🛰️ Daily AI Frontier
‹ back to 2026-08-17

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

Research Efficiency & Systems

Ranking

Overall 80
Content 95
Popularity 44

Observed public metrics from 1 member.

Merged summary

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.

Sources (1)

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

arXiv cs.LG Hanfeng Lu, Tianyu Feng, Suyi Li, Yuheng Zhao, Wei Gao, Shaopan Xiong, Ju Huang, Siran Yang, Jiamang Wang, Lin Qu, Wei Wang 2026-08-14 arXiv:2608.14498
Public signals Hugging Face upvotes 0
Providers: Hugging Face · Upvotes 0 OpenAlex · N/A Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-15 14:32:41.618004 UTC

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.
item →