Serving Latency Percentiles (Plotly)

{
  "data": [
    {"type": "bar", "name": "p50", "marker": {"color": "#bae6fd"},
     "x": ["Small 8B", "Mid 70B", "Frontier", "Small 8B + rerank", "Router (mixed)"],
     "y": [210, 640, 1480, 390, 470]},
    {"type": "bar", "name": "p95", "marker": {"color": "#38bdf8"},
     "x": ["Small 8B", "Mid 70B", "Frontier", "Small 8B + rerank", "Router (mixed)"],
     "y": [430, 1290, 3120, 780, 1010]},
    {"type": "bar", "name": "p99", "marker": {"color": "#0369a1"},
     "x": ["Small 8B", "Mid 70B", "Frontier", "Small 8B + rerank", "Router (mixed)"],
     "y": [880, 2340, 5600, 1420, 1980]}
  ],
  "layout": {
    "title": "Time to first token by serving configuration (ms)",
    "barmode": "group",
    "yaxis": {"title": "Milliseconds"},
    "shapes": [
      {"type": "line", "xref": "paper", "x0": 0, "x1": 1, "y0": 2000, "y1": 2000,
       "line": {"color": "#dc2626", "dash": "dot", "width": 2}}
    ],
    "annotations": [
      {"xref": "paper", "x": 0, "y": 2000, "xanchor": "left", "yshift": 12,
       "text": "p99 budget 2,000 ms", "showarrow": false, "font": {"color": "#dc2626"}}
    ]
  }
}