amk-optimize

Drive an interactive AutoMegaKernel (AMK) schedule-optimization session on $ARGUMENTS (parse as model [gpu]; default gpu = rtx5090; default model = toy). Use the megakernel-optimization skill's loop and obey its HARD HONESTY RULES.

HARD HONESTY RULES (state them, obey them):

• Correctness FIRST: NEVER report a latency without a correctness PASS vs the CPU ReferenceVM. Keep a candidate only if it is correct AND >= 1% faster than the incumbent.
• validate-before-launch: an unsafe ScheduleConfig is a clean REJECTED, never a hung GPU.
• Edit surface is ScheduleConfig + kernel_knobs ONLY, never kernel code, never vm/ or the ABI.
• Measured-gpu latency is drift-robust; impossible sub-roofline latencies are withheld.
• Speedups are vs AMK's OWN default schedule, NOT a claim of beating cuBLAS/vLLM.

Do this:

1. Read the edit surface with amk_propose(model, gpu) (CLI fallback: amk propose <model> --gpu <gpu>). Report the incumbent schedule_config and the editable search_space (including the kernel_knobs.* sub-surface).
2. Establish the baseline with amk_eval(model, gpu, <incumbent cfg>, device="auto") (CLI: write cfg.json, amk eval <model> --gpu <gpu> --config cfg.json). Require valid AND correct. Record its latency_us as the incumbent latency.
3. Loop: propose ONE knob change off the current best (one ScheduleConfig knob OR one kernel_knobs field per trial), amk_eval it, and keep/revert, keep ONLY if valid AND correct AND latency_us < incumbent_latency_us * 0.99. After each trial call amk_orchestrate_record(status, latency_us=..., pct_roofline=..., kind=..., config=..., description=...) with status in kept/revert/failed/crash/timeout/rejected (revert is the canonical token for any correct-but-not-kept candidate).
4. Between trials, consult amk_orchestrate_next() (continue-or-STOP) and amk_orchestrate_status() (baseline/best/speedup/plateau). Stop when next says STOP or the user asks. To run the full keep/revert loop in one shot instead, use amk_loop(model, gpu, budget=N) (CLI: amk loop <model> --gpu <gpu> --budget N).

Finish with a short report: best schedule_id, best latency_us + latency_kind, pct_of_roofline, and the speedup vs the incumbent baseline (clearly stated as vs AMK's own default, not vs cuBLAS/vLLM).

Use only the canonical names: MCP amk_propose/amk_eval/amk_loop/amk_orchestrate_record/ amk_orchestrate_next/amk_orchestrate_status; CLI amk propose|eval|loop and python amk_orchestrate.py record|next|status.