{
  "schema": 1,
  "status": "designed",
  "workload": {
    "name": "rms_scaled_softmax",
    "dtype": "float32",
    "columns": 131072,
    "primary_rows": 16,
    "main_rows": [
      1,
      16,
      32
    ],
    "capacity_rows": 64,
    "illustrative_tile_width": 8192,
    "validation_shapes": [
      [
        1,
        1
      ],
      [
        3,
        257
      ],
      [
        2,
        8193
      ],
      [
        2,
        131075
      ]
    ],
    "finite_inputs_only": true,
    "input_output_alias": false,
    "fixtures": [
      "random",
      "zero",
      "constant",
      "dominant",
      "alternating",
      "tile_offsets",
      "common_offset",
      "near_zero"
    ],
    "seeds": {
      "pilot": 11,
      "preflight": 23,
      "timing": 41,
      "held_out": 53
    },
    "eps": 1e-05,
    "maximum_absolute_input": 1024
  },
  "variants": [
    {
      "id": "baseline",
      "kernels": 9
    },
    {
      "id": "local_materialize",
      "kernels": 3
    },
    {
      "id": "local_recompute",
      "kernels": 3
    },
    {
      "id": "local_materialize2",
      "kernels": 2
    },
    {
      "id": "local_recompute2",
      "kernels": 2
    },
    {
      "id": "streaming",
      "kernels": 1
    },
    {
      "id": "retained",
      "kernels": "slabs"
    },
    {
      "id": "roundtrip_off",
      "kernels": "slabs"
    },
    {
      "id": "roundtrip_on",
      "kernels": "slabs"
    },
    {
      "id": "input_reload",
      "kernels": "slabs"
    },
    {
      "id": "both_reload",
      "kernels": "slabs"
    },
    {
      "id": "cudnn",
      "kernels": "measured"
    },
    {
      "id": "row_retained",
      "kernels": "cooperatively admitted slabs",
      "synchronization": "device-scope release/acquire row barriers"
    }
  ],
  "arithmetic": {
    "exp": "expf",
    "fast_math": false,
    "fmad": false
  },
  "tuning": {
    "tile_widths": [
      4096,
      8192,
      16384,
      131072
    ],
    "threads": [
      64,
      128,
      256
    ],
    "families": [
      "baseline",
      "local_materialize",
      "local_recompute",
      "local_materialize2",
      "local_recompute2",
      "streaming",
      "retained",
      "row_retained"
    ],
    "max_geometries_per_family_per_shape": 12,
    "rounds": 5,
    "batch_invocations": 16,
    "selection_metric": "median_graph_completion_per_invocation",
    "freeze_before_held_out_measurements": true,
    "scratch_bytes": 256,
    "shared_capacity_model": "Primary local kernels: 256 bytes; retained: 4*tile_width+256 bytes. No padded local control.",
    "cooperative_admission": "compiled occupancy API times queried SM count",
    "oversize_batch_policy": "sequential independent row slabs"
  },
  "logical_global_accesses": {
    "scope": "Dominant arrays only; logical accesses, not HBM traffic. Local2 and streaming repeat full-row work and are measured separately.",
    "baseline_bytes_per_element": 36,
    "materialize_bytes_per_element": 20,
    "recompute_bytes_per_element": 16,
    "retained_bytes_per_element": 8
  },
  "validation": {
    "oracle": "Independent float64 RMS scaling then stable row softmax on saved float32 inputs",
    "atol": 1e-07,
    "rtol": 2e-05,
    "row_sum_atol": 2e-05,
    "row_l1_max": 2e-05,
    "denominator_rtol": 2e-05,
    "changing_input_replays": 1000,
    "sanitizers": [
      "memcheck",
      "racecheck",
      "synccheck"
    ]
  },
  "timing": {
    "allocations": 2,
    "warmups": 10,
    "rounds": 21,
    "batch_invocations": 32,
    "metrics": [
      "host_single_completion_us",
      "host_batch_enqueue_us",
      "host_batch_completion_us",
      "cuda_event_batch_us"
    ],
    "main_cache_policy": "warm same-variant replay",
    "cache_eviction_rows": [
      16,
      64
    ],
    "eviction_buffer_bytes_minimum": 268435456,
    "eviction_l2_multiplier": 4
  },
  "proposed_collection": {
    "gpu": "H100!",
    "max_active_gpu_containers": 1,
    "automatic_retries": 0,
    "warm_pool": 0,
    "pilot_gpu_seconds_max": 240,
    "timing_gpu_seconds_max_per_allocation": 180,
    "evidence_gpu_seconds_max": 360,
    "total_gpu_function_seconds_max": 900,
    "include_failed_attempts_in_budget": true,
    "compile_and_package_on_cpu": true,
    "teardown_check_required": true
  },
  "prior_device_observation": {
    "scope": "capacity planning only; query the newly allocated device before launch",
    "source": "experiments/graphs_mlp/artifacts/20261019T233947Z/evidence/ncu-1024-fused_graph.metrics.csv",
    "sha256": "d23165b3c4ec6507a40ff0b7b57565e88c90fb585a2829145112081ef6d89872",
    "sm_count": 132,
    "shared_bytes_per_sm": 233472,
    "reserved_shared_bytes_per_block": 1024,
    "shared_bytes_per_block_optin": 232448,
    "registers_per_sm": 65536,
    "threads_per_sm": 2048,
    "blocks_per_sm": 32
  },
  "hypothesis": {
    "primary_rows": 16,
    "target_event_interval_speedup": 1.25,
    "allocations_required": 2,
    "not_a_completion_requirement": true,
    "primary_comparator": "pilot-fastest of all 3/2-kernel local families and redundant single-kernel streaming"
  },
  "execution_revision": {
    "parent_run": "20261020T022150Z",
    "reason": "Same RMS-scaled softmax; compare row-scoped barriers with full-grid barriers",
    "shared_gpu_seconds_cap": 900
  }
}