PyTorch integrates SGLang into IBM Spyre plugin

With assistance from AI coding agents, a two-person team integrated the SGLang serving stack into IBM's Spyre enterprise AI accelerator in under a week. The resulting roughly 1,500-line plugin delivered about 3.6 times the throughput of the reference PyTorch implementation while performing on par with vLLM on measured workloads for Granite 3.3 8B and Gemma 4 26B A4B models. Prefix sharing through RadixCache worked without requiring any Spyre-specific modifications. Open article →