Submission · qwen3-8b

`727880e648af4ca9`

by Unsloth dynamic NF4 + FP8 KV (reference) · test split · runner kubernetes

NF4 dynamic, Unsloth + FP8 KV unsloth/Qwen3-8B-unsloth-bnb-4bit W4A16KV8

done

Pipeline

build done
contract done
quality done
efficiency done
score done

Result

pending LowBScore

accuracy gate: the worst domain's 95% upper bound on the drop is 6.39 points, over 5; speed or memory on the other workloads not measured

W·A·KV	W4A16KV8 declared · the service reads W4A16KV8
Accuracy	fail · mean drop 1.90 points
Drift from BF16	KL 0.037 nats per token
Speed gain	0.74× BF16 in the same pod (serve 0.74×) (gate: at least 0.95× less its error)
Memory gain	— BF16's peak (serve 1.98×) · 28.1 GiB against BF16's 55.6 GiB on the first workload
Per-domain results on the hidden split are not released.

0.52 GPU-hours

Raw record

{
  "agent": "zoo-pub-unsloth-dyn-nf4-kvfp8",
  "created_at": "2026-10-08T05:59:14Z",
  "error": null,
  "id": "727880e648af4ca9",
  "participant": "Unsloth dynamic NF4 + FP8 KV",
  "partition": "reference",
  "result": {
    "admitted": false,
    "anchor_efficiency": {
      "serve": 1982.2602300719493
    },
    "bf16_peak_gib": 55.624,
    "efficiency": {
      "serve": 1467.2778171133546
    },
    "gpu_hours": 0.518,
    "kl": {
      "kl": 0.03727583430931515,
      "kl_median": 0.03453909987822208,
      "kl_per_token": 0.036643892876721436,
      "kl_se": 0.0017799672098198027,
      "sequences": 256,
      "tokens": 62197
    },
    "mean_drop": 0.01901,
    "memories": {
      "interactive": null,
      "long_context": null,
      "serve": 1.9773
    },
    "memory": null,
    "memory_saved": 0.4942,
    "metric": "lowbit_score",
    "peak_gib": 28.132,
    "peak_memory": {
      "serve": {
        "device_peak_gib": 163.9,
        "kv_bytes_per_token": 73579,
        "kv_gib": 19.086,
        "kv_usage": 0.1255,
        "other_gib": 1.775,
        "peak_gib": 28.132,
        "weights_gib": 7.27
      }
    },
    "precision": {
      "activations": 16,
      "kv": 8,
      "label": "W4A16KV8",
      "weights": 4,
      "weights_average": 5.64
    },
    "qualified": false,
    "quality": {
      "tier": "fail",
      "valid_domains": 5
    },
    "reason": "accuracy gate: the worst domain\u0027s 95% upper bound on the drop is 6.39 points, over 5; speed or memory on the other workloads not measured",
    "score": null,
    "score_se": null,
    "speed": null,
    "speed_serve": 0.7402,
    "speeds": {
      "interactive": null,
      "long_context": null,
      "serve": 0.7402
    },
    "task_id": "qwen3-8b",
    "timed_audit": null
  },
  "runner": "kubernetes",
  "solution": {
    "checkpoint": "unsloth/Qwen3-8B-unsloth-bnb-4bit",
    "declared_precision": true,
    "detected_precision": "W4A16KV8",
    "method": "NF4 dynamic, Unsloth + FP8 KV",
    "precision": "W4A16KV8"
  },
  "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": "artifact",
  "updated_at": "2026-10-08T22:33:20Z"
}