mtplx_runtime.json

{
  "arch_id": "qwen3-next-mtp",
  "artifact_role": "maximum-speed-flat4-candidate",
  "base_trunk": "mlx-community/Qwen3.6-27B-4bit",
  "exactness_baseline": {
    "attention_impl": "mlx_vector_paged",
    "context": 64,
    "gate": "phase0h-paged-verifier-exactness",
    "max_abs_diff": 0.0,
    "mode": "decode-from-stock-prefix",
    "sample_agreement": 1.0,
    "status": "passed",
    "topk_overlap_ratio": 1.0,
    "total_variation": 0.0,
    "verify_tokens": 4
  },
  "generic_vlm_loader_compatible": true,
  "mtp_depth_max": 3,
  "mtp_file": "mtp/weights.safetensors",
  "mtp_sidecar": "native MTP sidecar",
  "mtp_sidecar_file": "mtp/weights.safetensors",
  "mtplx_version": "0.1.0-preview",
  "recommended_draft_lm_head": {
    "bits": 3,
    "group_size": 64,
    "mode": "affine"
  },
  "recommended_draft_sampler": {
    "temperature": 0.7,
    "top_k": 20,
    "top_p": 0.95
  },
  "recommended_profile": "sustained",
  "sampler": {
    "temperature": 0.6,
    "top_k": 20,
    "top_p": 0.95
  },
  "speed_evidence": {
    "acceptance_by_depth": [
      0.9,
      0.7593,
      0.6476
    ],
    "depth": 3,
    "verdict": "mtp_depth_wins",
    "profile": "sustained",
    "forge_verify_rows": [
      {
        "acceptance_by_position": [],
        "depth": 0,
        "finish_reasons": {
          "stop": 1
        },
        "hit_token_budget": false,
        "hit_token_budget_count": 0,
        "multiplier_vs_ar": 1.0,
        "quality_passed": true,
        "tok_s": 13.66,
        "tok_s_decode": 14.49,
        "tok_s_e2e": 13.66,
        "verify_time_s": null
      },
      {
        "acceptance_by_position": [
          0.9233
        ],
        "depth": 1,
        "finish_reasons": {
          "stop": 1
        },
        "hit_token_budget": false,
        "hit_token_budget_count": 0,
        "multiplier_vs_ar": 1.4381,
        "quality_passed": true,
        "tok_s": 19.65,
        "tok_s_decode": 20.9,
        "tok_s_e2e": 19.65,
        "verify_time_s": null
      },
      {
        "acceptance_by_position": [
          0.9496,
          0.7526
        ],
        "depth": 2,
        "finish_reasons": {
          "stop": 1
        },
        "hit_token_budget": false,
        "hit_token_budget_count": 0,
        "multiplier_vs_ar": 1.4594,
        "quality_passed": true,
        "tok_s": 19.94,
        "tok_s_decode": 20.79,
        "tok_s_e2e": 19.94,
        "verify_time_s": null
      },
      {
        "acceptance_by_position": [
          0.9,
          0.7593,
          0.6476
        ],
        "depth": 3,
        "finish_reasons": {
          "stop": 1
        },
        "hit_token_budget": false,
        "hit_token_budget_count": 0,
        "multiplier_vs_ar": 1.3208,
        "quality_passed": true,
        "tok_s": 18.05,
        "tok_s_decode": 19.71,
        "tok_s_e2e": 18.05,
        "verify_time_s": null
      }
    ],
    "greedy_diagnostic": {
      "tok_s": 13.66
    },
    "tok_s": [
      18.05
    ],
    "draft_sampler": {
      "temperature": 0.7,
      "top_k": 20,
      "top_p": 0.95
    },
    "enable_thinking": false,
    "mtp_history_policy": "committed",
    "measurement_notes": "Re-measured 2026-07-18 on Apple M5 Pro 64GB, MTPLX 1.0.4, mtp_history_policy=committed (CLI default), draft_temperature=0.7 (model recommended_draft_sampler), enable_thinking=false, warm (4 warmup prompts), 8 measured prompts from calibration_coding suite, max_tokens=192, median tok/s. tok_s_e2e = generated_tokens/total_elapsed (wall-clock, includes prefill). tok_s_decode = decode-only. multiplier_vs_ar uses e2e. The prior baked numbers (acceptance 1.0/0.98/0.94, tok/s 63) were on Apple M5 Max 128GB and are not reproducible on M5 Pro 64GB due to lower memory bandwidth; preserved below as historical_m5max. Acceptance is stable across runs (greedy draft gives 0.87/0.74/0.64, recommended 0.7 draft gives 0.90/0.76/0.65 -- within noise).",
    "historical_m5max": {
      "acceptance_by_depth": [
        1.0,
        0.9795918367346939,
        0.9387755102040817
      ],
      "accepted_by_depth": [
        49,
        48,
        46
      ],
      "correction_tokens": 3,
      "depth": 3,
      "draft_lm_head": "3-bit affine group64 draft-only head",
      "draft_sampler": {
        "temperature": 0.7,
        "top_k": 20,
        "top_p": 0.95
      },
      "drafted_by_depth": [
        49,
        49,
        49
      ],
      "enable_thinking": false,
      "greedy_diagnostic": {
        "accepted_by_depth": [
          49,
          49,
          43
        ],
        "correction_tokens": 0,
        "draft_sampler": {
          "temperature": 0.0,
          "top_k": 0,
          "top_p": 1.0
        },
        "drafted_by_depth": [
          51,
          51,
          51
        ],
        "target_sampler": {
          "temperature": 0.0,
          "top_k": 0,
          "top_p": 1.0
        },
        "tok_s": 60.108346853627076,
        "verify_calls": 51
      },
      "older_3bit_tok_s": [
        60.03822277200243,
        60.06063543439426
      ],
      "previous_best_tok_s": [
        61.52744855840259,
        60.89966247416089
      ],
      "profile": "sustained",
      "timestamp": "2026-05-03T23:07:00+0100",
      "tok_s": [
        63.05651084048216,
        62.885564723941826
      ],
      "verify_calls": 49
    }
  },
  "verified_on": {
    "hardware": "Apple M5 Pro",
    "machine_arch": "arm64",
    "macos": "26.5.2",
    "model": "Qwen3.6-27B-MTPLX-Optimized-Speed-Vision",
    "timestamp": "2026-07-18T00:00:00+00:00"
  }
}