How-to guides¶
Task-oriented recipes. Each assumes you have completed the getting-started tutorial.
- Choose a mixer — pick
mixer=from your state shape and per-token budget; CPU-only, no GPU needed. - Control scan dispatch — force the eager or Triton backend via
MINGRU_SCAN; a CUDA GPU is needed only to exercise the Triton path itself. - Enable time decay — turn on
decay=/delta_tfor a mixer or stack, including the delta-mixer's eager-only path; CPU-only. - Run the benchmarks — reproduce the Triton scan speedup/memory artifacts; needs a CUDA GPU.
- Reproduce the evidence — replay a recorded accuracy row under the frozen
torch==2.5.1pin from a repo checkout; CPU-only.