Compiler-First State Space Duality and Portable $O(1)$ Autoregressive Caching for Inference
Researchers have developed a new method for optimizing Mamba-2 inference, focusing on compiler-first state space duality. This approach enables portable autoregressive caching with $O(1)$ complexity, eliminating the need for custom CUDA or Triton kernels. The resulting single-source inference path, implemented in JAX, demonstrates significant speedups on Google Cloud TPUs and NVIDIA GPUs, achieving high hardware utilization and matching reference perplexity scores. AI
IMPACT Enables faster and more portable inference for large state space models, potentially reducing deployment costs and complexity.