vLLM on GB200 DEP32 tracks DeepSeek's decode curve

DeepSeek-V4.1-Flash decode at 131K context, DSpark with 3 draft tokens at a fixed 3.25 accepted tokens per step. Up and to the right is better.

vLLM, GB200 DEP32 DeepInfer, EP64DP64

vLLM: full decode step on 32 GB200 GPUs (384 + 32 experts), real weights and routing, 131K context, EPLB on. DeepInfer: read off DeepSeek's chart (EP64DP64, 384 + 64 experts). Throughput = batch × tokens/s per user.