AIRLinf adds Cosmos3 support for fine-tuning and robot evaluation
LMSYS reports 3.33× end-to-end evaluation throughput with SGLang, which batches inference for 128 parallel environments on eight GPUs.
TLDR
LMSYS announced that RLinf, an open-source framework for embodied intelligence and AI agents, now supports Cosmos3 from fine-tuning to robot evaluation. It reports 3.33× end-to-end evaluation throughput with SGLang, describing a setup covering 500 episodes of the full LIBERO-10 evaluation with 128 parallel environments and eight GPUs. LMSYS says RLinf overlaps CPU simulation with GPU inference to reduce waiting between stages.
Combined views
1.1K
2 Sources, first seen 6h ago
RLinf adds Cosmos3 support for fine-tuning and robot evaluation
LMSYS reports 3.33× end-to-end evaluation throughput with SGLang, which batches inference for 128 parallel environments on eight GPUs.