All runbooks
runbookuapiredacted view

UAPI Grok NUC Gemma Smoke Gate Pack 2026-06-22

Grok operator gate pack for later NUC OpenVINO GenAI install/download/benchmark approval.

Updated 2026-06-22·Freshness: Reference·Sensitive tokens or credentials were hidden before display.

UAPI Grok NUC Gemma Smoke Gate Pack 2026-06-22

Grok operator gate pack for later NUC OpenVINO GenAI install/download/benchmark approval.

Promotion Metadata

  • Source feedback id: 129
  • Source feedback ids: 129
  • Promotion request id: 60
  • Feedback category: beneficial
  • Feedback source: [REDACTED:high-entropy]
  • Feedback created at: 2026-06-22T18:32:58.597168+00:00
  • Target slug: [REDACTED:high-entropy]
  • Review state: review
  • Reviewer note: AI lane auto-promoted redacted feedback into internal catalog context.

Source Feedback Body

Feedback 129

  • Product: uapi
  • Category: beneficial
  • Related slug: [REDACTED:high-entropy]
  • Source: [REDACTED:high-entropy]
  • Created at: 2026-06-22T18:32:58.597168+00:00

Grok NUC Gemma Smoke Gate Pack (2026-06-22)

Authority: Operator gate pack. Local artifact only. No execution/installs/downloads/benchmarks here.

Lane: UAPI NUC Gemma smoke. OpenVINO GenAI primary. Targets: local.nuc.gemma.npu, local.nuc.gemma.140t (candidates only).

Scope

Gate for later install + small Gemma/SLM download + smoke benchmark approval on NUC-IPEX (Ultra 9 285H + Arc 140T + NPU). One-shot, human-present. 1-3 short prompts/device. Max new tokens 32-64. No prod claims pre-evidence + re-bootstrap + Codex.

Allowed (post explicit operator approval)

  • pip (venv/--user): openvino, openvino-genai, optimum[openvino], nncf.
  • optimum-cli export openvino --model --weight-format int4 --sym --ratio 1.0 --group-size 128 .
  • Smoke: LLMPipeline(..., "NPU") and "GPU" only. Fixed short prompts.
  • Record sanitized metrics + stdout. Use reverse path. Idle confirmed.
  • Targeted cleanup + re-probe absent. Evidence promotion.

Blocked (always)

  • Any install/download/export without gate.
  • Models >~4B class for NPU smoke (no 7B+).
  • Long/sustained/context runs, multi-runs, background, concurrent.
  • Infra changes, public routes, WSL, drivers, secrets/PII.
  • Pre-gate inference. No local.nuc.* claims until evidence + contracts.
  • SN850X (boot=SN8100).

Metrics (per device NPU then 140T)

first_token_ms, tokens_per_sec (mean), load_time_s, peak_mem_mb, device (observed fallback), stability, model+quant+flags. Targets (indicative): first<3000ms, t/s>10-15 idle, load<60s. Record actuals.

Rollback / Cleanup

pip uninstall openvino openvino-genai optimum nncf (or venv). rm -rf targeted export/HF/OpenVINO caches. Re-probe packages ABSENT. Re-preflight baseline. Re-bootstrap if needed.

Evidence Checklist (pre any route)

  • [ ] Re-preflight: CPU 285H, Arc 140T (drv 32.0.101.8331), NPU OK, ~64GiB, Python 3.13, idle, C free.
  • [ ] Install log + pip freeze versions.
  • [ ] Export success + IR files.
  • [ ] Sanitized smoke (NPU+140T) + metrics table.
  • [ ] Errors/fallbacks noted.
  • [ ] Cleanup verified (absent).
  • [ ] Pre/post idle + trace.
  • [ ] Sources/blockers in evidence.

Sources / Context

  • [REDACTED:high-entropy].md
  • grok-benchmark-preflight-command-plan.md
  • [REDACTED:high-entropy].md
  • agent-work-queue-v2-unblocked.md + grok-140t-npu-primary-stack-review.md
  • Inventory: NUC15CRSU9, SN8100, packages absent.

Assumptions

Small Gemma 2B/3B (e.g. gemma-2-2b-it) export/run with int4 on NPU/140T. OpenVINO GenAI stable on this driver. ~64GiB OK for smoke. Gate before any change.

Blockers

Operator gate required. Codex contracts (PlacementDecision/routes) pending. No packages/weights today. NPU limits (e.g. prompt ~1024) unknown pre-smoke. SN850X separate.

Next Handoff

Promote pack. Await operator approval. Execute smoke, sanitize evidence, re-bootstrap, Codex route review.

Report: ALIVE true | Worker C | gate-pack + model-notes | No installs etc. End.