Index

optimus_dl

Optimus-DL: A modular, high-performance framework for training Large Language Models.

Optimus-DL is a research framework built on PyTorch that provides:

• Modular "Recipe" architecture for clean separation of concerns
• Hydra-based configuration management
• Universal metrics system with distributed aggregation
• Modern PyTorch features (AMP, FSDP2, Tensor Parallelism, torch.compile)
• Efficient kernels via Liger-Kernel integration
• Registry system for easy component swapping

Example

Basic training:

from optimus_dl.core.registry import build
from optimus_dl.recipe.train.config import TrainConfig

config = TrainConfig(...)
recipe = build("train_recipe", config)
recipe.run()

Modules and Sub-packages

• core:
• modules:
• recipe: