# Minimum compute guidance for the Babelbit private validator.
#
# This file is intentionally based on observed CPU-only audio scoring behavior
# in this repo rather than generic cloud sizing.
#
# Benchmark basis:
#   .venv/bin/python tests/benchmarks/stress_stt.py \
#     --miners 50 100 150 200 250 \
#     --durations-sec 60 \
#     --audio-kind mixed \
#     --stt-model faster-whisper-small \
#     --stt-device cpu \
#     --warmup 0 \
#     --repeat 1 \
#     --json
#
# Repo defaults used by the runner:
#   BB_AUDIO_SCORING_STT_MODEL=faster-whisper-small
#   BB_AUDIO_SCORING_STT_DEVICE=cpu
#
# Practical takeaway from testing:
# - 4 vCPU / 8 GB RAM on DigitalOcean Regular Intel was too slow for validator
#   scoring and would take hours on larger batches.
# - 8 vCPU / 16 GB RAM on DigitalOcean Premium AMD completed the same
#   250-miner / 60-second synthetic STT batch in about 24 minutes.
# - Because this benchmark is STT-only, these numbers should be treated as the
#   floor for acceptable validator performance, not a generous recommendation.

specs:
  minimum:
    cpu: 8
    ram_gb: 16
    gpu: 0
    storage_gb: 80
    network: moderate
    os: ubuntu-22.04
    profile: cpu-only validator floor for default audio scoring
    rationale:
      - Matches the first tested machine that completed default CPU STT scoring in a reasonable window.
      - Sized for faster-whisper-small on CPU plus validator services and cache growth.
      - Anything below this, especially 4 vCPU / 8 GB RAM, should be treated as unsupported for production scoring.

  recommended:
    cpu: 8
    ram_gb: 16
    gpu: 0
    storage_gb: 120
    network: high
    os: ubuntu-22.04
    profile: stable cpu-only validator for current scoring defaults
    rationale:
      - Same compute tier as the successful AMD benchmark box.
      - Leaves more room for Docker layers, Hugging Face cache, score artifacts, and logs.
      - Suitable when you intend to keep the current default STT model and device settings.

  below_minimum_observed:
    cpu: 4
    ram_gb: 8
    gpu: 0
    status: not recommended
    platform_estimate: DigitalOcean Regular Intel
    cpu_family_estimate: Intel Broadwell Xeon class
    rationale:
      - Observed on a DigitalOcean Regular Intel VM.
      - `lscpu` reported `GenuineIntel`, family `6`, model `79`, which is consistent with Broadwell-era Xeon hosts under KVM.
      - Expected scoring time stretched into hours for the same workload shape.
      - Not appropriate for timely validator scoring.

benchmark_summary:
  workload:
    miners: 250
    utterance_duration_sec: 60
    total_audio_sec: 15000
    audio_kind: mixed
    stt_model: faster-whisper-small
    stt_device: cpu
  observed_on_successful_box:
    cpu: 8 vCPU
    ram_gb: 16
    platform_estimate: DigitalOcean Premium AMD
    cpu_family_estimate: AMD EPYC Rome / Zen 2 class
    wall_sec: 1402.095
    stt_wall_sec: 1396.263
    max_rss_mb: 2517.895
    audio_sec_per_wall_sec: 10.698
    notes:
      - `lscpu` reported `AuthenticAMD`, family `23`, model `49`, and `DO-Premium-AMD`, which is consistent with a Premium AMD droplet backed by EPYC-class CPUs.
      - About 23 to 24 minutes for the STT portion of scoring.
      - This benchmark does not include the full scoring stack such as embeddings, metadata lookup, and persistence.

runner_defaults:
  stt_model: faster-whisper-small
  stt_device: cpu
  embedder: all-MiniLM-L6-v2

notes:
  - If runner defaults change to a larger STT model, revisit this file.
  - If you need materially faster scoring windows, move above this floor rather than below it.
  - Storage should account for OS, virtualenv, Docker images, logs, score artifacts, and Hugging Face model cache.