Metaseq is a flexible, high-performance framework for training and serving large-scale sequence models, such as language models, translation systems, and instruction-tuned LLMs. Built on top of PyTorch, it provides distributed training, model sharding, mixed-precision computation, and memory-efficient checkpointing to support models with hundreds of billions of parameters. The framework was used internally at Meta to train models like OPT (Open Pre-trained Transformer) and serves as a reference implementation for scaling transformer architectures efficiently across GPUs and nodes. It supports both pretraining and fine-tuning workflows with data pipelines for text, multilingual corpora, and custom tokenization schemes. Metaseq also includes APIs for evaluation, generation, and model serving, enabling seamless transitions from training to inference.
Features
- Distributed training and inference for large-scale transformer models
- Support for model, data, and pipeline parallelism across multiple GPUs and nodes
- Mixed-precision training and memory-efficient checkpointing
- Pretraining and fine-tuning workflows for text and multilingual data
- APIs for text generation, evaluation, and serving large models
- Reference implementation for Meta’s OPT and other large language models