkpointing to preserve the state of large-scale training jobs across distributed compute resources.
- Parallel Checkpoint Streams - Writes large model states across multiple storage nodes simultaneously to minimize training downtime during synchronization.
- Chain Replication - Ensures strong data integrity across distributed nodes using a chain replication sequence for all write operations.
- Strongly Consistent Data Stores - Combines chain replication and apportioned queries to provide a strongly consistent distributed storage layer.
- Distributed Storage Engines - Manages the low-level persistence and replication of data across a cluster of nodes optimized for AI workloads.
- Disaggregated Storage Pooling - Deploys a disaggregated distributed file system using SSDs and RDMA to share storage resources across multiple nodes.
- Storage-Compute Architectures - Utilizes a disaggregated architecture that separates compute resources from storage hardware for independent scaling.
- RDMA Networking - Uses remote direct memory access to move data between nodes with minimal CPU involvement for low-latency storage access.
- Distributed File Systems - Implements a distributed file system that spreads data across multiple nodes for high-performance AI workloads.
- Distributed Training - Provides the storage infrastructure to stream training datasets across compute nodes with random access.
- Inference Acceleration - Caches key and value vectors for decoder layers to reduce redundant computations and increase processing capacity.
- Random-Access Dataset Streaming - Allows compute nodes to pull specific training samples on demand without requiring pre-shuffled files or manual prefetching.
- LLM KV Cache Stores - Provides a distributed storage solution for caching decoder layer key and value vectors to optimize inference.
- Training Sample Streaming - Streams training samples across compute nodes using random access to eliminate the need for manual prefetching or shuffling.
- Replica Load Balancers - Implements mechanisms to distribute read requests across multiple replicas to prevent bottlenecks and increase retrieval throughput.
- Inference State Caching - Caches decoder layer key and value vectors to eliminate redundant computations and increase processing capacity during AI inference.
- AI & Machine Learning - Distributed file system for AI training.
- Attention Optimization - High-performance file system for distributed model training and inference.
- Data Infrastructure - Parallel file system optimized for high-bandwidth SSD and RDMA network access.
- Training and Orchestration - High-performance file system for AI training and inference.
Star history
How this analysis was created: This summary and feature list were written by an AI model that read the project's README and public documentation pages. Each feature links to the documentation it came from; stars, license and language come straight from the GitHub API. The model does not read the source code, and the analysis is refreshed when the project is re-analysed. Learn more on our About page.
