| name | ferret-kernel-system |
| description | Use when you need a NEW or optimized MPK CUDA kernel (a per-task `.cuh` under include/mirage/persistent_kernel/tasks/) that must PROVABLY beat a target WITHOUT the kernel-optimizer cheating its own correctness test. This is the frozen-gate Ferret system — a 3-agent orchestration (dispatcher → independent test-writer → optimizer) that makes "the optimizer ships a simplified-math kernel and self-reports a passing test" structurally impossible. Triggers — variants in Chinese ("调 ferret 优化一个 kernel", "让 agent 写个 kernel 还得保证不简化", "用 frozen gate 的方式生成 kernel"): the user/main-thread wants a kernel optimized or generated by an agent, especially when correctness/no-simplification matters more than fire-and-forget speed. |
Ferret frozen-gate kernel system — how to invoke + why
The problem it solves (read first)
The OLD ferret wrote its OWN correctness test. It once shipped a DeepSeek-V3
attention kernel with SIMPLIFIED math (theta-10000 rope not YaRN, 1/sqrt(576) not
the YaRN mscale scale, head-sum o_proj skipping W_UV, no kv_a_layernorm) and
self-reported cosine 1.0 — against its own simplified reference. Marking its
own homework. The refactor takes judging + constraints OUT of the optimizer.
The 3 agents (in .claude/agents/)
| Agent | Role | You invoke? |
|---|
| ferret-kernel-agent (L1 dispatcher) | The ENTRY POINT. Pins ALL constraints, freezes the gate via the test-writer, runs the optimizer in-session, drives Codex review each round, does in-MPK faithful acceptance. | YES — this is the only one you invoke directly. |
| ferret-test-writer (L2a) | Writes the FROZEN, hash-locked gate vs a CANONICAL reference (never re-derived), checking INTERMEDIATE tensors. Spawned by L1, before the optimizer. | No (L1 spawns it) |
| ferret-optimizer (L2b) | In-session optimizer (replaces claude -p), judged ONLY by the frozen gate, can't simplify. Spawned by L1 each round. | No (L1 spawns it) |
How to invoke (the main-thread → subagent contract)
Invoke the ferret-kernel-agent subagent (via the Agent tool) with a COMPLETE
constraint contract — this is where nothing gets missed and no simplification slips
in. Give it, explicitly:
- TARGET — the exact MPK op + what it replaces; the baseline = the kernel being
replaced, benched the way MPK calls it (not an external SOTA unless it's the consumer).
- REAL-MATH CONTRACT — enumerate EVERY step the kernel must compute, NO
simplification (the test-writer turns this into intermediate checks).
- SHAPES — exact (M,K,N) at the real TP/EP regime (derive from the builder/weights).
- PRODUCTION COMPILE FLAGS —
-rdc=true / MPK_FORCE_RDC_TRUE=1, arch sm_100a,
single-stream / no-CUDA-graph / no-cta_group::2. State that FINAL acceptance is the
in-MPK faithful build; standalone -rdc=true is diagnostic only.
- ABI — the
__device__ task_impl signature + NS/NE.
- CANONICAL REFERENCE SOURCE — the already-trusted oracle to compare against
(the in-MPK task-chain output; the official HF model; an in-tree faithful test).
NEVER "let the agent derive it."
If the gate already exists (hash-locked in the workspace), tell L1 to reuse it
(hash-verify) and go straight to the optimizer loop.
The invariants that make it trustworthy (Codex-hardened — gate fidelity is load-bearing)
- The gate is built by the independent test-writer, not the optimizer.
- The reference is CANONICAL (validated against a trusted source), never re-derived.
- The gate checks INTERMEDIATES (golden vectors per stage), not just a final cosine
— a deep simplification (a dropped layernorm, a wrong rope base) is caught at the
first diverging stage (
first_failing_stage), not washed out.
- Multiple metrics + edge cases (long-context, boundary positions).
- The gate is sha256 hash-locked; L1 re-verifies the hash before EVERY round
(tamper = abort).
- The optimizer is judged ONLY by
gate/check.py. Codex reviews each round on two
axes — Integrity (did it simplify to pass?) + Plan (is the lever sound?).
- FINAL acceptance = in-MPK faithful build (
MPK_FORCE_RDC_TRUE=1, compiled into
the real megakernel). A standalone number never ships.
- Early stop = round incomplete, never success.
Proof it works
Re-run on the simplified attention: the optimizer (judged by the frozen gate) built a
CORRECT real-DSv3 fused attention in round 1 — GATE_RESULT {pass:true} on all 5
cases / all intermediates, gate untouched, dump provenance verified (W_UV BMM present,
real mscale²/sqrt(192) scale), all 4 simplifications fixed, Codex Integrity+Plan PASS.
The old simplified-kernel failure is now structurally impossible.
Reference
Full design + the /cd-mechanics + the hardening rationale: DESIGN.md (bundled
alongside this skill). The ferret runtime itself lives at
~/ferret/ (its CLAUDE.md is the optimizer's methodology). Pre-authored task specs:
~/ferret/tasks/.
