haipipe-nn-tuner

name	haipipe-nn-tuner
description	Layer 2 (Tuner) specialist of haipipe-nn: defines the hyperparameter search space and the tuner that drives sweeps. Called by /haipipe-nn; direct invocation works layer-scoped.
argument-hint	[function] [args...]
allowed-tools	Bash, Read, Write, Edit, Grep, Glob
metadata	{"version":"0.1.1","last_updated":"2026-07-04","summary":"Layer 2 (Tuner) specialist of haipipe-nn."}

Skill: haipipe-nn-tuner

Layer 2 specialist. Owns the tuner classes in code/hainn/tuner/ — search space declarations, hyperparameter sampling, and the loop that wraps L1 algorithms. Called by the /haipipe-nn orchestrator; can also be invoked directly.

Function axis: dashboard | review | generate | test

---

Commands

/haipipe-nn-tuner                      -> dashboard: registered tuners
/haipipe-nn-tuner dashboard            -> same
/haipipe-nn-tuner review [class_name]  -> structural review of a tuner class
/haipipe-nn-tuner generate             -> scaffold a new tuner / search space
/haipipe-nn-tuner test [class_name]    -> run a short tuner smoke sweep

---

Dispatch Table

Invocation    This skill's ref       Umbrella's fn doc
------------- ---------------------- ------------------------------
dashboard     ref/concepts.md        ../haipipe-nn/fn/fn-dashboard.md
review        ref/concepts.md        ../haipipe-nn/fn/fn-review.md
generate      ref/concepts.md +
              ../haipipe-nn-instance/
                ref/concepts.md      ../haipipe-nn/fn/fn-generate.md
test          ref/concepts.md        ../haipipe-nn/fn/fn-test.md
(no fn arg)   ref/concepts.md        (ref-only mode)

generate reads the L3 (instance) ref because the tuner's output (best-config + ckpt) becomes the input contract for ModelInstance.

---

Step-by-Step Protocol

Step 0: Read ../haipipe-nn/ref/overview.md. Mandatory. Step 1: Parse args. Function vocabulary: dashboard | review | generate | test. Step 2: Read this skill's ref/concepts.md for L2 specifics. Step 3: Read the umbrella fn doc. Step 4: For generate, also read ../haipipe-nn-instance/ref/concepts.md. Step 5: Execute, scoped to L2 (tuner only — no algorithm class edits). Step 6: Emit the structured tail.

---

Layer Scope

Owns:

• Tuner classes in code/hainn/tuner/
• Search-space declarations (param names, ranges, samplers)
• The sweep loop that orchestrates L1 algorithm trials

Upstream dependency (L1): Tuner search space must match the algorithm's __init__ signature. If a param is missing or extra, escalate to /haipipe-nn-algo review.

Hand-off contract (L2 -> L3): ModelInstance creates this Tuner via the model_tuner registry and calls fit() / save_model(key, model_dir); the Tuner owns training and serialization inside the instance's model_dir. Verify against ../haipipe-nn-instance/ref/concepts.md.