Rollplex: Cross-Phase GPU Spatial Sharing for Vision Language Model Post-Training

TL;DR - Rollplex accelerates synchronous reinforcement-learning post-training for vision-language models by overlapping prefix computation with rollout decoding. It improves GPU utilization without changing on-policy update semantics.

• Decomposes reference-scoring and training phases, moving response-independent video and prompt-prefix processing into underutilized rollout windows.
• Uses phase-aware HBM management to control tensor residency according to producer-consumer lifetimes.
• Shares compatible actor-weight storage across different tensor-parallel layouts, reconstructing only incompatible tensors.
• On 32 H800 GPUs, achieves 1.23×–1.30× speedups over serial colocation and 1.57×–2.24× over disaggregation at the same GPU budget.