{
  "execution_mode": "CPU algorithm with controlled inputs; any latency/capacity coefficients are assumptions",
  "verification_step": {
    "gamma": 4,
    "accepted_draft_tokens": 2,
    "emitted_tokens": [
      0,
      1,
      1
    ],
    "total_emitted": 3,
    "trace": [
      {
        "position": 0,
        "draft_token": 0,
        "p_target": 0.7,
        "q_draft": 0.65,
        "accept_prob": 1.0,
        "u_draw": 0.42,
        "outcome": "accepted",
        "emitted_token": 0
      },
      {
        "position": 1,
        "draft_token": 1,
        "p_target": 0.68,
        "q_draft": 0.7,
        "accept_prob": 0.9714,
        "u_draw": 0.51,
        "outcome": "accepted",
        "emitted_token": 1
      },
      {
        "position": 2,
        "draft_token": 2,
        "p_target": 0.18,
        "q_draft": 0.6,
        "accept_prob": 0.3,
        "u_draw": 0.74,
        "outcome": "rejected_and_resampled",
        "emitted_token": 1
      }
    ]
  },
  "distribution_check_pos2": {
    "target_p": [
      0.1,
      0.48,
      0.18,
      0.24
    ],
    "draft_q": [
      0.05,
      0.15,
      0.6,
      0.2
    ],
    "residual_p_minus_q_norm": [
      0.119,
      0.7857,
      0.0,
      0.0952
    ],
    "tv_distance_with_residual": 0.0,
    "tv_distance_without_residual": 0.1364
  },
  "batch_crossover": [
    {
      "batch_size": 1,
      "verify_batched_tokens": 5,
      "standard_decode_ms_per_token": 22.0,
      "spec_cycle_ms": 32.4,
      "alpha_80_expected_tokens": 3.3616,
      "alpha_80_ms_per_token": 9.64,
      "alpha_80_speedup": 2.28,
      "alpha_45_expected_tokens": 1.7846,
      "alpha_45_ms_per_token": 18.16,
      "alpha_45_speedup": 1.21
    },
    {
      "batch_size": 2,
      "verify_batched_tokens": 10,
      "standard_decode_ms_per_token": 22.0,
      "spec_cycle_ms": 32.4,
      "alpha_80_expected_tokens": 3.3616,
      "alpha_80_ms_per_token": 9.64,
      "alpha_80_speedup": 2.28,
      "alpha_45_expected_tokens": 1.7846,
      "alpha_45_ms_per_token": 18.16,
      "alpha_45_speedup": 1.21
    },
    {
      "batch_size": 4,
      "verify_batched_tokens": 20,
      "standard_decode_ms_per_token": 22.0,
      "spec_cycle_ms": 32.4,
      "alpha_80_expected_tokens": 3.3616,
      "alpha_80_ms_per_token": 9.64,
      "alpha_80_speedup": 2.28,
      "alpha_45_expected_tokens": 1.7846,
      "alpha_45_ms_per_token": 18.16,
      "alpha_45_speedup": 1.21
    },
    {
      "batch_size": 8,
      "verify_batched_tokens": 40,
      "standard_decode_ms_per_token": 22.0,
      "spec_cycle_ms": 35.76,
      "alpha_80_expected_tokens": 3.3616,
      "alpha_80_ms_per_token": 10.64,
      "alpha_80_speedup": 2.07,
      "alpha_45_expected_tokens": 1.7846,
      "alpha_45_ms_per_token": 20.04,
      "alpha_45_speedup": 1.1
    },
    {
      "batch_size": 16,
      "verify_batched_tokens": 80,
      "standard_decode_ms_per_token": 22.0,
      "spec_cycle_ms": 52.56,
      "alpha_80_expected_tokens": 3.3616,
      "alpha_80_ms_per_token": 15.64,
      "alpha_80_speedup": 1.41,
      "alpha_45_expected_tokens": 1.7846,
      "alpha_45_ms_per_token": 29.45,
      "alpha_45_speedup": 0.75
    },
    {
      "batch_size": 32,
      "verify_batched_tokens": 160,
      "standard_decode_ms_per_token": 22.0,
      "spec_cycle_ms": 86.16,
      "alpha_80_expected_tokens": 3.3616,
      "alpha_80_ms_per_token": 25.63,
      "alpha_80_speedup": 0.86,
      "alpha_45_expected_tokens": 1.7846,
      "alpha_45_ms_per_token": 48.28,
      "alpha_45_speedup": 0.46
    },
    {
      "batch_size": 64,
      "verify_batched_tokens": 320,
      "standard_decode_ms_per_token": 35.44,
      "spec_cycle_ms": 159.12,
      "alpha_80_expected_tokens": 3.3616,
      "alpha_80_ms_per_token": 47.33,
      "alpha_80_speedup": 0.75,
      "alpha_45_expected_tokens": 1.7846,
      "alpha_45_ms_per_token": 89.16,
      "alpha_45_speedup": 0.4
    }
  ]
}