Hybrid-state caching, prefix reuse, Mamba checkpointing, GPU-kernel selection, and speculative decoding are jointly tuned
1 more from this model
We additionally optimize hybrid-state caching and GPU kernels for efficient long-context and agentic serving