{
  "platform": "Windows-10-10.0.26200-SP0",
  "torch": "2.8.0+cpu",
  "threads": 4,
  "batch": 1,
  "dtype": "float32",
  "device": "CPU",
  "tokenizer_sha256": "ef7f54d89966a08485006e0f64eaf2b0b99217b65ad79f96979f12adbd9ca49d",
  "results": [
    {
      "variant": "flm",
      "checkpoint_sha256": "88a53a8f128c9f09cc053dc461ffab24ecdc008074ace4042ff212d14069c6db",
      "parameters": 600003,
      "prefill_tokens": 128,
      "decode_tokens": 128,
      "repeats": 5,
      "prefill_seconds": [
        0.03439710009843111,
        0.03152840002439916,
        0.030261600157245994,
        0.03125330014154315,
        0.03717739996500313
      ],
      "decode_seconds": [
        0.6264657001011074,
        0.6387574998661876,
        0.6235859000589699,
        0.7074986000079662,
        0.7052372999023646
      ],
      "median_prefill_tokens_per_second": 4059.8317675791827,
      "median_decode_tokens_per_second": 200.38903656992605,
      "state": {
        "logical_tensor_bytes": 8192,
        "allocated_tensor_bytes": 8192,
        "scope": "Persistent sequence tensors only; excludes Python bookkeeping, model weights, temporary activations and adapters"
      }
    },
    {
      "variant": "gru",
      "checkpoint_sha256": "7f7ff50277a01569949d305ec55f3f9f17f6a73df569fb3f150d90072b127ee4",
      "parameters": 595408,
      "prefill_tokens": 128,
      "decode_tokens": 128,
      "repeats": 5,
      "prefill_seconds": [
        0.012766000116243958,
        0.011999999871477485,
        0.01188710005953908,
        0.011333400150761008,
        0.011942100012674928
      ],
      "decode_seconds": [
        0.03666500002145767,
        0.03823269996792078,
        0.03591539990156889,
        0.03761060000397265,
        0.035102800000458956
      ],
      "median_prefill_tokens_per_second": 10718.382852609278,
      "median_decode_tokens_per_second": 3491.0677737648934,
      "state": {
        "logical_tensor_bytes": 800,
        "allocated_tensor_bytes": 800,
        "scope": "Persistent sequence tensors only; excludes Python bookkeeping, model weights, temporary activations and adapters"
      }
    },
    {
      "variant": "transformer",
      "checkpoint_sha256": "7fe956869f6bccf8971e2d6764af488524cd714067095e170e7c2a96f94ca24b",
      "parameters": 607468,
      "prefill_tokens": 128,
      "decode_tokens": 128,
      "repeats": 5,
      "prefill_seconds": [
        0.007492100121453404,
        0.005500699859112501,
        0.0038805000949651003,
        0.003912499872967601,
        0.004148199921473861
      ],
      "decode_seconds": [
        0.1768930999096483,
        0.17518860008567572,
        0.17736539989709854,
        0.20661469991318882,
        0.1742275001015514
      ],
      "median_prefill_tokens_per_second": 30856.75773180224,
      "median_decode_tokens_per_second": 723.6008643942504,
      "state": {
        "logical_tensor_bytes": 164160,
        "allocated_tensor_bytes": 164160,
        "scope": "Persistent sequence tensors only; excludes Python bookkeeping, model weights, temporary activations and adapters"
      }
    }
  ],
  "protocol": "Call only after all training processes have exited. One warmup and five repeated identical original-text trials; tokenization excluded; forced identical token inputs across models.",
  "caveat": "This measures the implemented eager PyTorch path, including FLM recurrent-weight preparation per forward call. It is not a hardware-independent FLOP, peak-memory or energy comparison. Normal OS background activity can affect timings."
}