June 2026 monthly summary for vllm-ascend (vllm-project/vllm-ascend). This month focused on delivering Ascend-optimized weight transfer across distributed RL training and inference, establishing zero-copy IPC pathways, and providing end-to-end RLHF demonstrations to validate performance and reliability. The work enhances cross-node productivity and throughput for RL workloads on Ascend NPU.

March 2026 monthly summary for volcengine/verl. Key accomplishments include delivering Mooncake backend for the checkpoint engine with performance benchmarks and a dedicated Mooncake Transfer Engine-based weight synchronization class, and fixing test initialization by updating CheckpointEngineManager to use a config argument instead of backend to support both kimi and mooncake backends. Documentation and benchmarks were updated to reflect new capabilities and performance profiles. These changes improve readiness for Mooncake workloads, reinforce cross-backend compatibility, and provide clear performance visibility for stakeholders.