gnome-jarvis/vendor/agent-harness/lib/device.js

/**
 * Device selection: Metal (macOS), Vulkan (Linux/Windows NVIDIA+AMD), CPU fallback.
 * CUDA/ROCm LLM backends are not shipped by QVAC.
 * GPU is always attempted unless the caller explicitly requests CPU.
 */

function metricValue(metric, fallback) {
  if (metric == null) return fallback;
  if (typeof metric !== 'object') return metric;
  if (metric.status === 'supported') return metric.value;
  return fallback;
}

function flattenDrivers(drivers) {
  const out = {};
  if (!drivers || typeof drivers !== 'object') return out;
  for (const key of Object.keys(drivers)) {
    const v = drivers[key];
    if (v === true) out[key] = true;
    else if (v && v.status === 'supported' && v.value) out[key] = true;
  }
  return out;
}

function normalizeResources(raw) {
  if (!raw || typeof raw !== 'object') {
    return { totalRamBytes: 0, vramBytes: 0, gpus: [], drivers: {} };
  }
  if (Array.isArray(raw.gpus) || raw.vramBytes != null || raw.vram != null) {
    const gpus = raw.gpus || [];
    const drivers = Object.keys(raw.drivers || {}).length ? flattenDrivers(raw.drivers) : {};
    if (!Object.keys(drivers).length) {
      for (const g of gpus) Object.assign(drivers, flattenDrivers(g && g.drivers));
    }
    return {
      totalRamBytes: Number(raw.totalRamBytes) || 0,
      vramBytes: vramBytes(raw),
      gpus,
      drivers,
      capabilities: raw.capabilities,
      gpu: raw.gpu,
    };
  }
  const caps = raw.capabilities || {};
  const gpuList = metricValue(caps.gpus, []) || [];
  const gpus = gpuList.map((g) => ({
    id: g.id,
    name: metricValue(g.name, null),
    type: metricValue(g.type, null),
    memory: metricValue(g.memoryTotalBytes, 0),
    drivers: flattenDrivers(g.drivers),
  }));
  const drivers = {};
  for (const g of gpus) Object.assign(drivers, g.drivers || {});
  let maxVram = 0;
  for (const g of gpus) {
    const m = Number(g.memory) || 0;
    if (m > maxVram) maxVram = m;
  }
  return {
    totalRamBytes: Number(metricValue(caps.memory && caps.memory.totalBytes, 0)) || 0,
    vramBytes: maxVram,
    gpus,
    drivers,
    capabilities: { gpu: gpus.length > 0, vulkan: !!drivers.vulkan, metal: !!drivers.metal },
  };
}

function pickDevice(requested, resources) {
  const want = String(requested || 'auto').toLowerCase();
  if (want === 'cpu') return { device: 'cpu', gpu_layers: 0 };
  return { device: 'gpu', gpu_layers: 99 };
}

function hasGpu(resources) {
  const res = resources && resources.capabilities && !Array.isArray(resources.gpus)
    ? normalizeResources(resources)
    : resources;
  if (!res) return false;
  const drivers = res.drivers || {};
  if (drivers.metal || drivers.vulkan || drivers.cuda || drivers.opencl) return true;
  const gpus = res.gpus || [];
  if (gpus.length > 0) return true;
  const cap = res.capabilities || {};
  if (cap.gpu || cap.vulkan || cap.metal) return true;
  if (res.gpu) return true;
  return false;
}

function backendLabel(resources) {
  const res = resources && resources.capabilities && !Array.isArray(resources.gpus)
    ? normalizeResources(resources)
    : resources;
  const gpus = (res && res.gpus) || [];
  const dedicated = gpus.find((g) => g && (g.type === 2 || String(g.type).toLowerCase() === 'dedicated'));
  const gpu = dedicated || gpus[0] || null;
  const gpuName =
    (gpu && gpu.name) ||
    (res && res.gpu && (res.gpu.name || res.gpu.deviceName)) ||
    null;
  const vram = (gpu && gpu.memory) || (res && (res.vram || res.vramBytes)) || null;
  const drivers = (res && res.drivers) || (gpu && gpu.drivers) || {};
  if (process.platform === 'darwin' || drivers.metal) {
    return { backend: 'metal', backendId: 1, deviceName: gpuName, vram };
  }
  if (drivers.vulkan || hasGpu(res)) {
    return { backend: 'vulkan', backendId: 3, deviceName: gpuName, vram };
  }
  return { backend: 'cpu', backendId: 0, deviceName: gpuName, vram };
}

function summarizeGpuProbe(info) {
  const res = normalizeResources(info);
  return {
    gpus: res.gpus || [],
    drivers: res.drivers || {},
    length: (res.gpus && res.gpus.length) || 0,
  };
}

/**
 * Normalize GPU memory reports to bytes. bare-gpu-info may yield bytes, MiB, or GiB.
 */
function vramBytes(resources) {
  if (!resources) return 0;
  const raw =
    resources.vramBytes != null
      ? resources.vramBytes
      : resources.vram != null
        ? resources.vram
        : resources.gpus && resources.gpus[0] && resources.gpus[0].memory;
  const n = Number(raw);
  if (!Number.isFinite(n) || n <= 0) return 0;
  if (n < 1024) return Math.round(n * 1024 * 1024 * 1024);
  if (n < 1024 * 1024) return Math.round(n * 1024 * 1024);
  return Math.round(n);
}

/**
 * Keep KV-cache context from dominating VRAM. Weights still use gpu_layers.
 */
function capCtxSize(want, resources, onGpu) {
  const requested = Math.max(512, Number(want) || 8192);
  if (!onGpu) return Math.min(requested, 8192);
  const gb = vramBytes(resources) / 1e9;
  let max = requested;
  if (gb > 0 && gb < 6) max = 4096;
  else if (gb > 0 && gb < 10) max = 8192;
  else if (gb > 0 && gb < 16) max = 16384;
  return Math.min(requested, max);
}

function mmprojOnGpu(opts, resources, onGpu) {
  if (!onGpu) return false;
  if (opts && opts.mmprojUseGpu === false) return false;
  return true;
}

function gpuLayers(opts, dev) {
  const raw = opts && opts.gpu_layers != null ? opts.gpu_layers : dev && dev.gpu_layers;
  const n = Number(raw);
  if (Number.isFinite(n) && n >= 0) return n;
  return dev && dev.device === 'gpu' ? 99 : 0;
}

module.exports = {
  pickDevice,
  hasGpu,
  backendLabel,
  summarizeGpuProbe,
  vramBytes,
  capCtxSize,
  mmprojOnGpu,
  normalizeResources,
  gpuLayers,
};