elizaos--eliza/plugins/plugin-local-inference/scripts/local-inference-thresholds.json

{
  "_comment": "Per-(backend, variant) minimum tok/s floors for the local-inference ablation gate. Floors are intentionally generous on first-pass CI to avoid false-fails on shared/cold runners; tighten over time as we collect baselines per runner class. Use `default` as the per-backend fallback when a variant-specific floor is not listed.",
  "cpu": {
    "default": 3,
    "baseline_f16_kv": 5,
    "turbo4_polar_kv": 4,
    "turbo3_polar_kv": 4,
    "qjl_tcq_forced": 3,
    "mtp_only": 4,
    "mtp_turbo4_polar": 3,
    "all_mtp_qjl_tcq": 3
  },
  "linux-x64-cpu": {
    "_comment": "Linux x86_64 CPU managed fork floor. Keep separate from generic cpu so desktop regressions do not hide behind mobile/CI baselines.",
    "default": 3,
    "baseline_f16_kv": 5,
    "turbo4_polar_kv": 4,
    "turbo3_polar_kv": 4,
    "qjl_tcq_forced": 3,
    "mtp_only": 4,
    "mtp_turbo4_polar": 3,
    "all_mtp_qjl_tcq": 3
  },
  "linux-x86_64-cpu": {
    "_comment": "Canonical-arch alias for linux-x64-cpu.",
    "default": 3,
    "baseline_f16_kv": 5,
    "turbo4_polar_kv": 4,
    "turbo3_polar_kv": 4,
    "qjl_tcq_forced": 3,
    "mtp_only": 4,
    "mtp_turbo4_polar": 3,
    "all_mtp_qjl_tcq": 3
  },
  "win32-x64-cpu": {
    "_comment": "Runtime process.platform key for Windows x86_64 CPU.",
    "default": 3,
    "baseline_f16_kv": 5,
    "turbo4_polar_kv": 4,
    "turbo3_polar_kv": 4,
    "qjl_tcq_forced": 3,
    "mtp_only": 4,
    "mtp_turbo4_polar": 3,
    "all_mtp_qjl_tcq": 3
  },
  "win32-x86_64-cpu": {
    "_comment": "Canonical-arch alias for Windows x86_64 CPU.",
    "default": 3,
    "baseline_f16_kv": 5,
    "turbo4_polar_kv": 4,
    "turbo3_polar_kv": 4,
    "qjl_tcq_forced": 3,
    "mtp_only": 4,
    "mtp_turbo4_polar": 3,
    "all_mtp_qjl_tcq": 3
  },
  "windows-x64-cpu": {
    "_comment": "Managed artifact key for Windows x86_64 CPU builds.",
    "default": 3,
    "baseline_f16_kv": 5,
    "turbo4_polar_kv": 4,
    "turbo3_polar_kv": 4,
    "qjl_tcq_forced": 3,
    "mtp_only": 4,
    "mtp_turbo4_polar": 3,
    "all_mtp_qjl_tcq": 3
  },
  "windows-x86_64-cpu": {
    "_comment": "Canonical-arch alias for managed Windows x86_64 CPU builds.",
    "default": 3,
    "baseline_f16_kv": 5,
    "turbo4_polar_kv": 4,
    "turbo3_polar_kv": 4,
    "qjl_tcq_forced": 3,
    "mtp_only": 4,
    "mtp_turbo4_polar": 3,
    "all_mtp_qjl_tcq": 3
  },
  "cuda": {
    "default": 50,
    "baseline_f16_kv": 80,
    "turbo4_polar_kv": 70,
    "turbo3_polar_kv": 70,
    "qjl_tcq_forced": 50,
    "mtp_only": 80,
    "mtp_turbo4_polar": 70,
    "all_mtp_qjl_tcq": 50
  },
  "rocm": {
    "default": 30,
    "baseline_f16_kv": 50,
    "turbo4_polar_kv": 40,
    "turbo3_polar_kv": 40,
    "qjl_tcq_forced": 30,
    "mtp_only": 50,
    "mtp_turbo4_polar": 40,
    "all_mtp_qjl_tcq": 30
  },
  "metal": {
    "default": 20,
    "baseline_f16_kv": 30,
    "turbo4_polar_kv": 25,
    "turbo3_polar_kv": 25,
    "qjl_tcq_forced": 20,
    "mtp_only": 30,
    "mtp_turbo4_polar": 25,
    "all_mtp_qjl_tcq": 20
  },
  "vulkan": {
    "_comment": "Vulkan thresholds depend heavily on the device — discrete dGPU vs Mali/Adreno mobile chips have orders-of-magnitude different floors. Floors here target a low-end discrete GPU baseline; tighten per-device when concrete runners come online. Mobile-class Adreno/Mali should be benched separately under android-vulkan and given their own threshold block once we have data.",
    "default": 25,
    "baseline_f16_kv": 40,
    "turbo4_polar_kv": 35,
    "turbo3_polar_kv": 35,
    "qjl_tcq_forced": 25,
    "mtp_only": 40,
    "mtp_turbo4_polar": 35,
    "all_mtp_qjl_tcq": 25
  },
  "android-arm64-cpu": {
    "_comment": "Physical Android arm64 CPU floor for the in-process/mobile libllama path. Kept separate from desktop cpu so mobile thermal and big.LITTLE behavior do not water down desktop regressions.",
    "default": 1,
    "baseline_f16_kv": 2,
    "turbo4_polar_kv": 1,
    "turbo3_polar_kv": 1,
    "qjl_tcq_forced": 1,
    "mtp_only": 1,
    "mtp_turbo4_polar": 1,
    "all_mtp_qjl_tcq": 1
  },
  "android-x86_64-cpu": {
    "_comment": "Cuttlefish/emulator Android x86_64 CPU floor. Diagnostic only until a model-staged APK completes the full chat smoke.",
    "default": 1,
    "baseline_f16_kv": 2,
    "turbo4_polar_kv": 1,
    "turbo3_polar_kv": 1,
    "qjl_tcq_forced": 1,
    "mtp_only": 1,
    "mtp_turbo4_polar": 1,
    "all_mtp_qjl_tcq": 1
  },
  "android-x64-cpu": {
    "_comment": "Node process.arch compatibility alias for android-x86_64-cpu.",
    "default": 1,
    "baseline_f16_kv": 2,
    "turbo4_polar_kv": 1,
    "turbo3_polar_kv": 1,
    "qjl_tcq_forced": 1,
    "mtp_only": 1,
    "mtp_turbo4_polar": 1,
    "all_mtp_qjl_tcq": 1
  },
  "android-arm64-vulkan": {
    "_comment": "Physical Android Vulkan floor for Adreno/Mali. Intentionally not inherited from desktop vulkan; tighten by GPU family once Adreno and Mali graph-dispatch reports exist.",
    "default": 5,
    "baseline_f16_kv": 8,
    "turbo4_polar_kv": 6,
    "turbo3_polar_kv": 6,
    "qjl_tcq_forced": 5,
    "mtp_only": 8,
    "mtp_turbo4_polar": 6,
    "all_mtp_qjl_tcq": 5
  },
  "android-x86_64-vulkan": {
    "_comment": "Android x86_64 Vulkan floor for real ChromeOS/ARCVM GPU silicon. Cuttlefish SwiftShader runs are diagnostic and should not be used as recordable evidence.",
    "default": 5,
    "baseline_f16_kv": 8,
    "turbo4_polar_kv": 6,
    "turbo3_polar_kv": 6,
    "qjl_tcq_forced": 5,
    "mtp_only": 8,
    "mtp_turbo4_polar": 6,
    "all_mtp_qjl_tcq": 5
  },
  "android-x64-vulkan": {
    "_comment": "Node process.arch compatibility alias for android-x86_64-vulkan.",
    "default": 5,
    "baseline_f16_kv": 8,
    "turbo4_polar_kv": 6,
    "turbo3_polar_kv": 6,
    "qjl_tcq_forced": 5,
    "mtp_only": 8,
    "mtp_turbo4_polar": 6,
    "all_mtp_qjl_tcq": 5
  }
}