Announcement
SGLang reportedly merges Megakernel support
A post claims throughput close to 1,000 tokens per second per user on the GB300 NVL72 AgentX harness.
TLDR
SemiAnalysis says SGLang has merged Megakernel support, claiming close to 1,000 tokens per second per user on the GB300 NVL72 AgentX harness and a 1.5x increase in interactivity on MI355X. The post says the support already works out of the box in SGLang Docker images and expresses hope of bringing it to official AgentX.
Combined views
—
3 Sources, first seen ago
