TLDR Tech / Practical
VibeSys Optimizes Qwen3.5-397B on MI300A, Outperforming SGLang
The SyFi Lab at the University of Washington utilized its VibeSys system to construct and enhance a serving engine for the Qwen3.5-397B-A17B model across four AMD MI300A units. The automated process, spanning over 105 agent hours, delivered 2,242 tokens per second of goodput for multi-turn conversations, a 2.33-fold increase compared to a modified SGLang setup that could not operate directly on the hardware.
Read the full story