TorchSpec Enables Disaggregated Speculative Decoding Training at Scale

zhyncs42 · x · 2026-07-30

The TorchSpec team published an article on their open-source project for large-scale disaggregated speculative decoding training. As frontier models like Kimi K2.5 and GLM 5 scale to hundreds of billions of parameters, efficient inference has become a critical systems challenge.

The article highlights that training draft models (e.g., EAGLE-3) requires transferring massive volumes of hidden states from the target model (a single 128K-token sample requires 7GB). Traditional methods of precomputing and storing these on disk create severe I/O bottlenecks. TorchSpec proposed and implemented a disaggregated training architecture to solve this. The project has been adopted by companies like DigitalOcean and CoreWeave, and Tencent Hunyuan has even extended it with a new algorithm called AngelSpec.

Original post →

More from Infra

Infra channel →