Daily AI Updates: December 7, 2025

Main Content

• SGLang's Miles Training Framework: Adds FSDP backend support to improve LLM training efficiency, achieving numerical alignment with Megatron and compatibility with context parallelism.

• SGLang-Diffusion: Integrates the Cache-DiT caching framework, delivering an inference speedup of 20% to 165%, with a 46% acceleration in specific configurations.

• Gemini Enterprise: Introduces an internal codebase search agent feature, enabling users to locate precise functions and documentation with simple queries, reducing the need for code rewriting.