Skip to main content

Soki Fukuda · createcentury

Low-level optimization for AI on Apple Silicon. Currently writing Metal Shading Language kernels for state space models.

  • mamba-metal — Mamba's selective scan in MSL. Full feature parity with the official CUDA kernel; 11× faster prefill than mlx-lm at 5,601-token prompts.
  • mamba-metal-swift — Swift port of the kernels, library + iOS demo. Real Mamba inference on iPhone GPU.
  • mamba-vs-transformer-edge — On-device benchmarks: Mamba vs Transformer with detailed memory profiling.

Writing

  • Blog — English technical posts (Mamba, JAX, kernels, …)
  • Personal — Personal notes (日本語)

Elsewhere