{
  "status": "ok",
  "meta": {
    "device": "cuda",
    "dtype": "torch.float16",
    "hidden": 2048,
    "hidden_effective": 2048,
    "warmup": 20,
    "iters": 120,
    "bucket_lengths": [
      256
    ],
    "batch_sizes": [
      64
    ],
    "torch_compile": false,
    "compile_mode": "reduce-overhead",
    "compile_backend": "inductor",
    "inductor_gemm_backends": "",
    "inductor_gemm_backends_applied": false,
    "allow_tf32": true,
    "inference_mode": true,
    "use_inplace_residual": true,
    "pad_hidden_to": 16,
    "use_cuda_graph_arena": true,
    "single_shape_seq_len": 0,
    "skip_batch_scaling": false,
    "single_pass_after_warmup": false
  },
  "global_counters": {
    "compile_calls": 0,
    "graph_build_calls": 1
  },
  "bucket_results": [
    {
      "batch": 64,
      "seq_len": 256,
      "samples_ms": [
        14.9099,
        15.153,
        15.537,
        15.7995,
        15.8008,
        15.8105,
        15.8067,
        15.8097,
        15.8025,
        15.8218,
        15.8113,
        15.809,
        15.81,
        15.8152,
        15.8105,
        15.8123,
        15.8108,
        15.8112,
        15.8104,
        15.8092,
        15.8157,
        15.7958,
        15.6983,
        15.6966,
        15.7024,
        15.6954,
        15.7004,
        15.6888,
        15.6886,
        15.6876,
        15.6914,
        15.6839,
        15.6818,
        15.6925,
        18.9786,
        15.7139,
        15.5959,
        15.8143,
        15.8022,
        15.8089,
        15.806,
        15.8112,
        15.8148,
        15.8161,
        15.817,
        15.8087,
        15.8152,
        15.8182,
        15.8115,
        15.8147,
        15.8067,
        15.8124,
        15.9246,
        16.0962,
        16.3503,
        16.538,
        16.434,
        16.4293,
        16.4336,
        16.433,
        16.3861,
        16.291,
        16.1799,
        16.181,
        16.1916,
        16.1262,
        16.06,
        16.0658,
        15.9387,
        15.9355,
        15.9392,
        15.9337,
        15.9323,
        15.9331,
        15.9343,
        15.9416,
        15.9357,
        15.932,
        15.9377,
        15.9313,
        15.8861,
        15.8132,
        15.8096,
        15.8156,
        15.967,
        15.8442,
        15.8115,
        15.8161,
        15.8158,
        15.8104,
        15.8173,
        15.8147,
        15.8186,
        15.8188,
        15.8144,
        15.8162,
        15.8106,
        15.8126,
        15.817,
        15.8115,
        15.8118,
        15.8122,
        15.8162,
        15.8125,
        15.814,
        15.8102,
        15.9661,
        15.8109,
        15.8138,
        15.8106,
        15.8284,
        15.9326,
        16.0391,
        23.9481,
        16.0619,
        15.7742,
        15.7795,
        15.9328,
        16.049,
        16.1554
      ],
      "latency_ms": {
        "mean": 15.9618,
        "p50": 15.8147,
        "p95": 16.4293,
        "min": 14.9099,
        "max": 23.9481
      },
      "compile_calls_during_steady_state": 0,
      "graph_build_calls_during_steady_state": 0,
      "memory_pre": {
        "allocated_bytes": 151294976,
        "reserved_bytes": 719323136,
        "peak_allocated_bytes": 688165888,
        "allocation_events": 221,
        "segment_alloc_events": 12
      },
      "memory_post": {
        "allocated_bytes": 151286784,
        "reserved_bytes": 719323136,
        "peak_allocated_bytes": 688165888,
        "allocation_events": 221,
        "segment_alloc_events": 12
      },
      "allocation_event_delta": 0,
      "segment_alloc_event_delta": 0,
      "memory_stable": true
    }
  ],
  "batch_scaling": [
    {
      "batch": 64,
      "seq_len": 256,
      "samples_ms": [
        15.0014,
        14.9862,
        15.7169,
        16.0168,
        16.0235,
        16.0221,
        16.0221,
        16.0242,
        16.0267,
        15.9562,
        15.9231,
        15.9287,
        15.9207,
        15.9333,
        15.9271,
        15.93,
        15.9293,
        16.029,
        15.9291,
        15.9276,
        15.9328,
        15.8071,
        15.8025,
        15.7992,
        15.8019,
        15.8027,
        15.7954,
        15.8011,
        15.8048,
        15.8027
      ],
      "latency_ms": {
        "mean": 15.8441,
        "p50": 15.9271,
        "p95": 16.0267,
        "min": 14.9862,
        "max": 16.029
      },
      "compile_calls_during_steady_state": 0,
      "graph_build_calls_during_steady_state": 0,
      "memory_pre": {
        "allocated_bytes": 151294976,
        "reserved_bytes": 719323136,
        "peak_allocated_bytes": 688165888,
        "allocation_events": 333,
        "segment_alloc_events": 18
      },
      "memory_post": {
        "allocated_bytes": 151286784,
        "reserved_bytes": 719323136,
        "peak_allocated_bytes": 688165888,
        "allocation_events": 333,
        "segment_alloc_events": 18
      },
      "allocation_event_delta": 0,
      "segment_alloc_event_delta": 0,
      "memory_stable": true
    }
  ]
}