{
  "kind": "frozen_register_tiling_selection",
  "recorded_at": "2026-10-02T18:44:42.366590+00:00",
  "training_process_id": 81341,
  "training_seed": 1741,
  "validation_seed": 9473,
  "compiler_implementation_sha256": "f255b038cc297ff6ba053b68cc787f7e16d855c0b665ab82b8d5ea33c463f8be",
  "benchmark_source_sha256": "47a379ae8881b98ece7df1f415afed0c5f2b189f9459bf34a14198a22d112675",
  "baseline_export_sha256": "ddda22c512e09d77a855c6cbbd49b6f1bac26d5db71197d8c05b0e4f0ff257f9",
  "tuning_report_sha256": "46e22bc7e34901d8ff4c3532a4c96fa0a171f84264065eb9a8d9c0a94e27dca5",
  "tuning_report": "/Users/andreslavescu/Documents/projects/meTile/benchmarks/results/m5-rmsnorm-tiling-tuning.json",
  "gate_policy": {
    "aligned_throughput_gpu_speedup_minimum": 1.1,
    "aligned_throughput_batches": [
      32,
      256
    ],
    "every_case_wall_latency_ratio_maximum": 1.03,
    "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
    "reference": "frozen prechange static-width register4 striped kernel",
    "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
  },
  "precision_comparison": {
    "class": "same_storage_precision",
    "same_weight_representation": true,
    "bitwise_exact": false,
    "primary_policy": "identical storage dtype, FP32 sum/normalization/weight multiply, final output cast",
    "primary_mlx_comparator": "mx.compile explicit full-FP32 graph then final output cast",
    "mlx_fast_caveat": "secondary only: FP16 normalization rounds to storage before weight multiply; checked against its own rounding reference",
    "mlx_fast_source": "https://github.com/ml-explore/mlx/blob/v0.32.0/mlx/backend/metal/kernels/rms_norm.metal",
    "relaxed_precision": false
  },
  "cases": [
    {
      "name": "float16_1x1009",
      "dtype": "float16",
      "batches": 1,
      "width": 1009
    },
    {
      "name": "float16_32x1009",
      "dtype": "float16",
      "batches": 32,
      "width": 1009
    },
    {
      "name": "float16_256x1009",
      "dtype": "float16",
      "batches": 256,
      "width": 1009
    },
    {
      "name": "float16_1x1024",
      "dtype": "float16",
      "batches": 1,
      "width": 1024
    },
    {
      "name": "float16_32x1024",
      "dtype": "float16",
      "batches": 32,
      "width": 1024
    },
    {
      "name": "float16_256x1024",
      "dtype": "float16",
      "batches": 256,
      "width": 1024
    },
    {
      "name": "float32_1x1009",
      "dtype": "float32",
      "batches": 1,
      "width": 1009
    },
    {
      "name": "float32_32x1009",
      "dtype": "float32",
      "batches": 32,
      "width": 1009
    },
    {
      "name": "float32_256x1009",
      "dtype": "float32",
      "batches": 256,
      "width": 1009
    },
    {
      "name": "float32_1x1024",
      "dtype": "float32",
      "batches": 1,
      "width": 1024
    },
    {
      "name": "float32_32x1024",
      "dtype": "float32",
      "batches": 32,
      "width": 1024
    },
    {
      "name": "float32_256x1024",
      "dtype": "float32",
      "batches": 256,
      "width": 1024
    }
  ],
  "variants": [
    {
      "name": "register4_striped",
      "elements_per_thread": 4,
      "layout": "striped"
    },
    {
      "name": "register4_blocked4",
      "elements_per_thread": 4,
      "layout": "blocked4"
    },
    {
      "name": "register8_striped",
      "elements_per_thread": 8,
      "layout": "striped"
    },
    {
      "name": "register8_blocked4",
      "elements_per_thread": 8,
      "layout": "blocked4"
    },
    {
      "name": "register16_striped",
      "elements_per_thread": 16,
      "layout": "striped"
    },
    {
      "name": "register16_blocked4",
      "elements_per_thread": 16,
      "layout": "blocked4"
    },
    {
      "name": "register32_striped",
      "elements_per_thread": 32,
      "layout": "striped"
    },
    {
      "name": "register32_blocked4",
      "elements_per_thread": 32,
      "layout": "blocked4"
    },
    {
      "name": "register2_striped",
      "elements_per_thread": 2,
      "layout": "striped"
    }
  ],
  "selection": {
    "scope": "one layout/count policy for every dtype, width and batch size",
    "rule": "minimize training wall/ragged guard failures, then maximize worst aligned batch32/256 GPU speedup, then its geometric mean; ties follow declared variant order",
    "selected": {
      "name": "register4_striped",
      "elements_per_thread": 4,
      "layout": "striped"
    },
    "scores": [
      {
        "variant": {
          "name": "register4_striped",
          "elements_per_thread": 4,
          "layout": "striped"
        },
        "eligible": true,
        "regression_guard_failures": 0,
        "minimum_aligned_throughput_gpu_speedup": 1.0086978283172867,
        "geomean_aligned_throughput_gpu_speedup": 1.045561445420347,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": []
        }
      },
      {
        "variant": {
          "name": "register4_blocked4",
          "elements_per_thread": 4,
          "layout": "blocked4"
        },
        "eligible": true,
        "regression_guard_failures": 6,
        "minimum_aligned_throughput_gpu_speedup": 0.9237474428973753,
        "geomean_aligned_throughput_gpu_speedup": 0.9718185576864061,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": [
            "float16_1x1009",
            "float16_32x1009",
            "float16_256x1009",
            "float32_1x1009",
            "float32_32x1009",
            "float32_256x1009"
          ]
        }
      },
      {
        "variant": {
          "name": "register8_striped",
          "elements_per_thread": 8,
          "layout": "striped"
        },
        "eligible": true,
        "regression_guard_failures": 2,
        "minimum_aligned_throughput_gpu_speedup": 0.9882831778336025,
        "geomean_aligned_throughput_gpu_speedup": 1.0321442986083251,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": [
            "float16_1x1009",
            "float32_1x1009"
          ]
        }
      },
      {
        "variant": {
          "name": "register8_blocked4",
          "elements_per_thread": 8,
          "layout": "blocked4"
        },
        "eligible": true,
        "regression_guard_failures": 6,
        "minimum_aligned_throughput_gpu_speedup": 0.8814602420184451,
        "geomean_aligned_throughput_gpu_speedup": 0.9208678007594686,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": [
            "float16_1x1009",
            "float16_32x1009",
            "float16_256x1009",
            "float32_1x1009",
            "float32_32x1009",
            "float32_256x1009"
          ]
        }
      },
      {
        "variant": {
          "name": "register16_striped",
          "elements_per_thread": 16,
          "layout": "striped"
        },
        "eligible": true,
        "regression_guard_failures": 4,
        "minimum_aligned_throughput_gpu_speedup": 0.9542795309833677,
        "geomean_aligned_throughput_gpu_speedup": 1.00466326240963,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": [
            "float16_1x1009",
            "float16_32x1009",
            "float32_1x1009",
            "float32_32x1009"
          ]
        }
      },
      {
        "variant": {
          "name": "register16_blocked4",
          "elements_per_thread": 16,
          "layout": "blocked4"
        },
        "eligible": true,
        "regression_guard_failures": 6,
        "minimum_aligned_throughput_gpu_speedup": 0.766666666588103,
        "geomean_aligned_throughput_gpu_speedup": 0.8232797524681654,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": [
            "float16_1x1009",
            "float16_32x1009",
            "float16_256x1009",
            "float32_1x1009",
            "float32_32x1009",
            "float32_256x1009"
          ]
        }
      },
      {
        "variant": {
          "name": "register32_striped",
          "elements_per_thread": 32,
          "layout": "striped"
        },
        "eligible": true,
        "regression_guard_failures": 4,
        "minimum_aligned_throughput_gpu_speedup": 0.8381200151101741,
        "geomean_aligned_throughput_gpu_speedup": 0.9307341209660918,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": [
            "float16_1x1009",
            "float16_32x1009",
            "float32_1x1009",
            "float32_32x1009"
          ]
        }
      },
      {
        "variant": {
          "name": "register32_blocked4",
          "elements_per_thread": 32,
          "layout": "blocked4"
        },
        "eligible": true,
        "regression_guard_failures": 8,
        "minimum_aligned_throughput_gpu_speedup": 0.5679061317298304,
        "geomean_aligned_throughput_gpu_speedup": 0.727493385335174,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [
            "float16_1x1009",
            "float32_1x1009"
          ],
          "failed_ragged_gpu_cases": [
            "float16_1x1009",
            "float16_32x1009",
            "float16_256x1009",
            "float32_1x1009",
            "float32_32x1009",
            "float32_256x1009"
          ]
        }
      },
      {
        "variant": {
          "name": "register2_striped",
          "elements_per_thread": 2,
          "layout": "striped"
        },
        "eligible": true,
        "regression_guard_failures": 1,
        "minimum_aligned_throughput_gpu_speedup": 1.017094119129839,
        "geomean_aligned_throughput_gpu_speedup": 1.0256124991977755,
        "training_gate": {
          "passed": false,
          "policy": {
            "aligned_throughput_gpu_speedup_minimum": 1.1,
            "aligned_throughput_batches": [
              32,
              256
            ],
            "every_case_wall_latency_ratio_maximum": 1.03,
            "every_ragged_case_gpu_latency_ratio_maximum": 1.03,
            "reference": "frozen prechange static-width register4 striped kernel",
            "promotion": "opt-in unless every heldout guard passes; no automatic default changes"
          },
          "failed_aligned_throughput_gpu_cases": [
            "float16_32x1024",
            "float16_256x1024",
            "float32_32x1024",
            "float32_256x1024"
          ],
          "failed_wall_cases": [],
          "failed_ragged_gpu_cases": [
            "float16_256x1009"
          ]
        }
      }
    ],
    "heldout_results_used": false
  }
}