Following the release of GB300 with LongContext blog, here lists the potential items for future optimization on large scale serving with NVL72 systems:

• Enhancement of WideEP: [Feature] Large-EP MoE Redesign #22829
• Upgrade DeepEp branch to HybridEP fix[minimax]: support deepep with minimax models #19468
• (Cancelled due to slow speed) Try disable parallelism for mtp layers to avoid communication Add --speculative-moe-tp-size=1 flag to run Deepseek MTP layer with no parallelism. #16590
• More optimization on Spec v2 (MTP+overlap scheduler)[Feature] Overlap Spec Support #11762
  • Try to remove the data sync at target verify batch
  • Kernel fusions at metadata preparation
• DWDP [Feature] Distributed Weight Data Parallelism (DWDP) for Sparse MoE Models #22084
• Data Parallel Attention load balance [Feature] Load Balance Refactor for DP-Attention #16080
• Enable Dynamo+SGLang with Kubernetes
• Elastic EP/Fault Tolerance https://www.lmsys.org/blog/2026-03-25-eep-partial-failure-tolerance/
• Lower launching time feat: add coordinated checkpoint prefetch for network filesystem loading #20843
• Refining documents and cookbooks
• Support newer models like GLM-5/Kimi-K2/Qwen-3.5...