AutoMegaKernel: Command for Claude Code

One-shot compile and verify a model into a single persistent CUDA megakernel with AutoMegaKernel (AMK) on $ARGUMENTS (parse as model [gpu]; default gpu = rtx5090; default model = toy). This is THE product path: import -> lower -> validate (deadlock + race free) -> verify vs eager -> build the GPU megakernel -> measure against the HBM roofline -> emit a correct megakernel

• report.

HARD HONESTY RULES (state them, obey them):

• Correctness FIRST: the compile path verifies vs eager PyTorch AND the CPU ReferenceVM; a latency is NEVER reported without a correctness PASS.
• validate-before-launch: an unsafe schedule is a clean REJECTED (deadlock/race-free proof), never a hung GPU.
• The edit surface (if you tune afterward) is ScheduleConfig + kernel_knobs ONLY, never kernel code, never vm/ or the frozen ABI.
• Measured-gpu latency is drift-robust; impossible sub-roofline latencies are withheld.
• Any speedup is vs AMK's OWN baseline, NOT a claim of beating cuBLAS/vLLM.

Do this:

1. Confirm the environment with amk_doctor() (CLI: amk doctor), torch/cuda, device name, nvcc, registered targets. A CUDA GPU + nvcc is required to build and measure the megakernel.
2. Compile and verify in one shot via the CLI: amk compile <model> --gpu <gpu> --regime single-stream (equivalently uv run python amk_cli.py compile <model> --gpu <gpu> --regime single-stream). This lowers, statically validates, verifies correctness vs eager, builds the GPU megakernel, and measures it against the weights / HBM_bandwidth roofline, emitting the megakernel + report.
3. Report the result: correctness PASS (vs eager + ReferenceVM), measured per-token latency_us
   • latency_kind, pct_of_roofline, and the schedule_id. If correctness did not PASS, report the failure honestly and do NOT report a latency.

Optional next step: to tune the compiled schedule, run /amk-optimize <model> <gpu> (interactive propose -> eval -> keep/revert) or /amk-autoresearch <model> <gpu> (unattended). Speedups there are vs AMK's own default schedule, not vs cuBLAS/vLLM.

Use only the canonical names: CLI amk compile and amk doctor (MCP amk_doctor for the environment check).