March 2026 summary for yhyang201/sglang: Focused on robustness of distributed attention under varying data/tensor parallel configurations. Delivered a targeted fix that prevents dp_attention crashes when dp_size < tp_size during warmup, by aligning token counts to the tensor parallel size and adjusting batch sizing to maintain valid shapes. No new user-facing features released this month; emphasis was reliability, correctness, and maintainability of distributed execution. Result: fewer runtime crashes, smoother scale-out, and faster troubleshooting in multi-device setups.

January 2026 performance summary for kvcache-ai/sglang: Delivered substantial feature and reliability improvements across multimodal capabilities, reasoning parsing, and code quality. Implemented consolidated multimodal processing enhancements for K2-VL / KimiK25 (new image processor, vision tower integration, and video-chunk support with updated model configurations); extended OpenAI serving with a new reasoning parser 'kimi_k2' to determine if requests require reasoning; fixed a dimensionality issue in MoonViT3dPretrainedModel by squeezing the hidden states; completed code cleanup and logging improvements (refactored logging, removed debug prints, cleaned unused functions, updated tokeniser request logging). These changes improved inference versatility and reliability, reduced debugging time, and improved maintainability, enabling broader adoption of multimodal capabilities and faster iteration cycles for future features.