import { kvBytesPerToken, kvCacheForContext, DEFAULT_ARCHITECTURES } from "@agentium/core";

const llama70b = DEFAULT_ARCHITECTURES["llama-3.1-70b"];

// Per-token KV cost
console.log(kvBytesPerToken(llama70b, "bf16"));  // 327,680 bytes
console.log(kvBytesPerToken(llama70b, "fp8"));   // 163,840 bytes

// Full 128K context
const full = kvCacheForContext(llama70b, 131_072, "bf16");
console.log(`${full.gb.toFixed(1)} GB`);  // 40.0 GB

import { planCapacity, DEFAULT_ARCHITECTURES, DEFAULT_GPU_SPECS } from "@agentium/core";

const plan = planCapacity(
  DEFAULT_ARCHITECTURES["llama-3.1-70b"],
  { gpu: DEFAULT_GPU_SPECS["h100-sxm"], gpuCount: 4, nandPerGpuGb: 0, nandBandwidthGBs: 7 },
  { extreme: 1, heavy: 2, medium: 3, light: 4 },
  "fp8", "bf16",
);

console.log(`HBM slots:     ${plan.hbmSlots}`);
console.log(`TTFT breach:   ${plan.ttftBreachPoint} users`);
console.log(`Monthly cost:  $${plan.monthlyGpuCostUsd.toLocaleString()}`);

import { compareConfigs, DEFAULT_ARCHITECTURES, DEFAULT_GPU_SPECS } from "@agentium/core";

const results = compareConfigs(
  DEFAULT_ARCHITECTURES["llama-3.1-70b"],
  [
    { label: "4× H100",          hardware: { gpu: DEFAULT_GPU_SPECS["h100-sxm"], gpuCount: 4, nandPerGpuGb: 0, nandBandwidthGBs: 7 } },
    { label: "4× H100 + 4TB SSD", hardware: { gpu: DEFAULT_GPU_SPECS["h100-sxm"], gpuCount: 4, nandPerGpuGb: 4000, nandBandwidthGBs: 7 } },
    { label: "8× H100",          hardware: { gpu: DEFAULT_GPU_SPECS["h100-sxm"], gpuCount: 8, nandPerGpuGb: 0, nandBandwidthGBs: 7 } },
  ],
  { extreme: 1, heavy: 2, medium: 3, light: 4 },
);

for (const r of results) {
  console.log(`${r.label}: ${r.plan.totalSessions} sessions, $${r.monthlyCost}/mo, $${r.costPerSession.toFixed(2)}/session`);
}

# With OpenAI (default)
OPENAI_API_KEY=sk-... npx tsx examples/capacity/live-monitor.ts

# With a self-hosted vLLM endpoint
VLLM_BASE_URL=https://your-endpoint/v1 \
VLLM_API_KEY=your-key \
npx tsx examples/capacity/live-monitor.ts

── Observed Workload ──────────────────────────────────────
  Sessions:           light=4, medium=0, heavy=0, extreme=0
  Total tokens:       13,067
  Avg tokens/session: 3,267
  Estimated KV (bf16):3.988 GB

── Theoretical Capacity (8× NVIDIA H100 SXM) ──────────────
  Total HBM:          640 GB
  Weight memory:      35 GB (int4 AWQ)
  Free for KV:        600 GB
  HBM slots:          112
  Total sessions:     112

── Latency Estimates ──────────────────────────────────────
  TTFT breach point:  629 concurrent users (5s SLA)

── Cost ──────────────────────────────────────────────────
  Monthly GPU cost:   $23,302
  Per slot/day:       $6.94 (112 slots)
  Est. cost/1K tok:   $1.2384 (self-hosted)

── Headroom Check ────────────────────────────────────────
  ✓ 108 more sessions can fit in HBM (4/112 used)
  ✓ TTFT safe — 4 users well below 629-user breach point

import { restoreLatency, kvCacheForContext, DEFAULT_ARCHITECTURES } from "@agentium/core";

const llama70b = DEFAULT_ARCHITECTURES["llama-3.1-70b"];
const kvGb = kvCacheForContext(llama70b, 32768, "fp8").gb;

console.log(`Gen4 NVMe: ${(restoreLatency(kvGb, 7) / 1000).toFixed(1)}s`);
console.log(`Gen5 NVMe: ${(restoreLatency(kvGb, 14) / 1000).toFixed(1)}s`);