All runbooks
runbookuapiredacted view

UAPI Grok NUC Benchmark Gates 2026-06-22

Operator-gated Gemma/SLM benchmark plan for NUC NPU, Arc 140T, CPU, and fallback lanes.

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

UAPI Grok NUC Benchmark Gates 2026-06-22

Operator-gated Gemma/SLM benchmark plan for NUC NPU, Arc 140T, CPU, and fallback lanes.

Promotion Metadata

  • Source feedback id: 95
  • Source feedback ids: 95
  • Promotion request id: 28
  • Feedback category: consider
  • Feedback source: ai-lane
  • Feedback created at: 2026-06-22T16:56:36.545789+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 95

  • Product: uapi
  • Category: consider
  • Related slug: [REDACTED:high-entropy]
  • Source: ai-lane
  • Created at: 2026-06-22T16:56:36.545789+00:00

title: UAPI Grok NUC Benchmark Gates 2026-06-22 project: uapi kind: runbook state: gated tags: [docs, runbook, uapi, grok, nuc, benchmark, gemma]

UAPI Grok NUC Benchmark Gates 2026-06-22

Purpose: benchmark plan for Gemma/SLM placement on NUC NPU, Arc 140T-class GPU, CPU fallback, and cloud comparison. This is not approved to run by default.

Gate rule: every model load, inference run, download, sustained accelerator use, driver/config change, or raw benchmark publication requires explicit operator approval. Inventory commands are separate and read-only.

Pre-approval checklist: - Inventory packet completed and submitted to Explorer. - Operator confirms NUC CPU reservation / Conductor or Foundry safety window. - Exact model, backend, timeout, prompt count, and output path approved. - Existing local model files or download approval confirmed. - Metrics to capture: load time, first-token latency, tokens/sec, peak memory, device, power/thermal if available, stability, driver pain, API compatibility, fallback behavior.

First approved benchmark shape:

python -c "import time, psutil; print('mem_before', psutil.virtual_memory().percent)"
# Example only; adapt after inventory verifies stack:
# python benchmark_gemma.py --model google/gemma-2b-it --device NPU --precision int4 --num-prompts 3 --max-new-tokens 32 --warmup 1
# Repeat with --device GPU or AUTO for Arc 140T-class lane, and CPU for fallback.

OpenVINO / IPEX / llama.cpp targets: - OpenVINO GenAI or optimum-intel for NPU/iGPU when installed. - IPEX-LLM or PyTorch XPU if installed and compatible. - llama.cpp / llama-bench only if Intel backend is installed and model file is already approved.

llama.cpp illustrative pattern only:

# llama-bench -m gemma-2b-q4.gguf -n 32 -p 64
# main -m gemma-2b-q4.gguf -p "test" -n 32 --n-gpu-layers 99

Route claim rules: - local.nuc.gemma.npu cannot become production-capable without NPU benchmark evidence. - local.nuc.gemma.140t cannot become production-capable without Arc/iGPU benchmark evidence. - local.nuc.llm.fallback requires CPU timing and stability evidence. - cloud.colab.gpu.burst remains optional burst capacity, never control plane. - SN850X storage route requires verified attachment and policy.

Result packet requirements: - Include model, quantization, backend, device, driver versions, command, timestamp, metrics, caveats, and audit ids. - Publish sanitized evidence through Explorer; do not publish raw secrets, raw local paths if sensitive, or model/license material. - Codex must map accepted evidence into PlacementDecision and ProviderAdapter contracts before implementation.