Custom fused sampler kernel boosts vLLM throughput by 15%

Developer mmastrac built a custom fused sampler kernel for DiffusionGemma on vLLM, boosting standard text generation throughput by roughly 15%.

2026-09-23 ~ 2026-09-23 · 2 related posts

• Custom fused sampler kernel adds 15% tok/s throughput in vLLM for DiffusionGemma — generativist · 2026-09-23
• Custom fused sampler kernel adds 15% tok/s for DiffusionGemma on vLLM — imjustnewatai · 2026-09-23