TorchSpec: Speculative Decoding Training at Scale

Introduction Over the past year, large language models have rapidly expanded in both scale and capability. Frontier models such as Kimi K2.5, GLM 5, and Qwen 3.5 now operate with…