February 2026 highlights: Delivered Sampling API Enhancements to support seed and offset as scalar or 1D tensor inputs, enabling per-call seeds/offsets and better CUDA graph compatibility. Fixed CUDA Graph integration issues in the sampling path and centralized input validation to enforce correct dtype, device, shape/length, and batch semantics. Updated documentation and usage guidance, including union-type signatures and CUDA-graph notes. Added/updated tests with all tests passing, reinforcing robustness.

December 2025 monthly summary for bytedance-iaas/sglang. Focused on delivering Auxiliary Hidden State support in Eagle v2 to enhance model performance and inference flexibility. This feature enables capturing auxiliary hidden states during inference, aligning with Eagle v2 roadmap and expanding use cases for sgLang in production environments.

November 2025 performance summary focused on delivering scalable multimodal capabilities in IBM/vllm. Implemented M-RoPE support for the Eagle model to enhance multimodal input handling, with dynamic argument dimensions for improved tensor operations and better Torch compilation compatibility. Added CUDA graph support through MRope integration to optimize performance and stability during inference. These changes align with our roadmap for robust, production-ready multimodal models and position the repository for higher throughput workloads.

June 2025 monthly summary for NVIDIA/TensorRT-LLM focusing on enabling secure CI access for IzzyPutterman and aligning CI workflow with contributor permissions to improve feedback loops and PR velocity.