{
 "herkunft": {
  "datei": "/home/rulemaker/guardrail-eval/harness/runs/sweetspot-zellen.jsonl",
  "zellen": 112,
  "phasen": [
   "a0",
   "a1",
   "a2",
   "a4",
   "a5"
  ],
  "erfasst": [
   "2026-10-28T22:30:31+0100"
  ],
  "zuletzt": [
   "2026-10-30T10:47:41+0100"
  ]
 },
 "messgroessen": {
  "zeit_bis_erstes_token_s": "Median aus zwei Laeufen mit frischem Zufallsprompt und acht erzeugten Token; Zeit vom Absenden bis zum ersten Inhaltsteilchen",
  "prefill_tok_s": "Differenzgroesse: (lange Promptlaenge - mittlere Promptlaenge) / (Wartezeit lang - Wartezeit mittel); der fixe request-Aufwand faellt heraus",
  "decode_tok_s": "erzeugte Token geteilt durch die gesamte Streamingzeit, Wartezeit auf das erste Token inklusive (bewusste Unterkante)",
  "rest_durchsatz_tok_s": "zwei bis acht kurze Anfragen (256 Token) gleichzeitig in dieselbe Warteschlange; erzeugte Token gesamt / Wandzeit",
  "prompt_token": "von der Serverantwort gemeldete Promptgroesse (nicht vom Messer geschaetzt)",
  "vram_max_mb": "hoechster in der Telemetrie gesehener Wert einer Karte, Fuenf-Sekunden-Raster",
  "nadeltest": "Codewort mit Zahl in halber und oberster Tiefe des Kontexts, je Zelle frischer Vollkontext-Prompt, Temperatur 0",
  "antwortzeit_s": "Zeit bis zur vollstaendigen Antwort des Nadeltests, nicht bis zum ersten Token"
 },
 "spalten": [
  "phase",
  "plattform",
  "quant",
  "kv_stufe",
  "kv_resident_mib",
  "kontext",
  "sessions",
  "rope",
  "zeit",
  "prompt_token",
  "zeit_bis_erstes_token_s",
  "prefill_tok_s",
  "decode_tok_s",
  "rest_durchsatz_tok_s",
  "vram_max_mb",
  "nadel_treffer",
  "nadel_versuche",
  "nadel_prompt_median",
  "nadel_antwortzeit_s"
 ],
 "teil": "2 (Rope-Stretching)",
 "beschraenkungen": [
  "Je Zelle ein Lauf der Zeitsonde; die Mediane der Tempospalten beruhen auf n=1 je Konfiguration",
  "Nadeltest bei gestrecktem Kontext mit zwei Tiefen (50 und 90 Prozent), beim nativen 262k-Vorlauf mit drei",
  "12 von 24 Sonden mussten nach einer HTTP-400-Ablehnung verkleinert werden; die Prompts landeten knapp unter dem Fenster",
  "Vollpass-Rate = Promptgroesse geteilt durch Zeit bis zur vollstaendigen Antwort (Denken aus, maximal 32 Antworttoken)",
  "Der KV-Speicherbedarf ist aus den Formatgroessen gerechnet, nicht gemessen",
  "GB10-Phase war auf das native Fenster geplant; 512k und 1 Mio sind dort nicht gemessen",
  "Die zwei k8v4-Doppelzellen bei 512k sind Einzelfaelle ohne Gegenversuch, die Ursache ist ungeklaert"
 ],
 "zellzeilen": [
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "int8",
   20480,
   1048576,
   1,
   "yarn x4",
   "2026-10-29T19:26:54+0100",
   260470,
   131.55,
   2029.2,
   47.6,
   null,
   22217,
   2,
   2,
   969935,
   813.5
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "int8",
   20480,
   524288,
   1,
   "yarn x2",
   "2026-10-30T02:27:14+0100",
   260470,
   127.45,
   2092.0,
   55.4,
   null,
   22201,
   2,
   2,
   484915,
   324.6
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "int8",
   20480,
   524288,
   2,
   "yarn x2",
   "2026-10-29T17:29:41+0100",
   260470,
   127.4,
   2092.1,
   54.9,
   77.4,
   22187,
   null,
   null,
   null,
   null
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "int8",
   65536,
   1048576,
   1,
   "yarn x4",
   "2026-10-29T20:52:20+0100",
   260470,
   132.94,
   2013.5,
   49.0,
   null,
   22225,
   2,
   2,
   969632,
   824.4
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "int8",
   65536,
   524288,
   1,
   "yarn x2",
   "2026-10-30T03:00:00+0100",
   260470,
   155.95,
   1606.9,
   54.1,
   null,
   22203,
   2,
   2,
   484816,
   324.6
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "int8",
   65536,
   524288,
   2,
   "yarn x2",
   "2026-10-29T17:34:34+0100",
   260470,
   128.88,
   2068.1,
   54.3,
   76.1,
   22189,
   null,
   null,
   null,
   null
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "k8v4",
   20480,
   1048576,
   1,
   "yarn x4",
   "2026-10-29T22:05:28+0100",
   260470,
   137.5,
   1947.8,
   50.4,
   null,
   22209,
   2,
   2,
   970535,
   824.5
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "k8v4",
   20480,
   524288,
   1,
   "yarn x2",
   "2026-10-30T03:34:45+0100",
   260470,
   133.98,
   1998.9,
   49.5,
   null,
   22189,
   2,
   2,
   523364,
   365.1
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "k8v4",
   20480,
   524288,
   2,
   "yarn x2",
   "2026-10-30T03:47:35+0100",
   260470,
   175.18,
   1854.3,
   12.8,
   16.5,
   22203,
   null,
   null,
   null,
   null
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "k8v4",
   65536,
   1048576,
   1,
   "yarn x4",
   "2026-10-29T23:23:23+0100",
   260470,
   138.6,
   1933.0,
   45.6,
   null,
   22207,
   2,
   2,
   1047411,
   919.0
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "k8v4",
   65536,
   524288,
   1,
   "yarn x2",
   "2026-10-29T16:23:06+0100",
   260470,
   161.95,
   1556.6,
   38.6,
   null,
   22203,
   2,
   2,
   484636,
   348.4
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "k8v4",
   65536,
   524288,
   2,
   "yarn x2",
   "2026-10-30T03:58:22+0100",
   260470,
   193.83,
   1444.9,
   8.7,
   9.7,
   22211,
   null,
   null,
   null,
   null
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   20480,
   1048576,
   1,
   "yarn x4",
   "2026-10-30T00:33:23+0100",
   260470,
   158.37,
   1595.4,
   50.2,
   null,
   22205,
   2,
   2,
   970201,
   778.7
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   20480,
   1048576,
   2,
   "yarn x4",
   "2026-10-30T01:49:16+0100",
   260470,
   131.67,
   2042.4,
   47.8,
   76.2,
   22191,
   null,
   null,
   null,
   null
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   20480,
   524288,
   1,
   "yarn x2",
   "2026-10-29T16:51:52+0100",
   260470,
   128.99,
   2082.6,
   45.0,
   null,
   22201,
   2,
   2,
   523795,
   347.9
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   20480,
   524288,
   2,
   "yarn x2",
   "2026-10-30T04:03:30+0100",
   260470,
   131.48,
   2043.3,
   46.8,
   70.9,
   22189,
   null,
   null,
   null,
   null
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   65536,
   1048576,
   1,
   "yarn x4",
   "2026-10-30T01:44:03+0100",
   260470,
   160.56,
   1588.9,
   50.7,
   null,
   22207,
   2,
   2,
   969377,
   789.0
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   65536,
   1048576,
   2,
   "yarn x4",
   "2026-10-30T01:55:08+0100",
   260470,
   132.53,
   2028.3,
   47.4,
   72.1,
   22207,
   null,
   null,
   null,
   null
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   65536,
   524288,
   1,
   "yarn x2",
   "2026-10-29T17:24:17+0100",
   260470,
   129.68,
   2071.0,
   49.4,
   null,
   22201,
   2,
   2,
   523759,
   419.3
  ],
  [
   "a2",
   "3090-strata",
   "IQ3_S",
   "q4_0",
   65536,
   524288,
   2,
   "yarn x2",
   "2026-10-30T04:08:39+0100",
   260470,
   130.1,
   2063.6,
   47.8,
   73.9,
   22199,
   null,
   null,
   null,
   null
  ],
  [
   "a4",
   "gb10-llama",
   "IQ3_S",
   "int8",
   1,
   262144,
   1,
   "none",
   "2026-10-29T18:44:57+0100",
   null,
   null,
   null,
   null,
   null,
   null,
   null,
   null,
   null,
   null
  ],
  [
   "a4",
   "gb10-llama",
   "IQ3_S",
   "int8",
   1,
   262144,
   1,
   "none",
   "2026-10-29T20:34:03+0100",
   null,
   379.39,
   697.9,
   null,
   null,
   null,
   3,
   3,
   261175,
   471.6
  ],
  [
   "a4",
   "gb10-llama",
   "IQ3_S",
   "q4_0",
   1,
   262144,
   1,
   "none",
   "2026-10-29T18:46:35+0100",
   null,
   null,
   null,
   null,
   null,
   null,
   null,
   null,
   null,
   null
  ],
  [
   "a4",
   "gb10-llama",
   "IQ3_S",
   "q4_0",
   1,
   262144,
   1,
   "none",
   "2026-10-29T21:51:18+0100",
   null,
   404.58,
   640.6,
   null,
   null,
   null,
   3,
   3,
   261341,
   502.0
  ]
 ],
 "yarn_einzel": [
  {
   "kv": "int8",
   "resident": 1,
   "kontext": 262144,
   "sessions": 1,
   "phase": "a4",
   "rope": "none",
   "plattform": "gb10-llama",
   "n": 0,
   "ttft_frisch_32k_s": 379.39,
   "decode_toks": null,
   "parallel_tok_s": null,
   "prefill_tok_s": 697.9,
   "prompt_token_frisch_32k_median": null,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": null,
   "nadel": {
    "treffer": 3,
    "versuche": 3,
    "prompt_token_median": 261175,
    "antwortzeit_median_s": 471.6,
    "tiefen": [
     0.1,
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "int8",
   "resident": 20480,
   "kontext": 1048576,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 131.55,
   "decode_toks": 47.6,
   "parallel_tok_s": null,
   "prefill_tok_s": 2029.2,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22217,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 969935,
    "antwortzeit_median_s": 813.5,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "int8",
   "resident": 20480,
   "kontext": 524288,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 127.45,
   "decode_toks": 55.4,
   "parallel_tok_s": null,
   "prefill_tok_s": 2092.0,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22201,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 484915,
    "antwortzeit_median_s": 324.6,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "int8",
   "resident": 65536,
   "kontext": 1048576,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 132.94,
   "decode_toks": 49.0,
   "parallel_tok_s": null,
   "prefill_tok_s": 2013.5,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22225,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 969632,
    "antwortzeit_median_s": 824.4,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "int8",
   "resident": 65536,
   "kontext": 524288,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 155.95,
   "decode_toks": 54.1,
   "parallel_tok_s": null,
   "prefill_tok_s": 1606.9,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22203,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 484816,
    "antwortzeit_median_s": 324.6,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "k8v4",
   "resident": 20480,
   "kontext": 1048576,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 137.5,
   "decode_toks": 50.4,
   "parallel_tok_s": null,
   "prefill_tok_s": 1947.8,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22209,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 970535,
    "antwortzeit_median_s": 824.5,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "k8v4",
   "resident": 20480,
   "kontext": 524288,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 133.98,
   "decode_toks": 49.5,
   "parallel_tok_s": null,
   "prefill_tok_s": 1998.9,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22189,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 523364,
    "antwortzeit_median_s": 365.1,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "k8v4",
   "resident": 65536,
   "kontext": 1048576,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 138.6,
   "decode_toks": 45.6,
   "parallel_tok_s": null,
   "prefill_tok_s": 1933.0,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22207,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 1047411,
    "antwortzeit_median_s": 919.0,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "k8v4",
   "resident": 65536,
   "kontext": 524288,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 161.95,
   "decode_toks": 38.6,
   "parallel_tok_s": null,
   "prefill_tok_s": 1556.6,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22203,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 484636,
    "antwortzeit_median_s": 348.4,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "q4_0",
   "resident": 1,
   "kontext": 262144,
   "sessions": 1,
   "phase": "a4",
   "rope": "none",
   "plattform": "gb10-llama",
   "n": 0,
   "ttft_frisch_32k_s": 404.58,
   "decode_toks": null,
   "parallel_tok_s": null,
   "prefill_tok_s": 640.6,
   "prompt_token_frisch_32k_median": null,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": null,
   "nadel": {
    "treffer": 3,
    "versuche": 3,
    "prompt_token_median": 261341,
    "antwortzeit_median_s": 502.0,
    "tiefen": [
     0.1,
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "q4_0",
   "resident": 20480,
   "kontext": 1048576,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 158.37,
   "decode_toks": 50.2,
   "parallel_tok_s": null,
   "prefill_tok_s": 1595.4,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22205,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 970201,
    "antwortzeit_median_s": 778.7,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "q4_0",
   "resident": 20480,
   "kontext": 524288,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 128.99,
   "decode_toks": 45.0,
   "parallel_tok_s": null,
   "prefill_tok_s": 2082.6,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22201,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 523795,
    "antwortzeit_median_s": 347.9,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "q4_0",
   "resident": 65536,
   "kontext": 1048576,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 160.56,
   "decode_toks": 50.7,
   "parallel_tok_s": null,
   "prefill_tok_s": 1588.9,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22207,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 969377,
    "antwortzeit_median_s": 789.0,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "q4_0",
   "resident": 65536,
   "kontext": 524288,
   "sessions": 1,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 129.68,
   "decode_toks": 49.4,
   "parallel_tok_s": null,
   "prefill_tok_s": 2071.0,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22201,
   "nadel": {
    "treffer": 2,
    "versuche": 2,
    "prompt_token_median": 523759,
    "antwortzeit_median_s": 419.3,
    "tiefen": [
     0.5,
     0.9
    ]
   }
  }
 ],
 "yarn_doppel": [
  {
   "kv": "int8",
   "resident": 20480,
   "kontext": 524288,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 127.4,
   "decode_toks": 54.9,
   "parallel_tok_s": 77.4,
   "prefill_tok_s": 2092.1,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22187
  },
  {
   "kv": "int8",
   "resident": 65536,
   "kontext": 524288,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 128.88,
   "decode_toks": 54.3,
   "parallel_tok_s": 76.1,
   "prefill_tok_s": 2068.1,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22189
  },
  {
   "kv": "k8v4",
   "resident": 20480,
   "kontext": 524288,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 175.18,
   "decode_toks": 12.8,
   "parallel_tok_s": 16.5,
   "prefill_tok_s": 1854.3,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22203
  },
  {
   "kv": "k8v4",
   "resident": 65536,
   "kontext": 524288,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 193.83,
   "decode_toks": 8.7,
   "parallel_tok_s": 9.7,
   "prefill_tok_s": 1444.9,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22211
  },
  {
   "kv": "q4_0",
   "resident": 20480,
   "kontext": 1048576,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 131.67,
   "decode_toks": 47.8,
   "parallel_tok_s": 76.2,
   "prefill_tok_s": 2042.4,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22191
  },
  {
   "kv": "q4_0",
   "resident": 20480,
   "kontext": 524288,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 131.48,
   "decode_toks": 46.8,
   "parallel_tok_s": 70.9,
   "prefill_tok_s": 2043.3,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22189
  },
  {
   "kv": "q4_0",
   "resident": 65536,
   "kontext": 1048576,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x4",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 132.53,
   "decode_toks": 47.4,
   "parallel_tok_s": 72.1,
   "prefill_tok_s": 2028.3,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22207
  },
  {
   "kv": "q4_0",
   "resident": 65536,
   "kontext": 524288,
   "sessions": 2,
   "phase": "a2",
   "rope": "yarn x2",
   "plattform": "3090-strata",
   "n": 1,
   "ttft_frisch_32k_s": 130.1,
   "decode_toks": 47.8,
   "parallel_tok_s": 73.9,
   "prefill_tok_s": 2063.6,
   "prompt_token_frisch_32k_median": 260470,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": 22199
  }
 ],
 "gb10": [
  {
   "kv": "int8",
   "resident": 1,
   "kontext": 262144,
   "sessions": 1,
   "phase": "a4",
   "rope": "none",
   "plattform": "gb10-llama",
   "n": 0,
   "ttft_frisch_32k_s": 379.39,
   "decode_toks": null,
   "parallel_tok_s": null,
   "prefill_tok_s": 697.9,
   "prompt_token_frisch_32k_median": null,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": null,
   "nadel": {
    "treffer": 3,
    "versuche": 3,
    "prompt_token_median": 261175,
    "antwortzeit_median_s": 471.6,
    "tiefen": [
     0.1,
     0.5,
     0.9
    ]
   }
  },
  {
   "kv": "q4_0",
   "resident": 1,
   "kontext": 262144,
   "sessions": 1,
   "phase": "a4",
   "rope": "none",
   "plattform": "gb10-llama",
   "n": 0,
   "ttft_frisch_32k_s": 404.58,
   "decode_toks": null,
   "parallel_tok_s": null,
   "prefill_tok_s": 640.6,
   "prompt_token_frisch_32k_median": null,
   "gleich_papier_ttft_s": null,
   "vram_max_mb": null,
   "nadel": {
    "treffer": 3,
    "versuche": 3,
    "prompt_token_median": 261341,
    "antwortzeit_median_s": 502.0,
    "tiefen": [
     0.1,
     0.5,
     0.9
    ]
   }
  }
 ]
}