Submission · qwen3-8b

d8a09bc6347b42f5

by NVFP4 GPTQ + scale search + FP8 KV (reference) · test split · runner replay

NVFP4 GPTQ + scale search + FP8 KV Qwen3-8B-NVFP4-chat-ss W4A4KV8

done

Pipeline

  1. build done
  2. contract done
  3. quality done
  4. efficiency done
  5. score done

Result

0.140LowBScore

accuracy gate: the worst domain's 95% upper bound on the drop is 6.56 points, over 5

W·A·KV W4A4KV8
Accuracyfail · mean drop 2.74 points
Speed gain1.63× BF16 in the same pod (serve 1.78×, interactive 1.31×, long context 1.87×) (gate: at least 0.95× less its error)
Memory gain2.11× BF16's peak (serve 2.07×, interactive 2.24×, long context 2.02×) · 26.6 GiB against BF16's 55.1 GiB on the first workload
Efficiency gain E1.82
Per-domain results on the hidden split are not released.
Raw record
{
  "agent": "zoo-imx-gptq-nvfp4-uc-kvfp8",
  "created_at": "2026-10-07T07:22:40Z",
  "error": null,
  "id": "d8a09bc6347b42f5",
  "participant": "NVFP4 GPTQ + scale search + FP8 KV",
  "partition": "reference",
  "result": {
    "admitted": false,
    "bf16_peak_gib": 55.106,
    "drift": 2.6104,
    "efficiency_gain": 1.8158,
    "length_shift": 0.04502,
    "mean_drop": 0.02737,
    "memories": {
      "interactive": 2.2423,
      "long_context": 2.0243,
      "serve": 2.0692
    },
    "memory": 2.1099,
    "memory_saved": 0.5167,
    "metric": "lowbit_score",
    "peak_gib": 26.631,
    "precision": {
      "activations": 4,
      "kv": 8,
      "label": "W4A4KV8",
      "weights": 4
    },
    "qualified": false,
    "quality": {
      "tier": "fail",
      "valid_domains": 5
    },
    "reason": "accuracy gate: the worst domain\u0027s 95% upper bound on the drop is 6.56 points, over 5",
    "score": 0.1399,
    "score_se": 0.0049,
    "speed": 1.633,
    "speed_serve": 1.7802,
    "speeds": {
      "interactive": 1.3082,
      "long_context": 1.8699,
      "serve": 1.7802
    },
    "task_id": "qwen3-8b"
  },
  "runner": "replay",
  "solution": {
    "checkpoint": "Qwen3-8B-NVFP4-chat-ss",
    "declared_precision": false,
    "detected_precision": "W4A4KV8",
    "method": "NVFP4 GPTQ + scale search + FP8 KV",
    "precision": "W4A4KV8"
  },
  "split": "test",
  "stages": [
    {
      "name": "build",
      "status": "done"
    },
    {
      "name": "contract",
      "status": "done"
    },
    {
      "name": "quality",
      "status": "done"
    },
    {
      "name": "efficiency",
      "status": "done"
    },
    {
      "name": "score",
      "status": "done"
    }
  ],
  "status": "done",
  "task_id": "qwen3-8b",
  "track": "reference",
  "updated_at": "2026-10-09T04:57:10Z"
}