TEMPO: Makespan-Aware Expert-Parallel Load Balancing Across Memory- and Compute-Bound Regimes

arXiv:2608.13057v2 Announce Type: replace-cross
Abstract: In expert-parallel (EP) MoE serving, every layer synchronizes at the slowest GPU. Dispatchers balance token counts (EPLB, LPLB, UltraEP) or activated-expert counts (METRO), assuming expert time is linear in one. Measurements on two datacenter GPU generations show it is neither: below $n^* approx 156$–$168$ tokens, HBM weight streaming dominates—cost attaches to $activated replicas$, not tokens; above it, grouped GEMM rounds tokens to 128-tile $M$-tiles, so $splitting$ an expert adds padded compute. A max-affine profile $t=max(a+bG,,c+beta N)$ captures both regimes. Realistic decode batches hold hot experts in the linear regime and cold in the flat $simultaneously$; recorded batches show proxy dispatches differ by $1.4$–$1.6times$ in modeled block time (p95 up to $1.7times$), and $which$ proxy wins flips with the regime. We formalize per-batch dispatch as a fixed-charge makespan problem—NP-hard on two fully replicated GPUs, polynomial in degenerate limits—and present TEMPO, a makespan-aware dispatcher solving it in milliseconds off the critical path; its SGLang integration runs out-of-process and fuses dispatch with count collection into one in-graph kernel. Anchored by an 8-GPU Testbed A microbenchmark, TEMPO stays within $1%$ of the best fixed baseline everywhere and wins by up to $15.5%$ where regimes mix. End-to-end on Testbed B, Qwen3-235B (inside the win region) gains $4$–$6%$ throughput and cuts p99 latency by $sim 15.6%$; DeepSeek-V3 (outside, communication-dominated) shows only mechanism cost. A phase diagram, not a universal win, is the claim: it predicts both outcomes before deployment.