For the most part each provider implemented their own inference engines. This release will help them all adopt the common set of optimisations to serve DeepSeek Flash optimally.

Interesting, thanks!