{
  "schema_version": 1,
  "protocol": {
    "scope": "complete recurrent block forward, input/parameter backward and optimizer update",
    "optimizer": "torch.optim.AdamW(fused=True), states initialized before timing",
    "dtype": "BF16 autocast, FP32 parameters and recurrent carry",
    "order": "AB/BA",
    "compile_and_optimizer_state_initialization_excluded": true
  },
  "environment": {
    "torch": "2.1.2+cu118",
    "cuda": "11.8",
    "gpu": "NVIDIA H800 PCIe"
  },
  "cases": [
    {
      "case": "state2048_short",
      "batch": 4,
      "length": 2048,
      "config": {
        "width": 2048,
        "rnn_width": 2048,
        "depth": 1
      },
      "rows": [
        {
          "architecture": "rglru",
          "backend": "fused_grouped_prefix32_hierarchical_chunk16",
          "controller": null,
          "optimizer_step": {
            "samples_ms": [
              24.802595376968384,
              24.760689586400986,
              24.909019470214844,
              24.78039637207985,
              24.81362223625183,
              24.837687611579895,
              24.88834038376808,
              24.7415229678154,
              25.14628693461418,
              24.76516366004944
            ],
            "median_ms": 24.808108806610107,
            "mean_ms": 24.84453245997429,
            "std_ms": 0.11311406258685083,
            "mad_ms": 0.045182183384895325,
            "minimum_ms": 24.7415229678154,
            "maximum_ms": 25.14628693461418
          },
          "baseline_allocated_bytes": 1550969856,
          "baseline_reserved_bytes": 1606418432,
          "peak_allocated_bytes": 3240378880,
          "peak_reserved_bytes": 3533701120,
          "peak_allocated_delta_bytes": 1689409024,
          "peak_reserved_delta_bytes": 1927282688,
          "parameters": 50896896
        },
        {
          "architecture": "samu",
          "backend": "fused_output_fused_write_shared_sfu_chunk32",
          "controller": "fp32_fused_coords",
          "optimizer_step": {
            "samples_ms": [
              23.90008047223091,
              23.812387138605118,
              27.717124670743942,
              23.708980530500412,
              23.75074476003647,
              23.745745420455933,
              23.709382861852646,
              23.682843893766403,
              23.697882890701294,
              23.701224476099014
            ],
            "median_ms": 23.72756414115429,
            "mean_ms": 24.142639711499214,
            "std_ms": 1.1931383498276134,
            "mad_ms": 0.028010457754135132,
            "minimum_ms": 23.682843893766403,
            "maximum_ms": 27.717124670743942
          },
          "baseline_allocated_bytes": 1553065472,
          "baseline_reserved_bytes": 1608515584,
          "peak_allocated_bytes": 3107467776,
          "peak_reserved_bytes": 3391094784,
          "peak_allocated_delta_bytes": 1554402304,
          "peak_reserved_delta_bytes": 1782579200,
          "parameters": 50372612
        }
      ]
    },
    {
      "case": "state2560_medium",
      "batch": 1,
      "length": 8192,
      "config": {
        "width": 2560,
        "rnn_width": 2560,
        "depth": 1
      },
      "rows": [
        {
          "architecture": "rglru",
          "backend": "fused_grouped_prefix32_hierarchical_chunk16",
          "controller": null,
          "optimizer_step": {
            "samples_ms": [
              34.89868715405464,
              34.41610559821129,
              34.375760704278946,
              34.394122660160065,
              34.395188093185425,
              34.39483046531677,
              34.394003450870514,
              34.393176436424255,
              34.232962876558304,
              34.20514240860939
            ],
            "median_ms": 34.39406305551529,
            "mean_ms": 34.40999798476696,
            "std_ms": 0.17736235146927282,
            "mad_ms": 0.009713694453239441,
            "minimum_ms": 34.20514240860939,
            "maximum_ms": 34.89868715405464
          },
          "baseline_allocated_bytes": 2368228352,
          "baseline_reserved_bytes": 2975858688,
          "peak_allocated_bytes": 4454044672,
          "peak_reserved_bytes": 5073010688,
          "peak_allocated_delta_bytes": 2085816320,
          "peak_reserved_delta_bytes": 2097152000,
          "parameters": 79513600
        },
        {
          "architecture": "samu",
          "backend": "fused_output_fused_write_shared_sfu_grouped_prefix64_hierarchical_chunk32",
          "controller": "fp32_fused_coords",
          "optimizer_step": {
            "samples_ms": [
              33.11343118548393,
              33.14533457159996,
              33.158283680677414,
              33.10021385550499,
              33.10203552246094,
              33.1009104847908,
              33.08708593249321,
              33.04758295416832,
              33.05135667324066,
              33.07086229324341
            ],
            "median_ms": 33.100562170147896,
            "mean_ms": 33.09770971536636,
            "std_ms": 0.03422670100505065,
            "mad_ms": 0.021588057279586792,
            "minimum_ms": 33.04758295416832,
            "maximum_ms": 33.158283680677414
          },
          "baseline_allocated_bytes": 2372733440,
          "baseline_reserved_bytes": 2894069760,
          "peak_allocated_bytes": 4331364352,
          "peak_reserved_bytes": 4987027456,
          "peak_allocated_delta_bytes": 1958630912,
          "peak_reserved_delta_bytes": 2092957696,
          "parameters": 78694404
        }
      ]
    },
    {
      "case": "400m_block_long",
      "batch": 1,
      "length": 32768,
      "config": {
        "width": 1536,
        "rnn_width": 2048,
        "depth": 12
      },
      "rows": [
        {
          "architecture": "rglru",
          "backend": "fused_grouped_prefix32_hierarchical_chunk16",
          "controller": null,
          "optimizer_step": {
            "samples_ms": [
              49.09336939454079,
              49.62151125073433,
              48.90861362218857,
              49.49595034122467,
              48.72525855898857,
              48.7661212682724,
              49.08457025885582,
              48.89741539955139,
              49.16488751769066,
              48.675503581762314
            ],
            "median_ms": 48.996591940522194,
            "mean_ms": 49.04332011938095,
            "std_ms": 0.3018094814017261,
            "mad_ms": 0.19938312470912933,
            "minimum_ms": 48.675503581762314,
            "maximum_ms": 49.62151125073433
          },
          "baseline_allocated_bytes": 1137675264,
          "baseline_reserved_bytes": 2315255808,
          "peak_allocated_bytes": 6113490944,
          "peak_reserved_bytes": 7251951616,
          "peak_allocated_delta_bytes": 4975815680,
          "peak_reserved_delta_bytes": 4936695808,
          "parameters": 31230976
        },
        {
          "architecture": "samu",
          "backend": "fused_output_fused_write_shared_sfu_grouped_prefix64_hierarchical_chunk32",
          "controller": "fp32_fused_coords",
          "optimizer_step": {
            "samples_ms": [
              45.50177976489067,
              46.13611847162247,
              45.72005197405815,
              45.22400349378586,
              45.79009488224983,
              46.27338796854019,
              45.80125957727432,
              45.38346827030182,
              45.82751542329788,
              45.99413648247719
            ],
            "median_ms": 45.79567722976208,
            "mean_ms": 45.76518163084984,
            "std_ms": 0.31049831403069067,
            "mad_ms": 0.24617835879325867,
            "minimum_ms": 45.22400349378586,
            "maximum_ms": 46.27338796854019
          },
          "baseline_allocated_bytes": 1140831744,
          "baseline_reserved_bytes": 1914699776,
          "peak_allocated_bytes": 5713881088,
          "peak_reserved_bytes": 6916407296,
          "peak_allocated_delta_bytes": 4573049344,
          "peak_reserved_delta_bytes": 5001707520,
          "parameters": 30706692
        }
      ]
    }
  ]
}