MosaicML Composer Training
A large-scale training stack using Composer and the MosaicML toolkit to train and fine-tune models, including LLMs, efficiently on GPU clusters. It cuts training time and cost through proven optimizations.
MosaicML Composer Training
This stack focuses on efficient large-scale model training using Composer, MosaicML's open-source PyTorch training library, alongside the broader MosaicML toolkit for language and diffusion models. It is built to train and fine-tune deep learning models, including LLMs, faster and at lower cost through algorithmic and systems optimizations on GPU clusters.
Components
- Composer: A PyTorch training library with a flexible trainer, a library of speedup methods, automatic mixed precision, and efficient distributed training (FSDP). Its callback and algorithm system applies optimizations with minimal code change.
- MosaicML LLM Foundry / Streaming: Recipes and tooling for training LLMs, plus the Streaming dataset library that loads large datasets efficiently from object storage.
- GPU clusters on Kubernetes: Multi-node, multi-GPU infrastructure runs distributed training jobs.
- Amazon S3: Stores training datasets and checkpoints, streamed during training.
Strengths
- Training efficiency. Built-in speedup methods and FSDP cut training time and cost.
- Scalability. Efficient multi-node distributed training scales to large models.
- Streaming data. The Streaming library loads massive datasets from object storage without local copies.
- Practical recipes. LLM Foundry provides tested configurations for pretraining and fine-tuning.
Trade-offs
- GPU cost and ops. Large training runs require expensive, carefully managed GPU clusters.
- Expertise required. Distributed training, sharding, and tuning are advanced topics.
- PyTorch-centric. The toolkit targets PyTorch workflows.
- Not for inference. Serving belongs to a separate stack.
When to Use It
Choose this stack when you train or fine-tune deep learning models, especially LLMs, at scale and want to reduce training time and cost through proven optimizations. It fits teams with GPU clusters building or adapting foundation models. For inference-only needs or small models, simpler training fits. For efficient large-scale model training, Composer and the MosaicML toolkit are a strong foundation.