Qwen3.8-27B-FP8 / validation /validation_metrics.json
Chungulus's picture
Add FP8 validation evidence
6550026 verified
Raw
History Blame Contribute Delete
972 Bytes
{
"target": "FP8",
"reference": "/mnt/c/Users/chunk/OneDrive/Documents/quant3.8/qwen38-cuda-quants/hf-cache/source/Qwen3.8-27B--1d4bf0f2ff60",
"quant": "/mnt/c/Users/chunk/OneDrive/Documents/quant3.8/qwen38-cuda-quants/outputs/Qwen3.8-27B-FP8",
"created_utc": "2026-08-17T03:15:49+00:00",
"smoke_test": false,
"thresholds": {
"max_mean_kl": 0.2,
"max_p95_kl": 0.8,
"min_top1": 0.6,
"min_top5_overlap": 0.6,
"min_logit_cosine": 0.9,
"max_ppl_delta_percent": 50.0
},
"token_positions": 64,
"kl_direction": "KL(P_BF16 || P_QUANT)",
"kl_mean": 0.005752270109951496,
"kl_median": 0.0024972036480903625,
"kl_p95": 0.01725013740360737,
"top1_agreement": 0.96875,
"top5_set_overlap": 0.9750000238418579,
"reference_nll": 3.4673221111297607,
"quant_nll": 3.4618234634399414,
"nll_delta": -0.005498647689819336,
"ppl_delta_percent": -0.5483557797280714,
"logit_cosine_similarity": 0.9966187477111816,
"passed": true
}