{
  "platform": "macOS-27.0-arm64-arm-64bit",
  "machine": "arm64",
  "torch": "2.13.0",
  "method": "one process per model and device; auto_quant memory-linear attention for all three; median over 110 benchmark items under 1,024 tokens after 10 warm-up items; one ~8k- and one ~32k-token item. The BF16 checkpoint runs in FP32 here (CPU/MPS default). Memory is the resident-set growth over the process baseline; on MPS, weights in memory = torch.mps.current_allocated_memory() after load, and the peak = the Metal driver allocation (includes the allocator cache).",
  "runs": [
    {
      "model": "bf16",
      "device": "cpu",
      "dtype": "torch.float32",
      "load_seconds": 1.97,
      "rss_after_load_gb": 0.787,
      "mps_driver_after_load_gb": null,
      "short_items": 110,
      "median_ms_under_1k_tokens": 68.22,
      "long": [
        {
          "tokens": 7449,
          "seconds": 1.65,
          "peak_rss_gb": 1.33,
          "peak_mps_driver_gb": null
        },
        {
          "tokens": 29233,
          "seconds": 15.66,
          "peak_rss_gb": 2.12,
          "peak_mps_driver_gb": null
        }
      ]
    },
    {
      "model": "int8",
      "device": "cpu",
      "dtype": "torch.float32",
      "load_seconds": 1.77,
      "rss_after_load_gb": 0.37,
      "mps_driver_after_load_gb": null,
      "short_items": 110,
      "median_ms_under_1k_tokens": 69.84,
      "long": [
        {
          "tokens": 7449,
          "seconds": 1.49,
          "peak_rss_gb": 0.97,
          "peak_mps_driver_gb": null
        },
        {
          "tokens": 29233,
          "seconds": 13.44,
          "peak_rss_gb": 1.77,
          "peak_mps_driver_gb": null
        }
      ]
    },
    {
      "model": "int4",
      "device": "cpu",
      "dtype": "torch.float32",
      "load_seconds": 1.66,
      "rss_after_load_gb": 0.298,
      "mps_driver_after_load_gb": null,
      "short_items": 110,
      "median_ms_under_1k_tokens": 104.44,
      "long": [
        {
          "tokens": 7449,
          "seconds": 1.51,
          "peak_rss_gb": 0.9,
          "peak_mps_driver_gb": null
        },
        {
          "tokens": 29233,
          "seconds": 12.49,
          "peak_rss_gb": 1.73,
          "peak_mps_driver_gb": null
        }
      ]
    },
    {
      "model": "bf16",
      "device": "mps",
      "dtype": "torch.float32",
      "load_seconds": 2.0,
      "rss_after_load_gb": 0.233,
      "mps_tensors_after_load_gb": 0.557,
      "short_items": 110,
      "median_ms_under_1k_tokens": 24.45,
      "long": [
        {
          "tokens": 7449,
          "seconds": 0.46,
          "peak_rss_gb": 0.77,
          "peak_mps_driver_gb": 1.03
        },
        {
          "tokens": 29233,
          "seconds": 4.42,
          "peak_rss_gb": 0.62,
          "peak_mps_driver_gb": 2.02
        }
      ]
    },
    {
      "model": "int8",
      "device": "mps",
      "dtype": "torch.float32",
      "load_seconds": 1.79,
      "rss_after_load_gb": 0.455,
      "mps_tensors_after_load_gb": 0.14,
      "short_items": 110,
      "median_ms_under_1k_tokens": 29.83,
      "long": [
        {
          "tokens": 7449,
          "seconds": 0.47,
          "peak_rss_gb": 0.79,
          "peak_mps_driver_gb": 1.06
        },
        {
          "tokens": 29233,
          "seconds": 4.09,
          "peak_rss_gb": 0.54,
          "peak_mps_driver_gb": 2.05
        }
      ]
    },
    {
      "model": "int4",
      "device": "mps",
      "dtype": "torch.float32",
      "load_seconds": 1.77,
      "rss_after_load_gb": 0.392,
      "mps_tensors_after_load_gb": 0.072,
      "short_items": 110,
      "median_ms_under_1k_tokens": 37.83,
      "long": [
        {
          "tokens": 7449,
          "seconds": 0.47,
          "peak_rss_gb": 0.59,
          "peak_mps_driver_gb": 1.09
        },
        {
          "tokens": 29233,
          "seconds": 4.28,
          "peak_rss_gb": 0.54,
          "peak_mps_driver_gb": 2.07
        }
      ]
    }
  ]
}