{
 "schema_version": "1.0.0",
 "build": "cuda-qwen-1.1.4",
 "rows": [
  {
   "id": "cuda-decode-rtx-3090-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 967.5,
   "unit": "tok/s",
   "runs": 2,
   "min": 965.9,
   "max": 969.2,
   "spread_pct": 0.34,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-0.8b-balanced",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.27,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 880.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 877.6,
   "max": 883.9,
   "spread_pct": 0.72,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-0.8b-quality",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.42,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 972.0,
   "unit": "tok/s",
   "runs": 2,
   "min": 967.7,
   "max": 976.4,
   "spread_pct": 0.9,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-0.8b-fast",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.27,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 586.5,
   "unit": "tok/s",
   "runs": 2,
   "min": 584.8,
   "max": 588.2,
   "spread_pct": 0.58,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-2b-balanced",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.48,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 542.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 542.5,
   "max": 543.0,
   "spread_pct": 0.09,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-2b-quality",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.8,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 585.1,
   "unit": "tok/s",
   "runs": 2,
   "min": 584.2,
   "max": 586.0,
   "spread_pct": 0.31,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-2b-fast",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.48,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 376.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 376.5,
   "max": 377.2,
   "spread_pct": 0.19,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-4b-balanced",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.49,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 342.1,
   "unit": "tok/s",
   "runs": 2,
   "min": 342.0,
   "max": 342.1,
   "spread_pct": 0.03,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-4b-quality",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 6.27,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 376.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 376.2,
   "max": 377.4,
   "spread_pct": 0.32,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-4b-fast",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.49,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 190.2,
   "unit": "tok/s",
   "runs": 2,
   "min": 190.1,
   "max": 190.3,
   "spread_pct": 0.11,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-9b-balanced",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.44,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 172.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 172.6,
   "max": 172.9,
   "spread_pct": 0.17,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-9b-quality",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 9.82,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 190.4,
   "unit": "tok/s",
   "runs": 2,
   "min": 190.4,
   "max": 190.4,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen35-9b-fast",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.44,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/decode/20/1024",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 1024,
   "value": 1052.0,
   "unit": "tok/s",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/decode/20/1024",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 1024,
   "value": 560.4,
   "unit": "tok/s",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/decode/20/1024",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 1024,
   "value": 397.8,
   "unit": "tok/s",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/decode/20/1024",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 1024,
   "value": 170.2,
   "unit": "tok/s",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1317.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 1316.1,
   "max": 1319.5,
   "spread_pct": 0.26,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-0.8b-balanced",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.4,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1156.0,
   "unit": "tok/s",
   "runs": 2,
   "min": 1155.4,
   "max": 1156.5,
   "spread_pct": 0.1,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-0.8b-quality",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.55,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1316.3,
   "unit": "tok/s",
   "runs": 2,
   "min": 1315.2,
   "max": 1317.4,
   "spread_pct": 0.17,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-0.8b-fast",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.4,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 886.9,
   "unit": "tok/s",
   "runs": 2,
   "min": 879.2,
   "max": 894.7,
   "spread_pct": 1.75,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-2b-balanced",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.61,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 740.9,
   "unit": "tok/s",
   "runs": 2,
   "min": 740.9,
   "max": 741.0,
   "spread_pct": 0.01,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-2b-quality",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.93,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 895.1,
   "unit": "tok/s",
   "runs": 2,
   "min": 895.0,
   "max": 895.2,
   "spread_pct": 0.02,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-2b-fast",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.61,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 531.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 528.6,
   "max": 535.0,
   "spread_pct": 1.2,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-4b-balanced",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.62,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 441.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 441.8,
   "max": 441.8,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-4b-quality",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 6.4,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 535.4,
   "unit": "tok/s",
   "runs": 2,
   "min": 535.1,
   "max": 535.7,
   "spread_pct": 0.11,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-4b-fast",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.62,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 347.0,
   "unit": "tok/s",
   "runs": 2,
   "min": 346.7,
   "max": 347.3,
   "spread_pct": 0.17,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-9b-balanced",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.57,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 272.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 272.8,
   "max": 272.8,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-9b-quality",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 9.95,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 346.6,
   "unit": "tok/s",
   "runs": 2,
   "min": 346.1,
   "max": 347.2,
   "spread_pct": 0.32,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-4090-qwen35-9b-fast",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.57,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 1438.7,
   "unit": "tok/s",
   "runs": 1,
   "min": 1438.7,
   "max": 1438.7,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 868.8,
   "unit": "tok/s",
   "runs": 1,
   "min": 868.8,
   "max": 868.8,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 568.2,
   "unit": "tok/s",
   "runs": 1,
   "min": 568.2,
   "max": 568.2,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-4090-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 303.4,
   "unit": "tok/s",
   "runs": 1,
   "min": 303.4,
   "max": 303.4,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1529.1,
   "unit": "tok/s",
   "runs": 2,
   "min": 1527.6,
   "max": 1530.6,
   "spread_pct": 0.2,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-0.8b-balanced",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.51,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1437.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 1437.4,
   "max": 1438.1,
   "spread_pct": 0.05,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-0.8b-quality",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.66,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1530.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 1529.7,
   "max": 1532.0,
   "spread_pct": 0.15,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-0.8b-fast",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.51,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1138.0,
   "unit": "tok/s",
   "runs": 2,
   "min": 1137.7,
   "max": 1138.3,
   "spread_pct": 0.05,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-2b-balanced",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.72,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1040.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 1037.9,
   "max": 1043.4,
   "spread_pct": 0.53,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-2b-quality",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 4.03,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 1136.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 1135.3,
   "max": 1138.1,
   "spread_pct": 0.25,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-2b-fast",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.72,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 772.2,
   "unit": "tok/s",
   "runs": 2,
   "min": 772.1,
   "max": 772.3,
   "spread_pct": 0.03,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-4b-balanced",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.73,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 670.5,
   "unit": "tok/s",
   "runs": 2,
   "min": 670.4,
   "max": 670.5,
   "spread_pct": 0.01,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-4b-quality",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 6.5,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 771.4,
   "unit": "tok/s",
   "runs": 2,
   "min": 770.6,
   "max": 772.3,
   "spread_pct": 0.22,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-4b-fast",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.73,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 498.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 498.6,
   "max": 498.8,
   "spread_pct": 0.04,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-9b-balanced",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.67,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 423.4,
   "unit": "tok/s",
   "runs": 2,
   "min": 423.3,
   "max": 423.4,
   "spread_pct": 0.02,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-9b-quality",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 10.05,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/decode/21/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 498.5,
   "unit": "tok/s",
   "runs": 2,
   "min": 498.4,
   "max": 498.6,
   "spread_pct": 0.04,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen35-9b-fast",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.67,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen36-35b-a3b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/decode/21/128",
   "axis": "decode",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 449.9,
   "unit": "tok/s",
   "runs": 2,
   "min": 449.9,
   "max": 450.0,
   "spread_pct": 0.02,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen36-35b-a3b-balanced",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 21.67,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen36-35b-a3b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/decode/21/128",
   "axis": "decode",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 544.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 544.6,
   "max": 544.7,
   "spread_pct": 0.02,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-5090-qwen36-35b-a3b-fast",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 21.74,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 1675.8,
   "unit": "tok/s",
   "runs": 1,
   "min": 1675.8,
   "max": 1675.8,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 1107.2,
   "unit": "tok/s",
   "runs": 1,
   "min": 1107.2,
   "max": 1107.2,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 822.2,
   "unit": "tok/s",
   "runs": 1,
   "min": 822.2,
   "max": 822.2,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/decode/21/1024",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 434.6,
   "unit": "tok/s",
   "runs": 1,
   "min": 434.6,
   "max": 434.6,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-5090-qwen36-35b-a3b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/decode/21/1024",
   "axis": "decode",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 449.6,
   "unit": "tok/s",
   "runs": 1,
   "min": 449.6,
   "max": 449.6,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 4.18,
   "unit": "ms",
   "runs": 2,
   "min": 4.18,
   "max": 4.18,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 4.09,
   "unit": "ms",
   "runs": 2,
   "min": 4.06,
   "max": 4.12,
   "spread_pct": 1.47,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 4.09,
   "unit": "ms",
   "runs": 2,
   "min": 4.0,
   "max": 4.18,
   "spread_pct": 4.4,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 6.63,
   "unit": "ms",
   "runs": 2,
   "min": 6.6,
   "max": 6.67,
   "spread_pct": 1.06,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 6.9,
   "unit": "ms",
   "runs": 2,
   "min": 6.85,
   "max": 6.95,
   "spread_pct": 1.45,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 6.67,
   "unit": "ms",
   "runs": 2,
   "min": 6.66,
   "max": 6.67,
   "spread_pct": 0.15,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 13.51,
   "unit": "ms",
   "runs": 2,
   "min": 13.38,
   "max": 13.64,
   "spread_pct": 1.92,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 14.03,
   "unit": "ms",
   "runs": 2,
   "min": 13.72,
   "max": 14.33,
   "spread_pct": 4.35,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 13.11,
   "unit": "ms",
   "runs": 2,
   "min": 13.04,
   "max": 13.18,
   "spread_pct": 1.07,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 21.72,
   "unit": "ms",
   "runs": 2,
   "min": 21.7,
   "max": 21.74,
   "spread_pct": 0.18,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 23.25,
   "unit": "ms",
   "runs": 2,
   "min": 23.25,
   "max": 23.26,
   "spread_pct": 0.04,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 21.77,
   "unit": "ms",
   "runs": 2,
   "min": 21.74,
   "max": 21.8,
   "spread_pct": 0.28,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (20-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 4.379,
   "unit": "ms",
   "runs": 1,
   "min": 4.379,
   "max": 4.379,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-2b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 7.119,
   "unit": "ms",
   "runs": 1,
   "min": 7.119,
   "max": 7.119,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-4b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 13.902,
   "unit": "ms",
   "runs": 1,
   "min": 13.902,
   "max": 13.902,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 22.421,
   "unit": "ms",
   "runs": 1,
   "min": 22.421,
   "max": 22.421,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 3.001,
   "unit": "ms",
   "runs": 2,
   "min": 2.995,
   "max": 3.007,
   "spread_pct": 0.4,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 3.109,
   "unit": "ms",
   "runs": 2,
   "min": 3.107,
   "max": 3.111,
   "spread_pct": 0.13,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 3.006,
   "unit": "ms",
   "runs": 2,
   "min": 3.001,
   "max": 3.01,
   "spread_pct": 0.3,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 4.494,
   "unit": "ms",
   "runs": 2,
   "min": 4.369,
   "max": 4.618,
   "spread_pct": 5.54,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 4.694,
   "unit": "ms",
   "runs": 2,
   "min": 4.693,
   "max": 4.695,
   "spread_pct": 0.04,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 4.37,
   "unit": "ms",
   "runs": 2,
   "min": 4.367,
   "max": 4.373,
   "spread_pct": 0.14,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 8.73,
   "unit": "ms",
   "runs": 2,
   "min": 8.595,
   "max": 8.866,
   "spread_pct": 3.1,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 10.539,
   "unit": "ms",
   "runs": 2,
   "min": 10.539,
   "max": 10.539,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 8.68,
   "unit": "ms",
   "runs": 2,
   "min": 8.617,
   "max": 8.742,
   "spread_pct": 1.44,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 16.82,
   "unit": "ms",
   "runs": 2,
   "min": 16.707,
   "max": 16.934,
   "spread_pct": 1.35,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 21.617,
   "unit": "ms",
   "runs": 2,
   "min": 21.614,
   "max": 21.62,
   "spread_pct": 0.03,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 16.472,
   "unit": "ms",
   "runs": 2,
   "min": 16.461,
   "max": 16.484,
   "spread_pct": 0.14,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-0.8b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 3.013,
   "unit": "ms",
   "runs": 1,
   "min": 3.013,
   "max": 3.013,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-2b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 4.375,
   "unit": "ms",
   "runs": 1,
   "min": 4.375,
   "max": 4.375,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-4b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 8.604,
   "unit": "ms",
   "runs": 1,
   "min": 8.604,
   "max": 8.604,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-4090-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-4090/qwen35-9b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-4090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 16.616,
   "unit": "ms",
   "runs": 1,
   "min": 16.616,
   "max": 16.616,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 2.744,
   "unit": "ms",
   "runs": 2,
   "min": 2.74,
   "max": 2.749,
   "spread_pct": 0.33,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 2.942,
   "unit": "ms",
   "runs": 2,
   "min": 2.932,
   "max": 2.951,
   "spread_pct": 0.65,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 2.744,
   "unit": "ms",
   "runs": 2,
   "min": 2.735,
   "max": 2.753,
   "spread_pct": 0.66,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 3.782,
   "unit": "ms",
   "runs": 2,
   "min": 3.778,
   "max": 3.786,
   "spread_pct": 0.21,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 4.106,
   "unit": "ms",
   "runs": 2,
   "min": 4.104,
   "max": 4.108,
   "spread_pct": 0.1,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 3.828,
   "unit": "ms",
   "runs": 2,
   "min": 3.816,
   "max": 3.839,
   "spread_pct": 0.6,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 6.885,
   "unit": "ms",
   "runs": 2,
   "min": 6.882,
   "max": 6.887,
   "spread_pct": 0.07,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 7.904,
   "unit": "ms",
   "runs": 2,
   "min": 7.891,
   "max": 7.916,
   "spread_pct": 0.32,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 6.917,
   "unit": "ms",
   "runs": 2,
   "min": 6.91,
   "max": 6.924,
   "spread_pct": 0.2,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 11.057,
   "unit": "ms",
   "runs": 2,
   "min": 11.047,
   "max": 11.067,
   "spread_pct": 0.18,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 13.022,
   "unit": "ms",
   "runs": 2,
   "min": 13.013,
   "max": 13.032,
   "spread_pct": 0.15,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 11.085,
   "unit": "ms",
   "runs": 2,
   "min": 11.072,
   "max": 11.098,
   "spread_pct": 0.23,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen36-35b-a3b-balanced-128",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 61.251,
   "unit": "ms",
   "runs": 2,
   "min": 61.184,
   "max": 61.318,
   "spread_pct": 0.22,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen36-35b-a3b-fast-128",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/prompt_latency/21/128",
   "axis": "prompt_latency",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 128,
   "value": 61.268,
   "unit": "ms",
   "runs": 2,
   "min": 61.19,
   "max": 61.346,
   "spread_pct": 0.25,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-0.8b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 2.748,
   "unit": "ms",
   "runs": 1,
   "min": 2.748,
   "max": 2.748,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-2b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 3.786,
   "unit": "ms",
   "runs": 1,
   "min": 3.786,
   "max": 3.786,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-4b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 6.906,
   "unit": "ms",
   "runs": 1,
   "min": 6.906,
   "max": 6.906,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen35-9b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 11.118,
   "unit": "ms",
   "runs": 1,
   "min": 11.118,
   "max": 11.118,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-5090-qwen36-35b-a3b-balanced-1024",
   "cell": "nvidia-cuda/rtx-5090/qwen36-35b-a3b/prompt_latency/21/1024",
   "axis": "prompt_latency",
   "model": "qwen36-35b-a3b",
   "device": "rtx-5090",
   "target": "nvidia-cuda",
   "in": 21,
   "out": 1024,
   "value": 61.477,
   "unit": "ms",
   "runs": 1,
   "min": 61.477,
   "max": 61.477,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "fact_chatml (21-token ChatML prompt, seed 42, greedy)",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 4.864,
   "unit": "ms",
   "runs": 2,
   "min": 4.863,
   "max": 4.866,
   "spread_pct": 0.06,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 4.648,
   "unit": "ms",
   "runs": 2,
   "min": 4.644,
   "max": 4.652,
   "spread_pct": 0.17,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 4.892,
   "unit": "ms",
   "runs": 2,
   "min": 4.891,
   "max": 4.893,
   "spread_pct": 0.04,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 8.388,
   "unit": "ms",
   "runs": 2,
   "min": 8.382,
   "max": 8.394,
   "spread_pct": 0.14,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 9.018,
   "unit": "ms",
   "runs": 2,
   "min": 8.992,
   "max": 9.043,
   "spread_pct": 0.57,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 8.412,
   "unit": "ms",
   "runs": 2,
   "min": 8.395,
   "max": 8.429,
   "spread_pct": 0.4,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 17.218,
   "unit": "ms",
   "runs": 2,
   "min": 16.978,
   "max": 17.459,
   "spread_pct": 2.79,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 19.175,
   "unit": "ms",
   "runs": 2,
   "min": 18.92,
   "max": 19.43,
   "spread_pct": 2.66,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 17.208,
   "unit": "ms",
   "runs": 2,
   "min": 16.932,
   "max": 17.485,
   "spread_pct": 3.21,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 28.99,
   "unit": "ms",
   "runs": 2,
   "min": 28.958,
   "max": 29.022,
   "spread_pct": 0.22,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 30.951,
   "unit": "ms",
   "runs": 2,
   "min": 30.877,
   "max": 31.024,
   "spread_pct": 0.47,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/prompt_latency/20/128",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 29.211,
   "unit": "ms",
   "runs": 2,
   "min": 29.189,
   "max": 29.234,
   "spread_pct": 0.15,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/prompt_latency/921/1024",
   "axis": "prompt_latency",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 28.575,
   "unit": "ms",
   "runs": 1,
   "min": 28.575,
   "max": 28.575,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/prompt_latency/921/1024",
   "axis": "prompt_latency",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 58.492,
   "unit": "ms",
   "runs": 1,
   "min": 58.492,
   "max": 58.492,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/prompt_latency/921/1024",
   "axis": "prompt_latency",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 154.003,
   "unit": "ms",
   "runs": 1,
   "min": 154.003,
   "max": 154.003,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-a5000-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/prompt_latency/921/1024",
   "axis": "prompt_latency",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 269.87,
   "unit": "ms",
   "runs": 1,
   "min": 269.87,
   "max": 269.87,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-0.8b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 776.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 776.2,
   "max": 777.1,
   "spread_pct": 0.12,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-0.8b-balanced",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.22,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-0.8b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 714.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 712.2,
   "max": 717.1,
   "spread_pct": 0.69,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-0.8b-quality",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.37,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-0.8b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 771.9,
   "unit": "tok/s",
   "runs": 2,
   "min": 771.5,
   "max": 772.2,
   "spread_pct": 0.09,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-0.8b-fast",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 2.22,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-2b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 467.3,
   "unit": "tok/s",
   "runs": 2,
   "min": 465.8,
   "max": 468.9,
   "spread_pct": 0.66,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-2b-balanced",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.43,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-2b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 435.2,
   "unit": "tok/s",
   "runs": 2,
   "min": 434.8,
   "max": 435.6,
   "spread_pct": 0.18,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-2b-quality",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.75,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-2b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 464.6,
   "unit": "tok/s",
   "runs": 2,
   "min": 463.5,
   "max": 465.7,
   "spread_pct": 0.47,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-2b-fast",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 3.43,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-4b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 280.5,
   "unit": "tok/s",
   "runs": 2,
   "min": 279.3,
   "max": 281.7,
   "spread_pct": 0.86,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-4b-balanced",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.44,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-4b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 256.5,
   "unit": "tok/s",
   "runs": 2,
   "min": 255.9,
   "max": 257.1,
   "spread_pct": 0.47,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-4b-quality",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 6.22,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-4b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 278.9,
   "unit": "tok/s",
   "runs": 2,
   "min": 278.0,
   "max": 279.9,
   "spread_pct": 0.68,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-4b-fast",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 5.44,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-9b-balanced-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 151.0,
   "unit": "tok/s",
   "runs": 2,
   "min": 150.8,
   "max": 151.1,
   "spread_pct": 0.2,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-9b-balanced",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.39,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-9b-quality-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 137.8,
   "unit": "tok/s",
   "runs": 2,
   "min": 137.2,
   "max": 138.4,
   "spread_pct": 0.87,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-9b-quality",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 9.77,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-9b-fast-128",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/decode/20/128",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 20,
   "out": 128,
   "value": 149.1,
   "unit": "tok/s",
   "runs": 2,
   "min": 148.8,
   "max": 149.3,
   "spread_pct": 0.34,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-a5000-qwen35-9b-fast",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 8.39,
   "unit": "GiB",
   "runs": 2,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-0.8b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-0.8b/decode/921/1024",
   "axis": "decode",
   "model": "qwen35-0.8b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 916.1,
   "unit": "tok/s",
   "runs": 1,
   "min": 916.1,
   "max": 916.1,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-2b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-2b/decode/921/1024",
   "axis": "decode",
   "model": "qwen35-2b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 449.1,
   "unit": "tok/s",
   "runs": 1,
   "min": 449.1,
   "max": 449.1,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-4b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-4b/decode/921/1024",
   "axis": "decode",
   "model": "qwen35-4b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 228.6,
   "unit": "tok/s",
   "runs": 1,
   "min": 228.6,
   "max": 228.6,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-a5000-qwen35-9b-balanced-1024",
   "cell": "nvidia-cuda/rtx-a5000/qwen35-9b/decode/921/1024",
   "axis": "decode",
   "model": "qwen35-9b",
   "device": "rtx-a5000",
   "target": "nvidia-cuda",
   "in": 921,
   "out": 1024,
   "value": 124.8,
   "unit": "tok/s",
   "runs": 1,
   "min": 124.8,
   "max": 124.8,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-21",
   "kind": "carried",
   "carried_from": "cuda-qwen-llm-1.0.0-be389a65",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "acceptance prompt (20 tokens; not in the reproduction kit), seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen35-9b-c1",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/throughput/34/128",
   "axis": "throughput",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 101.4,
   "unit": "tok/s",
   "runs": 3,
   "min": 101.4,
   "max": 101.4,
   "spread_pct": 0.0,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen35-9b-c1",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/ttft/34/128",
   "axis": "ttft",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 55.6,
   "unit": "ms",
   "runs": 3,
   "min": 55.5,
   "max": 55.8,
   "spread_pct": 0.54,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen35-9b-c4",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/throughput/34/128",
   "axis": "throughput",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 366.4,
   "unit": "tok/s",
   "runs": 3,
   "min": 366.4,
   "max": 366.9,
   "spread_pct": 0.14,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen35-9b-c4",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/ttft/34/128",
   "axis": "ttft",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 176.2,
   "unit": "ms",
   "runs": 3,
   "min": 133.1,
   "max": 176.4,
   "spread_pct": 24.57,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen35-9b-c8",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/throughput/34/128",
   "axis": "throughput",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 663.0,
   "unit": "tok/s",
   "runs": 3,
   "min": 662.7,
   "max": 663.3,
   "spread_pct": 0.09,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen35-9b-c8",
   "cell": "nvidia-cuda/rtx-3090/qwen35-9b/ttft/34/128",
   "axis": "ttft",
   "model": "qwen35-9b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 308.8,
   "unit": "ms",
   "runs": 3,
   "min": 308.4,
   "max": 308.9,
   "spread_pct": 0.16,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen35-0.8b-c1",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/throughput/36/128",
   "axis": "throughput",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 36,
   "out": 128,
   "value": 531.9,
   "unit": "tok/s",
   "runs": 3,
   "min": 528.2,
   "max": 532.1,
   "spread_pct": 0.73,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "36-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen35-0.8b-c1",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/ttft/36/128",
   "axis": "ttft",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 36,
   "out": 128,
   "value": 24.6,
   "unit": "ms",
   "runs": 3,
   "min": 24.5,
   "max": 26.6,
   "spread_pct": 8.54,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "36-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen35-0.8b-c4",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/throughput/36/128",
   "axis": "throughput",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 36,
   "out": 128,
   "value": 1778.1,
   "unit": "tok/s",
   "runs": 3,
   "min": 1776.5,
   "max": 1780.1,
   "spread_pct": 0.2,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "36-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen35-0.8b-c4",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/ttft/36/128",
   "axis": "ttft",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 36,
   "out": 128,
   "value": 62.9,
   "unit": "ms",
   "runs": 3,
   "min": 62.6,
   "max": 63.9,
   "spread_pct": 2.07,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "36-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen35-0.8b-c8",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/throughput/36/128",
   "axis": "throughput",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 36,
   "out": 128,
   "value": 3125.2,
   "unit": "tok/s",
   "runs": 3,
   "min": 3124.7,
   "max": 3130.5,
   "spread_pct": 0.19,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "36-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen35-0.8b-c8",
   "cell": "nvidia-cuda/rtx-3090/qwen35-0.8b/ttft/36/128",
   "axis": "ttft",
   "model": "qwen35-0.8b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 36,
   "out": 128,
   "value": 92.6,
   "unit": "ms",
   "runs": 3,
   "min": 92.5,
   "max": 92.8,
   "spread_pct": 0.32,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "36-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen36-35b-a3b-balanced-128",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/decode/22/128",
   "axis": "decode",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 22,
   "out": 128,
   "value": 194.3,
   "unit": "tok/s",
   "runs": 3,
   "min": 194.0,
   "max": 195.4,
   "spread_pct": 0.72,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen36-35b-a3b-balanced",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/prompt_latency/22/128",
   "axis": "prompt_latency",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 22,
   "out": 128,
   "value": 107.25,
   "unit": "ms",
   "runs": 3,
   "min": 106.87,
   "max": 109.05,
   "spread_pct": 2.03,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen36-35b-a3b-balanced",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 21.44,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen36-35b-a3b-fast-128",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/decode/22/128",
   "axis": "decode",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 22,
   "out": 128,
   "value": 172.7,
   "unit": "tok/s",
   "runs": 2,
   "min": 172.3,
   "max": 173.1,
   "spread_pct": 0.46,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-prompt-rtx-3090-qwen36-35b-a3b-fast",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/prompt_latency/22/128",
   "axis": "prompt_latency",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 22,
   "out": 128,
   "value": 107.63,
   "unit": "ms",
   "runs": 2,
   "min": 107.43,
   "max": 107.82,
   "spread_pct": 0.36,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-prompt-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-resident-rtx-3090-qwen36-35b-a3b-fast",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/resident_memory//",
   "axis": "resident_memory",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": null,
   "out": null,
   "value": 21.51,
   "unit": "GiB",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-memory-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen36-35b-a3b-balanced-1024",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/decode/22/1024",
   "axis": "decode",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 22,
   "out": 1024,
   "value": 209.7,
   "unit": "tok/s",
   "runs": 1,
   "min": null,
   "max": null,
   "spread_pct": null,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-decode-rtx-3090-qwen36-35b-a3b-single-128",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/decode/22/128",
   "axis": "decode",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 22,
   "out": 128,
   "value": 197.9,
   "unit": "tok/s",
   "runs": 3,
   "min": 197.6,
   "max": 200.7,
   "spread_pct": 1.57,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-decode-v1",
   "prompt_set": "22-token ChatML prompt, seed 42, greedy",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen36-35b-a3b-c1",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/throughput/34/128",
   "axis": "throughput",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 160.9,
   "unit": "tok/s",
   "runs": 3,
   "min": 159.1,
   "max": 160.9,
   "spread_pct": 1.12,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen36-35b-a3b-c1",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/ttft/34/128",
   "axis": "ttft",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 182.7,
   "unit": "ms",
   "runs": 3,
   "min": 182.1,
   "max": 185.8,
   "spread_pct": 2.03,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen36-35b-a3b-c4",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/throughput/34/128",
   "axis": "throughput",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 222.4,
   "unit": "tok/s",
   "runs": 3,
   "min": 221.3,
   "max": 224.4,
   "spread_pct": 1.39,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen36-35b-a3b-c4",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/ttft/34/128",
   "axis": "ttft",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 625.5,
   "unit": "ms",
   "runs": 3,
   "min": 622.5,
   "max": 629.8,
   "spread_pct": 1.17,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-serve-rtx-3090-qwen36-35b-a3b-c8",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/throughput/34/128",
   "axis": "throughput",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 230.6,
   "unit": "tok/s",
   "runs": 3,
   "min": 229.4,
   "max": 230.8,
   "spread_pct": 0.61,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  },
  {
   "id": "cuda-ttft-rtx-3090-qwen36-35b-a3b-c8",
   "cell": "nvidia-cuda/rtx-3090/qwen36-35b-a3b/ttft/34/128",
   "axis": "ttft",
   "model": "qwen36-35b-a3b",
   "device": "rtx-3090",
   "target": "nvidia-cuda",
   "in": 34,
   "out": 128,
   "value": 1001.0,
   "unit": "ms",
   "runs": 3,
   "min": 998.5,
   "max": 1003.3,
   "spread_pct": 0.48,
   "bound": "exact",
   "build": "cuda-qwen-1.1.4",
   "measured": "2026-10-22",
   "kind": "carried",
   "carried_from": "cuda-qwen-1.0.0",
   "kernel_digest": null,
   "protocol": "cuda-own-serve-v1",
   "prompt_set": "34-token chat request, 128 tokens each, temperature 0, streamed",
   "condition": null
  }
 ]
}