config.yaml

version: "0.2.1"
mode: simulation
scenario_name: local-dry-run
seed: 7
artifacts_dir: artifacts
duration_s: 180
step_s: 1
sensor:
  kind: simulated
  support_level: mock-only
  gpu_count: 1
  noise_c: 0.35
  initial_temp_c: 72.0
  ambient_temp_c: 32.0
  rise_per_pressure: 1.7
  cooling_per_step: 0.4
  kv_relief_cooling: 2.0
  thermal_inertia: 0.82
  stale_after_s: 3
workload:
  kind: simulated
  base_request_rate: 6.0
  burst_request_rate: 10.5
  burst_start_s: 35
  burst_duration_s: 80
  prompt_length_tokens: 131072
  decode_tokens: 768
  base_latency_ms: 1450.0
  queue_sensitivity: 110.0
  thermal_latency_gain: 320.0
  throughput_per_batch: 2.2
  kv_growth_rate: 0.018
  kv_relief_factor: 0.22
  oscillation_penalty_ms: 75.0
policy:
  throttle_temp_c: 85.0
  recover_temp_c: 80.0
  min_batch_size: 16
  max_batch_size: 256
  initial_batch_size: 256
  throttle_step_ratio: 0.5
  recover_step_ratio: 2.0
  min_dwell_s: 8
  cooldown_s: 10
  anti_flap_window_s: 30
  max_actions_per_window: 3
  kv_migration_pct: 0.12
  degraded_retries_before_hold: 2
  dry_run: true
  enable_kv_migration: true
backends:
  batch:
    kind: mock
    support_level: mock-only
    base_url: http://localhost:8000/v1/admin
    timeout_s: 1.0
    experimental_note: "Switch to http-admin only after validating your serving stack."
  kv_migration:
    kind: mock
    support_level: mock-only
    base_url: http://localhost:8000/v1/admin
    timeout_s: 1.0
    experimental_note: "Modeled pressure-relief hook; replace with a real implementation only after validation."
metrics:
  prometheus: false
  port: 9091