The vLLM community has optimized serving for Qwen3.5's hybrid attention architecture (combining full-attention and GDN layers), achieving over 25K TPS per GPU on NVIDIA GB200 NVL72 systems. By addressing heterogeneous state transfer and accelerating GDN computation on Blackwell GPUs, this solution significantly boosts the performance ceiling of disaggregated serving.
Termsslug: gated-delta-network
Gated Delta Network
门控增量网络
A linear attention architecture variant combining gating mechanisms with delta updates.
Curated Intelligence
vLLM 官方博客T1·Tools
68 pts
Qwen 官方博客 (通义)T1·Infra
78 pts
The Qwen team has open-sourced FlashQLA, a high-performance linear attention library built on TileLang and optimized for Gated Delta Networks (GDN). By leveraging gate-driven intra-GPU sequence parallelism and hardware-friendly algebraic rewrites, it achieves 2-3x forward and 2x backward speedups on NVIDIA Hopper GPUs, significantly reducing training and inference costs for long-context models.