February 2026 monthly work summary for kvcache-ai/sglang. Key features delivered include Qwen3.5 model support with multimodal processing and architecture refinements; configurable Mamba state dtype via configuration files; block-wise FP8 quantization and model adaptation for large-scale models; and a distributed-precision bug fix for the Qwen3.5 dense model when TP_SIZE > 1, applying all-reduce fusion in the MLP. These changes improve scalability, efficiency, and hardware adaptability, enabling production-ready multimodal inference and larger-scale deployments. Commits highlighted: 27c447653d9cf0f63aea1c190b931be4875cbf86, 4ed2548427a0f01a969d6e518088bcb62a568f5d, 44603764d65e79d2406eab8d1928dfdec9290138, fa5698d7916497288af8fe5a5b57bc4ee7e6fb37, d38c0e537d95bfb78486c1185f68c90046ce0cc9.

January 2026 performance summary for kvcache-ai/sglang (2026-01). Delivered a targeted data-parallel size handling fix for the Qwen3 Vision Model, introducing an attention_reduction parameter and refactoring multiple modules to adopt it. This work resolves the dp size > 1 issue, stabilizing distributed training and improving throughput on multi-GPU runs. The change reduces training downtime and accelerates experimentation with vision-language workloads. The work was completed with a cross-module refactor and co-authored commit.

September 2025 monthly summary focusing on key accomplishments, business value, and technical achievements for sgLang (yhyang201/sglang).