Shar137/two / evaluation /BENCH_STATICFP8.json
Shar137's picture
download
raw
6.51 kB
{
"tier": "STATICFP8",
"pkg": "Coder390-static-fp8-block128-mm-mtp-v1",
"draft": "draft-dflash2-staticfp8-v1",
"gpu": "单张 RTX PRO 6000 Blackwell 96GB",
"server": "SGLang,context 32768,max-running-requests 16,mamba-ssm-dtype bfloat16",
"load": "生成上限 4096,开思考,采样用模型 generation_config 默认值",
"spec_args": {
"mtp": "NEXTN(等价 EAGLE),3 步,topk 1,4 个草稿 token,草稿=包内 BF16 MTP",
"dflash2": "DFLASH,草稿 draft-dflash2-staticfp8-v1,8 个草稿 token,block 8",
"nospec": "不开投机"
},
"rows": [
{
"spec": "mtp",
"concurrency": 1,
"requests": 4,
"wall_s": 79.2,
"total_tokens": 6507,
"aggregate_tps": 82.1,
"per_request_decode_tps_median": 91.1,
"ttft_median_s": 0.12,
"accept_len": 2.858,
"accept_rate": 0.619,
"speedup_vs_nospec": 1.82,
"source": "主测"
},
{
"spec": "mtp",
"concurrency": 4,
"requests": 24,
"wall_s": 54.5,
"total_tokens": 15982,
"aggregate_tps": 293.3,
"per_request_decode_tps_median": 84.0,
"ttft_median_s": 0.15,
"accept_len": 2.844,
"accept_rate": 0.615,
"latency_p50_s": 6.8,
"latency_max_s": 28.4,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 1.96
},
{
"spec": "mtp",
"concurrency": 8,
"requests": 48,
"wall_s": 95.3,
"total_tokens": 36240,
"aggregate_tps": 380.3,
"per_request_decode_tps_median": 71.1,
"ttft_median_s": 0.16,
"accept_len": 2.813,
"accept_rate": 0.604,
"latency_p50_s": 5.9,
"latency_max_s": 61.6,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 1.34
},
{
"spec": "mtp",
"concurrency": 16,
"requests": 96,
"wall_s": 117.2,
"total_tokens": 71779,
"aggregate_tps": 612.7,
"per_request_decode_tps_median": 60.1,
"ttft_median_s": 0.17,
"accept_len": 2.747,
"accept_rate": 0.582,
"latency_p50_s": 8.2,
"latency_max_s": 65.6,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 1.45
},
{
"spec": "dflash2",
"concurrency": 1,
"requests": 4,
"wall_s": 28.6,
"total_tokens": 4196,
"aggregate_tps": 146.7,
"per_request_decode_tps_median": 180.3,
"ttft_median_s": 0.12,
"accept_len": 4.289,
"accept_rate": 0.47,
"speedup_vs_nospec": 3.25,
"source": "主测"
},
{
"spec": "dflash2",
"concurrency": 4,
"requests": 24,
"wall_s": 45.5,
"total_tokens": 17512,
"aggregate_tps": 384.9,
"per_request_decode_tps_median": 124.4,
"ttft_median_s": 0.16,
"accept_len": 3.982,
"accept_rate": 0.425,
"latency_p50_s": 3.7,
"latency_max_s": 33.8,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 2.57
},
{
"spec": "dflash2",
"concurrency": 8,
"requests": 48,
"wall_s": 51.4,
"total_tokens": 33224,
"aggregate_tps": 646.3,
"per_request_decode_tps_median": 113.9,
"ttft_median_s": 0.15,
"accept_len": 4.131,
"accept_rate": 0.447,
"latency_p50_s": 4.3,
"latency_max_s": 36.9,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 2.28
},
{
"spec": "dflash2",
"concurrency": 16,
"requests": 96,
"wall_s": 64.4,
"total_tokens": 68093,
"aggregate_tps": 1057.5,
"per_request_decode_tps_median": 93.6,
"ttft_median_s": 0.24,
"accept_len": 4.017,
"accept_rate": 0.431,
"latency_p50_s": 5.9,
"latency_max_s": 41.8,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 2.5
},
{
"spec": "nospec",
"concurrency": 1,
"requests": 4,
"wall_s": 125.2,
"total_tokens": 5648,
"aggregate_tps": 45.1,
"per_request_decode_tps_median": 45.6,
"ttft_median_s": 0.11,
"accept_len": null,
"accept_rate": null,
"speedup_vs_nospec": 1.0,
"source": "主测"
},
{
"spec": "nospec",
"concurrency": 4,
"requests": 24,
"wall_s": 113.2,
"total_tokens": 16919,
"aggregate_tps": 149.5,
"per_request_decode_tps_median": 42.2,
"ttft_median_s": 0.13,
"accept_len": null,
"accept_rate": null,
"latency_p50_s": 11.2,
"latency_max_s": 98.0,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 1.0
},
{
"spec": "nospec",
"concurrency": 8,
"requests": 48,
"wall_s": 112.3,
"total_tokens": 31770,
"aggregate_tps": 283.0,
"per_request_decode_tps_median": 41.7,
"ttft_median_s": 0.12,
"accept_len": null,
"accept_rate": null,
"latency_p50_s": 10.1,
"latency_max_s": 81.6,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 1.0
},
{
"spec": "nospec",
"concurrency": 16,
"requests": 96,
"wall_s": 163.3,
"total_tokens": 69049,
"aggregate_tps": 422.8,
"per_request_decode_tps_median": 38.4,
"ttft_median_s": 0.13,
"accept_len": null,
"accept_rate": null,
"latency_p50_s": 11.5,
"latency_max_s": 105.4,
"source": "复测(6×C 请求)",
"speedup_vs_nospec": 1.0
}
],
"best": {
"mtp": {
"max_aggregate_tps_at": 16,
"max_aggregate_tps": 612.7,
"recommended_concurrency": 16,
"recommended_rule": "单请求速度≥C1 的 60% 时总吞吐最高",
"max_per_request_tps_at": 1,
"c1_per_request_tps": 91.1
},
"dflash2": {
"max_aggregate_tps_at": 16,
"max_aggregate_tps": 1057.5,
"recommended_concurrency": 8,
"recommended_rule": "单请求速度≥C1 的 60% 时总吞吐最高",
"max_per_request_tps_at": 1,
"c1_per_request_tps": 180.3
},
"nospec": {
"max_aggregate_tps_at": 16,
"max_aggregate_tps": 422.8,
"recommended_concurrency": 16,
"recommended_rule": "单请求速度≥C1 的 60% 时总吞吐最高",
"max_per_request_tps_at": 1,
"c1_per_request_tps": 45.6
}
},
"note": "MTP 与 DFLASH2 互斥,一次启动只能选一种。C1 为主测,C4/C8/C16 为加大请求数后的复测;复测前主测 DFLASH2 C8 受长尾请求影响偏低(456 tok/s),已弃用。"
}

Xet Storage Details

Size:
6.51 kB
·
Xet hash:
ac90f22c2e5770ee0d41a6940c0329fa01059d95358ba6a18875b6240ef0d831

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.