LMSYS optimizes DeepSeek-V4.1-Flash on H20-3e
Researchers optimized disaggregated DeepSeek-V4.1-Flash serving on H20 GPUs by tuning prefill and decode parallelism, incorporating FP8 DeepGEMM, breakable CUDA graphs, tensor-parallel MoE, and optimized indexer kernels. Compared with the baseline, the final configuration reduced mean end-to-end latency by 65.5% and increased output throughput by 2.91 times. Open article →