ResearchBenchmarksLLM Decoding on Llama-3-8BFollow0.5172Decode Time per StepPre30.500580.6127650.724950.837135Jun 4, 2025Evaluation ResultsMethodMethodLinksDecode Time per StepPre3Batchsize=1Batchsize=12025.060.5172XGrammarBatchsize=1Batchsize=12025.060.5531Pre3Batchsize=4Batchsize=42025.060.6537XGrammarBatchsize=4Batchsize=42025.060.9327