index.md

layout	permalink	hidden	header	excerpt	feature_row
splash	/	true	overlay_image overlay_filter /assets/images/def_banner.png rgba(255, 255, 255, 0.8)	Automatically Discovering Fast Parallelization Strategies for Distributed Deep Neural Network Training<br />	title excerpt url btn_class btn_label Joint Optimization FlexFlow uses a novel hierarchical search algorithm to jointly optimize algebraic transformations and parallelization while maintaining scalability. https://www.cs.cmu.edu/~zhihaoj2/papers/unity_osdi22.pdf btn--primary Learn more title excerpt url btn_class btn_label Flexible Parallelization FlexFlow supports parallelizing DNN training through combinations of the Sample, Operator, Attribute, and Parameter dimensions. https://cs.stanford.edu/~zhihao/papers/sysml19a.pdf btn--primary Learn more title excerpt url btn_class btn_label Speculative Inference FlexFlow accelerates generative LLM inference with speculative inference and token tree verification. /specInfer/ btn--primary Learn more

FlexFlow is a DNN framework that automatically discovers fast parallelization strategies for distributed DNN training. FlexFlow generalizes and goes beyond today's manually designed parallelization strategies (e.g., data and model parallelism) for distributed DNN training by exploring parallelization opportunities across different Samples, Operators, Attributes, and Parameters.

FlexFlow includes a novel execution simulator to evaluate the runtime performance of different strategies and uses an automated search algorithm to discover highly optimized strategies, which generally outperform today's manually designed strategies.

{% include feature_row %}