| name | hipfire-tester |
| description | Guide a tester through hipfire bring-up, serve smoke, claim-scoped harnesses, and benchmark reporting on AMD RDNA/CDNA GPUs. Use when the user wants a route-oriented test plan or an upstream-ready tester report. Retired batteries are not current evidence — use [`docs/VALIDATION.md`](../../../docs/VALIDATION.md). |
hipfire-tester
Route-oriented bring-up, serve, and reporting skill. Pick the narrowest
validation route for the claim under test; do not invent a universal matrix.
Load guide.md when running a full session or drafting a report.
Authority
Fail closed: unknown claim class → no improvised gate. Green no-GPU CI is
not GPU correctness. Registry presence is not runtime admission
(docs/admissions.yml records stay empty until
earned).
What this skill is not
- Not a universal GPU replacement gate.
- Not numerical/state parity (use an arch-owned oracle when one exists; else blocked).
- Not Redline product timed-arm proof from harness success alone.
- Not promotion authority for benches, Atlas rows, or historical checkpoints.
Quick start (bring-up only)
hipfire diag
hipfire pull qwen3.5:4b
hipfire run qwen3.5:4b "Explain WMMA in one paragraph."
If diag or first run fails, chain to hipfire-diag / hipfire-autoheal.
Claim → harness (summary)
Full map: docs/VALIDATION.md. Branch by
claim class immediately after identity/build — do not run a fixed linear
battery. Maintained manual harnesses that exist today:
| Claim class | Route |
|---|
New/changed numeric .hip | target/release/examples/test_kernels then the applicable model/path-level manual route for that arch/surface (docs/VALIDATION.md). Channel alone is not the complete route. |
| Forward / fusion / KV numerical or state | Path-specific parity/state oracle for that arch/surface; BLOCKED if none exists — never substitute serve harness or perf. |
| User-facing serve semantics | python3 scripts/serve_harness.py --model <path> … (after parity when numbers/state can change) |
| LFM2.5 chat / thinking frame | python3 scripts/serve_harness.py --model <path> --tag lfm2.5:… --sampling registry |
| Optional wrapper (Redline + serve; perf on by default) | scripts/gates.sh --model <path> --no-perf for harness-only; omit --no-perf only on a clean tree willing to let probe_commits.sh mutate checkout |
| Redline capture / shadow evidence | python3 scripts/redline_daemon_harness.py --model <path> + REDLINE.md ladder |
| Prefill/decode floor vs committed baselines | scripts/speed-gate.sh when that path’s policy applies |
| Fresh-process A/B | scripts/probe_commits.sh (default arm of gates.sh; disable with --no-perf. Mutates worktree — clean tree required) |
| Product-CLI synthetic daemon bench probe | hipfire bench <tag> / hipfire bench <tag> --matrix … — synthetic bench_prefill/bench_decode probes, not production-path authority |
| Production-path throughput report | Native serve_harness.py JSON / daemon done fields (prefill_tok_s, decode_tok_s, ttft_ms) per docs/methodology/perf-benchmarking.md and |
Retired batteries are not current evidence. Route claims through docs/VALIDATION.md; never require a retired battery for merge, promotion, or report pass/fail.
Report contract
Include identity, not folklore:
- GPU marketing name, gfx arch, ROCm/HIP version (
hipfire diag)
- hipfire version / binary path; daemon or bench binary md5 when claiming perf
- Exact model path or registry tag + file basename; draft path if DFlash-related
- Relevant config (
kv_cache / dflash_mode / thinking) and full command lines
- Prompt identity (path or md5) for any timed or semantic fixture
- Which VALIDATION route was run and what evidence class it covers
- Pass/fail per route; separate “measured” numbers from “admitted” claims
- Notes: hangs, OOM, first-run JIT cost, dirty worktree
Template and phase checklist: guide.md.
