March 2026 (sglang) focused on stability, correctness, and performance in core attention components. Key deliverable: a targeted fix to Rotary Embedding Handling in DeepseekV2AttentionMLA that eliminates the naive rotary forward overriding, reducing risk of incorrect rotary behavior and enhancing model performance under rotary embeddings. This change improves reliability for attention-heavy workloads and prepares the model for future scaling.

October 2025 monthly summary for JustinTong0323/sglang focused on feature delivery and code quality. Delivered a Tokenizer Batch Decoding Control feature, enabling granular decoding control via a new CLI flag and updating DetokenizerManager to switch to individual decoding when enabled. This work reduces the risk of performance regressions and behavior inconsistencies across workloads, while improving configurability and traceability for future enhancements. The change is tracked by commit 138ff23187a8c75f68ecc7afddf33f2d3ee494d4 and references issue #11944.

Month: 2025-09 | Repository: kvcache-ai/sglang Overview: Focused on reliability and compatibility improvement for MoE weight loading on NVFP4 target models when using Flashinfer. No new feature deliveries this month beyond critical stability fixes.