ModelNative quickstart

Start with the playground

Open /demos/modelnative, load the supplied example, change a scenario constraint and run. Download the resulting evidence JSON. The hosted adapter uses a bounded subprocess, not an animation.

Input contract

phases describes capability bits, memory demand, parallelism, work units, I/O and latency objectives. engines supplies available memory, health, capabilities and performance-model inputs. The public adapter caps engines at 16 and phases at 8. Latency and energy are model estimates from supplied data, not silicon benchmarks.

{
  "phases": [
    {
      "name": "prefill",
      "required_caps": 1,
      "memory_mb": 4096,
      "parallelism": 2,
      "work_units": 3000,
      "io_gb": 4,
      "latency_slo_us": 200
    },
    {
      "name": "decode",
      "required_caps": 5,
      "memory_mb": 2048,
      "parallelism": 2,
      "work_units": 1000,
      "io_gb": 2,
      "latency_slo_us": 150
    }
  ],
  "engines": [
    {
      "id": "m0",
      "valid": true,
      "health": 95,
      "caps": 7,
      "memory_mb": 3072,
      "tops": 100,
      "bandwidth_gbps": 800,
      "energy_per_work": 0.005
    },
    {
      "id": "m1",
      "valid": true,
      "health": 92,
      "caps": 7,
      "memory_mb": 3072,
      "tops": 90,
      "bandwidth_gbps": 600,
      "energy_per_work": 0.004
    }
  ]
}

Expected result

The sample produces separate prefill and decode plans. The full output below is captured from the unchanged archived engine. Memory-fit and SLO results describe this scenario only; they are not a claim about trained models or devices.

{
  "status": "planned",
  "phases": [
    {
      "phase": "prefill",
      "status": "planned",
      "engines": [
        "m0",
        "m1"
      ],
      "shards": 2,
      "memory_fit": true,
      "estimated_latency_us": 43.3333,
      "latency_slo_us": 200.0,
      "slo_met": true,
      "estimated_energy": 27.0
    },
    {
      "phase": "decode",
      "status": "planned",
      "engines": [
        "m0"
      ],
      "shards": 1,
      "memory_fit": true,
      "estimated_latency_us": 30.0,
      "latency_slo_us": 150.0,
      "slo_met": true,
      "estimated_energy": 5.0
    }
  ],
  "all_slos_met": true,
  "estimated_total_energy": 32.0
}

Use the licensed package

After entitled delivery, extract the named commercial release and run its CLI in an isolated Python environment. These commands refer to the actual archived release, not a package invented for the website.

python -m pip install ./dist/garmo_modelnative-1.2.0-py3-none-any.whl
python -m modelnative advanced-demo
python -m modelnative serve --port 8787

What this demonstrates

The algorithm is executing real software on synthetic input. Production datasets, calibration, integration performance, service levels and customer ROI require separate evaluation. The website adapter limits requests and does not expose the release source.