Collected molecules will appear here. Add from search or explore.
Distributed training and inference optimization library that enables scaling deep learning models to trillions of parameters through memory-saving techniques like ZeRO and various forms of parallelism.
stars
42,020
forks
4,776
DeepSpeed is a category-defining project in the LLM era. Its ZeRO (Zero Redundancy Optimizer) paper revolutionized how researchers handle memory in distributed training. With over 42k stars and deep integration into the Hugging Face ecosystem, it serves as the industry-standard infrastructure for training large-scale models. While PyTorch FSDP provides competition, DeepSpeed remains ahead in specialized kernels and heterogeneous memory management. Frontier labs are partners or heavy contributors rather than competitors in this specific software niche.
TECH STACK
INTEGRATION
pip_installable
READINESS