April 2026: Delivered a performance-focused feature in yhyang201/sglang by implementing XPU-specific support to apply router weights on inputs for Llama4 fused-experts. This change modulates inputs with router weights to improve inference throughput and efficiency on XPU backends. Commits: 1aee04e7df7da92b1ae5b89398755a24dab287ff.

March 2026 monthly summary for sgl-project/sglang. Focused on stabilizing floating-point precision checks in Triton attention unit tests. Delivered a safer bf16 precision validation by adjusting the assertion tolerance, improving test accuracy and reducing false negatives. The change is tracked under commit ac2819c81fcd15ac272dba8873784be21f34da46 (Fix assertion tolerance for bf16 precision in triton attention UT), with sign-off by Rahul Vijayaraghavan. This work enhances test reliability for the Triton attention path and contributes to the overall correctness and reliability of the sgLang project.