{
  "host": "H100 PCIe 80GB x2",
  "GPUs": 2,
  "concurrent_users": 1,
  "model": "GLM-5.3-Flash",
  "quantization": "GGUF UD-IQ3_XXS",
  "context_capacity": 40960,
  "generation_limit": 1024,
  "target_split": "layer 1:1",
  "MTP_on": {
    "spec_type": "draft-mtp",
    "max_draft_tokens": 3,
    "sampling": "probabilistic"
  },
  "MTP_off": "none",
  "groups": [
    {
      "input_tokens": 8193,
      "repetitions_each": 3,
      "off": {
        "actual_input_tokens": 8193,
        "repetitions": 3,
        "generation_tokens_per_second": {
          "all": [
            41.28226747667661,
            51.22410812255903,
            51.153848406658035
          ],
          "median": 51.153848406658035,
          "min": 41.28226747667661,
          "max": 51.22410812255903
        },
        "first_streamed_generated_text_seconds": {
          "all": [
            22.92088064807467,
            22.474210544954985,
            22.475983759970404
          ],
          "median": 22.475983759970404,
          "min": 22.474210544954985,
          "max": 22.92088064807467
        },
        "first_answer_text_seconds": {
          "all": [
            27.273759820032865,
            22.474210544954985,
            22.475983759970404
          ],
          "median": 22.475983759970404,
          "min": 22.474210544954985,
          "max": 27.273759820032865
        },
        "input_processing_seconds": {
          "all": [
            22.68219,
            22.430748,
            22.433433
          ],
          "median": 22.433433,
          "min": 22.430748,
          "max": 22.68219
        },
        "end_to_end_seconds": {
          "all": [
            41.30697966204025,
            41.69950583600439,
            40.37881925900001
          ],
          "median": 41.30697966204025,
          "min": 40.37881925900001,
          "max": 41.69950583600439
        },
        "capped_trials": [],
        "draft_acceptance_fraction": null
      },
      "on": {
        "actual_input_tokens": 8193,
        "repetitions": 3,
        "generation_tokens_per_second": {
          "all": [
            61.51962611014726,
            62.27434301329048,
            62.77850289778328
          ],
          "median": 62.27434301329048,
          "min": 61.51962611014726,
          "max": 62.77850289778328
        },
        "first_streamed_generated_text_seconds": {
          "all": [
            24.887945767026395,
            24.628569833934307,
            24.631712430040352
          ],
          "median": 24.631712430040352,
          "min": 24.628569833934307,
          "max": 24.887945767026395
        },
        "first_answer_text_seconds": {
          "all": [
            26.091206192970276,
            24.628569833934307,
            24.631712430040352
          ],
          "median": 24.631712430040352,
          "min": 24.628569833934307,
          "max": 26.091206192970276
        },
        "input_processing_seconds": {
          "all": [
            24.648375,
            24.562103999999998,
            24.565319
          ],
          "median": 24.565319,
          "min": 24.562103999999998,
          "max": 24.648375
        },
        "end_to_end_seconds": {
          "all": [
            39.339272454963066,
            41.010106497909874,
            40.88068579498213
          ],
          "median": 40.88068579498213,
          "min": 39.339272454963066,
          "max": 41.010106497909874
        },
        "capped_trials": [
          "ctx8192-r2",
          "ctx8192-r3"
        ],
        "draft_acceptance_fraction": {
          "all": [
            0.5265700483091788,
            0.5370212765957447,
            0.5474137931034483
          ],
          "median": 0.5370212765957447,
          "min": 0.5265700483091788,
          "max": 0.5474137931034483
        }
      },
      "generation_median_ratio": 1.217393118074476,
      "generation_median_change_percent": 21.739311807447592,
      "first_streamed_text_median_change_seconds": 2.155728670069948,
      "on_all_trial_means_above_50": true
    },
    {
      "input_tokens": 32768,
      "repetitions_each": 3,
      "off": {
        "actual_input_tokens": 32768,
        "repetitions": 3,
        "generation_tokens_per_second": {
          "all": [
            49.71899775054341,
            49.72888671075154,
            49.767699383269026
          ],
          "median": 49.72888671075154,
          "min": 49.71899775054341,
          "max": 49.767699383269026
        },
        "first_streamed_generated_text_seconds": {
          "all": [
            86.56187437300105,
            86.54652248800267,
            86.54690774192568
          ],
          "median": 86.54690774192568,
          "min": 86.54652248800267,
          "max": 86.56187437300105
        },
        "first_answer_text_seconds": {
          "all": [
            88.69535350007936,
            86.75351399497595,
            87.07444424391724
          ],
          "median": 87.07444424391724,
          "min": 86.75351399497595,
          "max": 88.69535350007936
        },
        "input_processing_seconds": {
          "all": [
            86.504044,
            86.488305,
            86.48918300000001
          ],
          "median": 86.48918300000001,
          "min": 86.488305,
          "max": 86.504044
        },
        "end_to_end_seconds": {
          "all": [
            107.13822314806748,
            103.05686300690286,
            101.47718459099997
          ],
          "median": 103.05686300690286,
          "min": 101.47718459099997,
          "max": 107.13822314806748
        },
        "capped_trials": [
          "ctx32768-r1"
        ],
        "draft_acceptance_fraction": null
      },
      "on": {
        "actual_input_tokens": 32768,
        "repetitions": 3,
        "generation_tokens_per_second": {
          "all": [
            55.7749175125499,
            53.98533985476238,
            57.01239519928313
          ],
          "median": 55.7749175125499,
          "min": 53.98533985476238,
          "max": 57.01239519928313
        },
        "first_streamed_generated_text_seconds": {
          "all": [
            94.31905742699746,
            94.38870209804736,
            94.3602513789665
          ],
          "median": 94.3602513789665,
          "min": 94.31905742699746,
          "max": 94.38870209804736
        },
        "first_answer_text_seconds": {
          "all": [
            96.27599957806524,
            95.75533016200643,
            94.7251296370523
          ],
          "median": 95.75533016200643,
          "min": 94.7251296370523,
          "max": 96.27599957806524
        },
        "input_processing_seconds": {
          "all": [
            94.26143300000001,
            94.330284,
            94.30095299999999
          ],
          "median": 94.30095299999999,
          "min": 94.26143300000001,
          "max": 94.330284
        },
        "end_to_end_seconds": {
          "all": [
            112.66147581802215,
            110.22721290506888,
            108.27052422100678
          ],
          "median": 110.22721290506888,
          "min": 108.27052422100678,
          "max": 112.66147581802215
        },
        "capped_trials": [
          "ctx32768-r1"
        ],
        "draft_acceptance_fraction": {
          "all": [
            0.5132450331125827,
            0.48799231508165225,
            0.5401540154015402
          ],
          "median": 0.5132450331125827,
          "min": 0.48799231508165225,
          "max": 0.5401540154015402
        }
      },
      "generation_median_ratio": 1.1215798543201891,
      "generation_median_change_percent": 12.157985432018915,
      "first_streamed_text_median_change_seconds": 7.813343637040816,
      "on_all_trial_means_above_50": true
    }
  ],
  "GPU_peak_MiB": {
    "off": {
      "0": 58033,
      "1": 56509
    },
    "on": {
      "0": 58273,
      "1": 59667
    }
  },
  "limits": [
    "GPU memory and speed measured on this H100 pair only; no RTX performance guarantee.",
    "Reported generation includes reasoning tokens; first streamed text and first answer text are separate.",
    "Trial-average >50 is not an instantaneous speed or service-level guarantee.",
    "MTP probabilistic sampling can produce different outputs even with the same seed.",
    "No removal of slow, capped, or qualitatively incorrect outputs.",
    "Fixed 3-draft comparison, not an exhaustive optimization search.",
    "Earlier stopped 1M/256K/layer/row campaigns are unchanged.",
    "ExLlamaV3 uses a different quantization and measurement interface; compare its MTP effect within its own paired conditions."
  ],
  "created_utc": "2026-10-30T08:48:13.521678+00:00",
  "source_comparison_sha256": "25ff1866a56061c3cf7978ea61d6fc4a5b608bdd39cd47a797d99735166ab64d",
  "quality_findings": [
    "MTP-off: one of six answers ended at1024-token cap; MTP-on: three of six. All retained.",
    "Technical inaccuracies include conflating TTFT/prefill and unsupported causal inference; no claim MTP caused these."
  ],
  "trial_average_not_instantaneous_rate": true,
  "generation_includes_reasoning_tokens": true
}