Understand this story
On November 12, 2026, the PyTorch Blog reported that a two-person IBM team, with the help of an AI coding agent, integrated SGLang into the IBM Spyre accelerator in under a week using a plugin of about 1,500 lines of code. The plugin has been open-sourced along with upstream SGLang, and RadixCache prefix sharing works without any Spyre-specific changes. Measurements show that SGLang on Spyre delivers about 3.6 times the throughput of the torch-spyre reference PyTorch inference implementation, and matches vLLM on Granite 3.3 8B and Gemma 4 26B A4B. There are currently no further updates.
AI-generated from public reports
Latest development · Report
IBM team integrates SGLang into Spyre in under a week with AI coding agentsDevelopments
1 confirmed developments
- Source:PyTorch Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 12
PyTorch BlogFirst-party
IBM team integrates SGLang into Spyre in under a week with AI coding agents
A two-person IBM team integrated SGLang into the IBM Spyre accelerator in under a week using a plugin of about 1,500 lines of code, with the help of AI coding agents. SGLang on Spyre delivers about 3.6x the throughput of the torch-spyre reference PyTorch inference implementation, and in measured tests on Granite 3.3 8B and Gemma 4 26B A4B it matches vLLM. The plugin has been open-sourced with upstream SGLang, and RadixCache prefix sharing works without any Spyre-specific changes.