{
  "device": "cpu",
  "dtype": "torch.float64",
  "seed": 2026,
  "torch": "2.6.0+cpu",
  "python": "3.12.14",
  "hardware": "x86_64",
  "cpu_threads": 1,
  "process_peak_rss_KiB": 358616,
  "notes": [
    "Single-process measurements, not speedup guarantees.",
    "Warm-up is excluded from evaluation.",
    "CPU RSS is process high-water mark; state tensor bytes are not total memory.",
    "Async recording overlaps disk writing; host staging is included on the caller."
  ],
  "measurements": [
    {
      "kind": "sampler",
      "method": "ULA eager",
      "dimension": 8,
      "chains": 32,
      "warmup_steps": 8,
      "warmup_seconds": 0.0037236469997878885,
      "evaluation_steps": 60,
      "evaluation_seconds": 0.009968960000151128,
      "seconds_per_step": 0.00016614933333585212,
      "live_state_tensor_bytes": 2048,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "sampler",
      "method": "ULA eager",
      "dimension": 64,
      "chains": 32,
      "warmup_steps": 8,
      "warmup_seconds": 0.002270383999530168,
      "evaluation_steps": 60,
      "evaluation_seconds": 0.013015965000704455,
      "seconds_per_step": 0.00021693275001174092,
      "live_state_tensor_bytes": 16384,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "sampler",
      "method": "ULA eager",
      "dimension": 256,
      "chains": 32,
      "warmup_steps": 8,
      "warmup_seconds": 0.0035125140002492117,
      "evaluation_steps": 60,
      "evaluation_seconds": 0.02658271100062848,
      "seconds_per_step": 0.00044304518334380795,
      "live_state_tensor_bytes": 65536,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "particles",
      "method": "SVGD exact blocked",
      "dimension": 8,
      "particles": 16,
      "pair_block": 32,
      "warmup_steps": 3,
      "warmup_seconds": 0.0031958150002537877,
      "evaluation_steps": 20,
      "evaluation_seconds": 0.01726601100017433,
      "seconds_per_step": 0.0008633005500087165,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "particles",
      "method": "SVGD exact blocked",
      "dimension": 8,
      "particles": 64,
      "pair_block": 32,
      "warmup_steps": 3,
      "warmup_seconds": 0.006657544000518101,
      "evaluation_steps": 20,
      "evaluation_seconds": 0.05084107800030324,
      "seconds_per_step": 0.002542053900015162,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "particles",
      "method": "SVGD exact blocked",
      "dimension": 8,
      "particles": 128,
      "pair_block": 32,
      "warmup_steps": 3,
      "warmup_seconds": 0.0240312450005149,
      "evaluation_steps": 20,
      "evaluation_seconds": 0.18414928899983352,
      "seconds_per_step": 0.009207464449991676,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "particles",
      "method": "SVGD exact blocked",
      "dimension": 32,
      "particles": 16,
      "pair_block": 32,
      "warmup_steps": 3,
      "warmup_seconds": 0.002507704999516136,
      "evaluation_steps": 20,
      "evaluation_seconds": 0.013973362999422534,
      "seconds_per_step": 0.0006986681499711267,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "particles",
      "method": "SVGD exact blocked",
      "dimension": 32,
      "particles": 64,
      "pair_block": 32,
      "warmup_steps": 3,
      "warmup_seconds": 0.009194703000503068,
      "evaluation_steps": 20,
      "evaluation_seconds": 0.05201439499978733,
      "seconds_per_step": 0.0026007197499893662,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "particles",
      "method": "SVGD exact blocked",
      "dimension": 32,
      "particles": 128,
      "pair_block": 32,
      "warmup_steps": 3,
      "warmup_seconds": 0.0250877809994563,
      "evaluation_steps": 20,
      "evaluation_seconds": 0.1693888199997673,
      "seconds_per_step": 0.008469440999988365,
      "cuda_peak_allocated_bytes": null
    },
    {
      "kind": "recording",
      "mode": "off",
      "dimension": 64,
      "chains": 32,
      "every": 1,
      "warmup_steps": 8,
      "warmup_seconds": 0.0021888430001126835,
      "evaluation_steps": 60,
      "evaluation_seconds": 0.014824393999333552,
      "seconds_per_step": 0.00024707323332222585,
      "includes_CPU_staging_and_flush": true,
      "disk_bytes": 0,
      "writer_stats": null
    },
    {
      "kind": "recording",
      "mode": "sync",
      "dimension": 64,
      "chains": 32,
      "every": 1,
      "warmup_steps": 8,
      "warmup_seconds": 0.00212412799919548,
      "evaluation_steps": 60,
      "evaluation_seconds": 0.0234712360006597,
      "seconds_per_step": 0.0003911872666776617,
      "includes_CPU_staging_and_flush": true,
      "disk_bytes": 1013133,
      "writer_stats": {
        "buffer_bytes": 0,
        "peak_buffer_bytes": 131136,
        "max_buffer_bytes": 524288,
        "committed_records": 60
      }
    },
    {
      "kind": "recording",
      "mode": "async",
      "dimension": 64,
      "chains": 32,
      "every": 1,
      "warmup_steps": 8,
      "warmup_seconds": 0.0017331379995084717,
      "evaluation_steps": 60,
      "evaluation_seconds": 0.02710519399988698,
      "seconds_per_step": 0.00045175323333144964,
      "includes_CPU_staging_and_flush": true,
      "disk_bytes": 1013133,
      "writer_stats": {
        "buffer_bytes": 0,
        "peak_buffer_bytes": 131136,
        "max_buffer_bytes": 524288,
        "committed_records": 60
      }
    },
    {
      "kind": "recording",
      "mode": "sparse",
      "dimension": 64,
      "chains": 32,
      "every": 10,
      "warmup_steps": 8,
      "warmup_seconds": 0.002716945999964082,
      "evaluation_steps": 60,
      "evaluation_seconds": 0.014084026999626076,
      "seconds_per_step": 0.00023473378332710125,
      "includes_CPU_staging_and_flush": true,
      "disk_bytes": 101935,
      "writer_stats": {
        "buffer_bytes": 0,
        "peak_buffer_bytes": 98352,
        "max_buffer_bytes": 524288,
        "committed_records": 6
      }
    }
  ],
  "functional_compile": {
    "first_call_including_compile_seconds": 4.910878255999705,
    "warmup_seconds": 0.00014932100020814687,
    "compiled_evaluation_seconds": 0.0006455869997807895,
    "eager_evaluation_seconds": 0.0011439649997555534,
    "steps": 60,
    "note": "Functional transition only: excludes potential, autograd and sampler snapshots"
  }
}