shenh10/DeepSeek_Simulator — measured-kernel DeepSeek V3/R1 decode throughput on H800/H20

Test-based simulator: REAL DeepGEMM + FlashMLA kernels run on the target GPU (H800 80G, H20 96G explicitly 'tested'), then decode-phase throughput is computed for data-parallel/expert-parallel configs that cannot fit on one node. Values are tokens/sec PER DEVICE aggregate decode at large batch (b_mla 32-256, d=EP degree 24-288). It is a measured-kernel model, not an end-to-end serving trace - treat as a simulated ceiling, status=draft.