Determinism-Preserving GPU Spatial Sharing with Vitamin-E

Abstract

GPU sharing faces a determinism--utilization tradeoff: fixed bindings can strand capacity as demand fluctuates, while resource-driven kernel reshaping improves utilization by altering a launch's parallel structure, potentially changing output bits. We rethink modern GPU scheduling and observe that it decouples logical structure from physical width: one unmodified launch spans a family of widths through changes in block placement and wave count. From this observation, we derive the parallel-structure invariant: for fixed-structure deterministic workloads, keeping each launch immutable makes its output bits independent of physical width. Guided by this invariant, Vitamin-E late-binds immutable launches to pooled physical contexts, preserving bitwise equality across allocations, whereas resource-driven reshaping can alter the selected token under temperature-zero greedy decoding. Across all workload--baseline comparisons, Vitamin-E achieves up to 3.50$\times$ the aggregate normalized LLM training throughput, 62.5\% lower inference p99 latency, and 1.43$\times$ the background-training throughput. With the same mechanism, \textsc{TPOT-First} reduces TPOT SLO violations by up to 46.1\% over \textsc{Throughput-Oriented} on three serving workloads, demonstrating mechanism effectiveness and policy flexibility.