Description

This project implements Christopher Owen's 'spark-ds41f' recipe to run the DeepSeek-V4.1-Flash model across four DGX Spark (GB10) boxes using a tensor parallel ring configuration. It achieves decode speeds of ~76-81 tok/s and over 8,300 tok/s prefill at 32K context, utilizing vLLM with custom TileLang kernels and a one-shot ring all-reduce transport.

Descriptions, tags, and model credits may be AI-generated or inferred from public sources and can be incomplete or wrong. Learn more.