Commit ·
9618d00
1
Parent(s): f77cfc0
上传四个性能优化题包及 Sol 与 Qwen3.8-Max 对比 (#1)
Browse files- 上传四个性能优化题包及 Sol 与 Qwen3.8-Max 对比 (aeccf416af9498c08d6fb0044f4a6e834ed20e0b)
- 统一四个性能题的中文题面与题包格式 (3d4dab99b84a44c55b1d251970875e6321501dc0)
Co-authored-by: dsad <superhhhsss32@users.noreply.huggingface.co>
- COMPARISON.zh-CN.md +66 -0
- InfraOptBench-complete-20261109.zip +3 -0
- MANIFEST.json +60 -0
- PyTorch-LICENSE.txt +84 -0
- README.md +43 -0
- THIRD_PARTY_NOTICES.md +9 -0
- cpu-ragged-gqa-v1.zip +3 -0
- multilora-windowed-v3.zip +3 -0
- results.jsonl +14 -0
- sglang-mature-v9.zip +3 -0
- sglang-n0-v9.zip +3 -0
- tasks.json +30 -0
COMPARISON.zh-CN.md
ADDED
|
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# GPT-6.1-sol 与 Qwen3.8-Max:四个会话的实际差异
|
| 2 |
+
|
| 3 |
+
本报告汇总四个会话、三个机制题族,其中 SGLang 服务题分成熟 B 和直接实现 N0 两种起点。这里比较具体封存代码,不作模型总体能力排名。本次整理没有新增模型调用、候选运行或 GPU 试验。
|
| 4 |
+
|
| 5 |
+
## 题目是什么、来自哪里
|
| 6 |
+
|
| 7 |
+
| 会话 | 题目 | 系统原本做什么,需要优化什么 |
|
| 8 |
+
| --- | --- | --- |
|
| 9 |
+
| 性能 (5) | SGLang 成熟起点 v9 | 固定 7B 模型混合处理短问题和长文章;已有成熟引擎,需改善在线准入、KV/前缀复用、短请求延迟及整体完成时间。 |
|
| 10 |
+
| 性能 (3) | 在线版本化多 LoRA 窗口 v3 | 在有限显存中计算当前窗口的多适配器投影,发布真实结果后才能读取未来窗口;版本、秩、路由、缩放和历史输出须一致。 |
|
| 11 |
+
| 继续 Qwen3.8 与 GPT 对比 | CPU Ragged-GQA | 在单 CPU 核上加速变长分组注意力,保持因果、滑动窗口、固定前缀、softcap 与全部 values/logsumexp。 |
|
| 12 |
+
| 性能 (2) | SGLang N0 起点 v9 | 与成熟服务题同一冻结请求和服务目标,但实际从直接 PyTorch FIFO 实现起步,允许引入固定版本引擎和批处理;加速分母是实际 N0。 |
|
| 13 |
+
|
| 14 |
+
评测机制参考固定 infrabench pipeline。在线窗口、请求组合和故障/性能约束由本项目构造,不冒充上游原样缺陷。CPU 与 LoRA 题使用合成张量而非完整语言模型;SGLang 执行 DeepSeek-R1-Distill-Qwen-7B BF16。解题模型的实际服务标识为 gpt-6.1-sol、qwen3.8-max;未由 API 独立核实服务端参数规模。
|
| 15 |
+
|
| 16 |
+
## 主要结果
|
| 17 |
+
|
| 18 |
+
| 任务及所评源码 | Sol | Qwen3.8-Max | 当前可支持的结论 |
|
| 19 |
+
| --- | --- | --- | --- |
|
| 20 |
+
| SGLang 成熟 B,首反馈前源码/零反馈提交 | 请求检查 224/224;合计相对参考 0.983×;公开三项短请求 TTFT 超标 | 请求检查 224/224;相对参考 0.966×;公开只剩前缀突发 TTFT 超标 | Qwen 更重视短请求调度,短请求延迟更接近目标;总耗时约比 Sol 多 1.76%,两者整体目标均未通过。只有单次观察,不能断言稳定快慢。 |
|
| 21 |
+
| LoRA,Sol 首反馈前 / Qwen 600 万额度 | 正确性公开 6/6、私有 18/18 | 正确性公开 6/6、私有 18/18 | 两者最终都能产生正确实现;测速同代码对照未获资格,速度暂不能作排名或模型失败证据。 |
|
| 22 |
+
| CPU GQA,Sol 历史 / Qwen 100 万额度 | 60/60;公开 13.96×、私有 13.99×,通过 | 60/60;公开 10.38×、私有 10.36×,通过 | 在这道题的已测代码上 Sol 加速更大;Qwen 提高额度后也完整通过。预算及接口不同,不能推广为总体差距。 |
|
| 23 |
+
| SGLang N0 | 零反馈源码 224/224,但公开 TTFT 仍超标;反馈后代码 7.028×/7.938×,仍未过全部目标 | 严格 0 反馈、120 分钟后封存仍是原 N0,独立评分等待中 | 目前只能确认 Qwen 没有留下优化修改;没有本次实测加速,不能写成 1×、0 分或行为失败。 |
|
| 24 |
+
|
| 25 |
+
加速比大于 1 才表示比对应分母更快。SGLang 成熟起点使用独立参考 suite/候选 suite;N0 使用同集合 N0/候选;CPU 和 LoRA 使用各自起点 B/候选的配对几何均值。这些分母不同,不能跨行平均加速比。
|
| 26 |
+
|
| 27 |
+
## 反馈轮次和预算
|
| 28 |
+
|
| 29 |
+
| 试跑/源码 | 实际送入模型的外部评测反馈 | 累计 token 额度 | 关键区别 |
|
| 30 |
+
| --- | ---: | ---: | --- |
|
| 31 |
+
| SGLang 成熟 Sol 首反馈前源码 | 源码时点为 0;原会话收到 1 次设施错误反馈 | 沿用历史协议 | 前后四个源码相同;独立复测的是封存代码,不能把原设施异常算模型失败。 |
|
| 32 |
+
| SGLang 成熟 Qwen/high | 0 | 沿用该轮协议 | 全新原生 Codex 读/改/交;约 107.6 分钟后提交。 |
|
| 33 |
+
| LoRA Sol/high 首反馈前 | 源码时点 0 | 100 万 | 原会话知道可申请一次反馈;本次没重跑 Sol。 |
|
| 34 |
+
| LoRA Qwen/medium | 0 | 100 万、300 万、600 万分别独立 | 100 万没有交付修改;300 万确有补丁但算错;600 万正确性通过。不能混合择优成同一轮结果。 |
|
| 35 |
+
| CPU Sol/high 历史稿 | 0 | 18 万 | 调用过一次评测入口并占用额度,但结果未送进下一次模型请求;约 7.2 万实际 tokens、4.3 分钟封存。 |
|
| 36 |
+
| CPU Qwen 原额度 | 0 | 18 万 | 约 16.16 万实际 tokens、39.3 分钟,没有落盘修改。 |
|
| 37 |
+
| CPU Qwen 提高额度 | 0 | 100 万 | 实际 899177 tokens、约 41.8 分钟、五次代码修改;单次输出 131072、思考预算 32768。 |
|
| 38 |
+
| SGLang N0 Sol | 0 的源码快照 / 1 的最终源码分列 | 沿用历史协议 | 源码不同,成绩不能交叉代替;原接口超时和未完成正式资格继续保留。 |
|
| 39 |
+
| SGLang N0 Qwen/high | 0 | 7200 秒时限 | 四个文件与起始 N0 完全相同;终评尚无完整成绩。 |
|
| 40 |
+
|
| 41 |
+
CPU 100 万额度与 LoRA 600 万额度是另行授权的新试跑,旧成绩保留。LoRA 提供方累计完成用量还包含预算边界后未交付的末响应;600 万轮实际交付约 583 万、提供方完成约 606 万,末响应没有编辑或新增反馈。不能把不同 token、思考和工具配置描述为只替换模型的严格对照。
|
| 42 |
+
|
| 43 |
+
## 实现行为差异
|
| 44 |
+
|
| 45 |
+
SGLang 成熟起点中,Sol 主要把累计 token 传输改成增量传输,没有解决原 LPM 调度。Qwen 增加短请求优先、冷长请求准入、首 token 驱动的放行及防饥饿逻辑。公开交互混合短 TTFT 从 Sol 的 1.918 秒降到 1.300 秒,KV 压力从 6.674 秒降到 0.480 秒,前缀突发从 3.945 秒降到 1.456 秒;最后一项仍超过 1.150 秒。短请求先输出与整批先完成是不同目标;准入等待及主机同步可能影响批处理和尾部排空,但各机制的因果贡献尚无消融证明。
|
| 46 |
+
|
| 47 |
+
LoRA 的 Qwen 300 万轮至少存在一处明确逻辑错误:对权重版本相同、缩放系数不同的请求合批,却统一使用第一份缩放。逐例诊断只有公开 3/6、私有 3/18;性能未执行。600 万轮已通过全部正确性与资源检查;不能把该早期错误归给后来的正确候选。
|
| 48 |
+
|
| 49 |
+
CPU 的 Qwen 高额度候选把逐查询/逐头计算改成共享 KV 头的矩阵运算,并用分块在线 softmax 合并;这次确实产出了优化,而非只给计划。Sol 的历史候选在较少实际 tokens 下完成、加速更大,但两者工具、思考、上下文及输出限额不同。
|
| 50 |
+
|
| 51 |
+
## LoRA 测速为何不能作为稳定排名
|
| 52 |
+
|
| 53 |
+
最新只复测已封存代码,公私各十次配对,未重新调用模型:
|
| 54 |
+
|
| 55 |
+
| 实现 | 公开加速和 95% 区间 | 私有加速和 95% 区间 |
|
| 56 |
+
| --- | --- | --- |
|
| 57 |
+
| Sol 零反馈 | 2.09× [1.60, 2.73] | 1.91× [1.55, 2.35] |
|
| 58 |
+
| Qwen 600 万 | 1.69× [1.17, 2.42] | 1.48× [1.19, 1.85] |
|
| 59 |
+
|
| 60 |
+
同代码 B1/B2 对照公开区间 [0.44,2.95]、私有 [0.62,1.48],均不满足区间包含 1 且相对宽度 ≤30% 的资格要求。所有正确性和资源检查通过,但测量本身不能稳定区分相同实现。遥测支持 GPU 动态频率参与波动,尚未排除 CPU/桌面负载、短测量和观测开销。这里保留原始点估计作为诊断,不能把性能门未过直接算作模型解题失败。
|
| 61 |
+
|
| 62 |
+
## 发布和结论范围
|
| 63 |
+
|
| 64 |
+
每个目录包含完整题面、冻结起点、公开输入、接口和公开评分代码。参考优化、私有输入、API 凭据及模型完整轨迹未公开。四个目录是四个会话对应的交付视图,不是四个相互独立的随机题目。
|
| 65 |
+
|
| 66 |
+
CPU 高额度结果已有有效五次配对;GPU 题仍有测量/作者资格或 H800 补验缺口。SGLang N0 的 Qwen 评分待完成,后续只有读取到真实独立结果才能更新。现有证据支持“具体优化策略、交付效率及预算敏感性不同”,不支持把这几次探索性结果汇总成通用胜率。
|
InfraOptBench-complete-20261109.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0f67b63f20815a01221e0d927f7e4146c1ae5c0b05b87e592a2eaa6f171d0d4c
|
| 3 |
+
size 813195
|
MANIFEST.json
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema": 1,
|
| 3 |
+
"repo": "dabingzz/InfraOptBench",
|
| 4 |
+
"snapshot_date": "2026-11-09",
|
| 5 |
+
"task_packages": [
|
| 6 |
+
"sglang-mature-v9",
|
| 7 |
+
"multilora-windowed-v3",
|
| 8 |
+
"cpu-ragged-gqa-v1",
|
| 9 |
+
"sglang-n0-v9"
|
| 10 |
+
],
|
| 11 |
+
"result_records": 14,
|
| 12 |
+
"files": {
|
| 13 |
+
"COMPARISON.zh-CN.md": {
|
| 14 |
+
"bytes": 7785,
|
| 15 |
+
"sha256": "c95da0aa3086233e1c3947902777772129cc1e8f90be5fda80cee457a76ee053"
|
| 16 |
+
},
|
| 17 |
+
"cpu-ragged-gqa-v1.zip": {
|
| 18 |
+
"bytes": 69750,
|
| 19 |
+
"sha256": "4927607867679cd23c2ac9d646d78d76d1e50743493545f0a45213577bd98e54"
|
| 20 |
+
},
|
| 21 |
+
"InfraOptBench-complete-20261109.zip": {
|
| 22 |
+
"bytes": 813195,
|
| 23 |
+
"sha256": "0f67b63f20815a01221e0d927f7e4146c1ae5c0b05b87e592a2eaa6f171d0d4c"
|
| 24 |
+
},
|
| 25 |
+
"multilora-windowed-v3.zip": {
|
| 26 |
+
"bytes": 80887,
|
| 27 |
+
"sha256": "c6f432bcf48b3239e9ecf6e340d38977f977a9d85b948d50d8c23711bc4cdb9b"
|
| 28 |
+
},
|
| 29 |
+
"PyTorch-LICENSE.txt": {
|
| 30 |
+
"bytes": 3464,
|
| 31 |
+
"sha256": "a8a2c6b67cf8a597ffaa5c82380a6bd1b82629c0417e00673d8013d662eb500a"
|
| 32 |
+
},
|
| 33 |
+
"README.md": {
|
| 34 |
+
"bytes": 3178,
|
| 35 |
+
"sha256": "a366bc772964ea09c45e23c518a91128f67341d2a92cc2a7d6731febcc8caf15"
|
| 36 |
+
},
|
| 37 |
+
"results.jsonl": {
|
| 38 |
+
"bytes": 12781,
|
| 39 |
+
"sha256": "13b92a8477831b053597f8a2cd4451ce1bb24ed597f1c6ef6c66fa37420829cb"
|
| 40 |
+
},
|
| 41 |
+
"sglang-mature-v9.zip": {
|
| 42 |
+
"bytes": 316585,
|
| 43 |
+
"sha256": "1ea67199e0ce3d796e8349c24d449ec29f712b919091b8d52de60ed3284a049f"
|
| 44 |
+
},
|
| 45 |
+
"sglang-n0-v9.zip": {
|
| 46 |
+
"bytes": 319718,
|
| 47 |
+
"sha256": "5d711c5c010af07fa697b01536976b88c2781b87d30a1ab2ad8ecd3721124840"
|
| 48 |
+
},
|
| 49 |
+
"tasks.json": {
|
| 50 |
+
"bytes": 1080,
|
| 51 |
+
"sha256": "d7290b162f0225f799d51b273a07067b211dc17beaa33b9d1d704310fda4a77f"
|
| 52 |
+
},
|
| 53 |
+
"THIRD_PARTY_NOTICES.md": {
|
| 54 |
+
"bytes": 1274,
|
| 55 |
+
"sha256": "a418d3e9452b6b81cba91dad8a32bf61ffcfa4b57e183388762d5b92425b7521"
|
| 56 |
+
}
|
| 57 |
+
},
|
| 58 |
+
"format_revision": "2026-11-10",
|
| 59 |
+
"complete_snapshot_file_count": 133
|
| 60 |
+
}
|
PyTorch-LICENSE.txt
ADDED
|
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
From PyTorch:
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2016- Facebook, Inc (Adam Paszke)
|
| 4 |
+
Copyright (c) 2014- Facebook, Inc (Soumith Chintala)
|
| 5 |
+
Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
|
| 6 |
+
Copyright (c) 2012-2014 Deepmind Technologies (Koray Kavukcuoglu)
|
| 7 |
+
Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
|
| 8 |
+
Copyright (c) 2011-2013 NYU (Clement Farabet)
|
| 9 |
+
Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
|
| 10 |
+
Copyright (c) 2006 Idiap Research Institute (Samy Bengio)
|
| 11 |
+
Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
|
| 12 |
+
|
| 13 |
+
From Caffe2:
|
| 14 |
+
|
| 15 |
+
Copyright (c) 2016-present, Facebook Inc. All rights reserved.
|
| 16 |
+
|
| 17 |
+
All contributions by Facebook:
|
| 18 |
+
Copyright (c) 2016 Facebook Inc.
|
| 19 |
+
|
| 20 |
+
All contributions by Google:
|
| 21 |
+
Copyright (c) 2015 Google Inc.
|
| 22 |
+
All rights reserved.
|
| 23 |
+
|
| 24 |
+
All contributions by Yangqing Jia:
|
| 25 |
+
Copyright (c) 2015 Yangqing Jia
|
| 26 |
+
All rights reserved.
|
| 27 |
+
|
| 28 |
+
All contributions by Kakao Brain:
|
| 29 |
+
Copyright 2019-2020 Kakao Brain
|
| 30 |
+
|
| 31 |
+
All contributions by Cruise LLC:
|
| 32 |
+
Copyright (c) 2022 Cruise LLC.
|
| 33 |
+
All rights reserved.
|
| 34 |
+
|
| 35 |
+
All contributions by Tri Dao:
|
| 36 |
+
Copyright (c) 2024 Tri Dao.
|
| 37 |
+
All rights reserved.
|
| 38 |
+
|
| 39 |
+
All contributions by Arm:
|
| 40 |
+
Copyright (c) 2021, 2023-2024 Arm Limited and/or its affiliates
|
| 41 |
+
|
| 42 |
+
All contributions from Caffe:
|
| 43 |
+
Copyright(c) 2013, 2014, 2015, the respective contributors
|
| 44 |
+
All rights reserved.
|
| 45 |
+
|
| 46 |
+
All other contributions:
|
| 47 |
+
Copyright(c) 2015, 2016 the respective contributors
|
| 48 |
+
All rights reserved.
|
| 49 |
+
|
| 50 |
+
Caffe2 uses a copyright model similar to Caffe: each contributor holds
|
| 51 |
+
copyright over their contributions to Caffe2. The project versioning records
|
| 52 |
+
all such contribution and copyright details. If a contributor wants to further
|
| 53 |
+
mark their specific copyright on a particular contribution, they should
|
| 54 |
+
indicate their copyright solely in the commit message of the change when it is
|
| 55 |
+
committed.
|
| 56 |
+
|
| 57 |
+
All rights reserved.
|
| 58 |
+
|
| 59 |
+
Redistribution and use in source and binary forms, with or without
|
| 60 |
+
modification, are permitted provided that the following conditions are met:
|
| 61 |
+
|
| 62 |
+
1. Redistributions of source code must retain the above copyright
|
| 63 |
+
notice, this list of conditions and the following disclaimer.
|
| 64 |
+
|
| 65 |
+
2. Redistributions in binary form must reproduce the above copyright
|
| 66 |
+
notice, this list of conditions and the following disclaimer in the
|
| 67 |
+
documentation and/or other materials provided with the distribution.
|
| 68 |
+
|
| 69 |
+
3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America
|
| 70 |
+
and IDIAP Research Institute nor the names of its contributors may be
|
| 71 |
+
used to endorse or promote products derived from this software without
|
| 72 |
+
specific prior written permission.
|
| 73 |
+
|
| 74 |
+
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
|
| 75 |
+
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
| 76 |
+
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
| 77 |
+
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
|
| 78 |
+
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
|
| 79 |
+
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
|
| 80 |
+
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
|
| 81 |
+
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
|
| 82 |
+
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
|
| 83 |
+
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
|
| 84 |
+
POSSIBILITY OF SUCH DAMAGE.
|
README.md
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- zh
|
| 4 |
+
- en
|
| 5 |
+
pretty_name: InfraOptBench conversation task packages and Sol–Qwen comparisons
|
| 6 |
+
task_categories:
|
| 7 |
+
- other
|
| 8 |
+
tags:
|
| 9 |
+
- ai-infrastructure
|
| 10 |
+
- code-optimization
|
| 11 |
+
- benchmark
|
| 12 |
+
- zero-feedback
|
| 13 |
+
configs:
|
| 14 |
+
- config_name: model_results
|
| 15 |
+
data_files:
|
| 16 |
+
- split: results
|
| 17 |
+
path: results.jsonl
|
| 18 |
+
---
|
| 19 |
+
|
| 20 |
+
# InfraOptBench:四个会话的性能优化题与模型对照
|
| 21 |
+
|
| 22 |
+
本数据集收录用户指定的四个会话的完整公开题包,以及 **GPT-6.1-sol 与 Qwen3.8-Max 的已核验成绩对比**。四个目录覆盖三个机制题族;SGLang 的成熟起点和 N0 起点分别保留,不能合并。
|
| 23 |
+
|
| 24 |
+
| 来源会话 | 题包 | 当前结果范围 |
|
| 25 |
+
| --- | --- | --- |
|
| 26 |
+
| 性能 (5) | [SGLang 成熟起点 v9](sglang-mature-v9.zip) | 两者请求正确,Qwen 短请求延迟更好、总耗时略慢;整体目标均未过。单次观察、资格待补。 |
|
| 27 |
+
| 性能 (3) | [在线多 LoRA 窗口 v3](multilora-windowed-v3.zip) | Sol 和 Qwen 600 万轮正确性全部通过;测速控制未获资格,原始提速仅作诊断。 |
|
| 28 |
+
| 继续 Qwen3.8 与 GPT 对比 | [CPU Ragged-GQA](cpu-ragged-gqa-v1.zip) | Sol 13.96×/13.99×,Qwen 高额度 10.38×/10.36×,两者 60/60 且性能通过。额度不同。 |
|
| 29 |
+
| 性能 (2) | [SGLang N0 起点 v9](sglang-n0-v9.zip) | Sol 代码有加速但未过全部目标;Qwen 未修改、独立评分待完成,不填虚构成绩。 |
|
| 30 |
+
|
| 31 |
+
详细报告:[COMPARISON.zh-CN.md](COMPARISON.zh-CN.md)。结构化结果:[results.jsonl](results.jsonl)。会话映射:[tasks.json](tasks.json)。原来源和许可说明:[THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md)。
|
| 32 |
+
|
| 33 |
+
每题的 TASK.md 给出完整题面;contract.json、tests/ 和 solution/ 保留对应冻结版本。历史预算/工具差异和评分限制在 reports/ 说明。报告、历史协议及模型结果不提供给新的解题会话;新试跑默认严格零反馈,模型仅读公开材料、编辑和提交,评分在封存关闭后独立执行。
|
| 34 |
+
|
| 35 |
+
这是任务源码与已有观测的发布,不是仅凭克隆即可复现所有历史服务器隔离和评分的机器镜像。需准备各题固定 CPU/GPU、软件版本和(SGLang)本地模型权重;不同环境须重新资格验证。没有在本次整理中新增模型或 GPU 实验。
|
| 36 |
+
|
| 37 |
+
不能把不同基线、不同预算、不同工具或未获测量资格的结果合并成严格模型排名。N0 的待评分项保持为空;同代码噪声不会计作模型错误。数据集尚未统一宣称正式准入,GPU H800 资格仍待完成。
|
| 38 |
+
|
| 39 |
+
## 四题统一格式
|
| 40 |
+
|
| 41 |
+
四个包的 TASK.md / TASK.zh.md 均使用相同的十节完整中文题面:目标、起点与环境、接口输入输出、正确性、可改范围、负载与日志、计时评分、预算、反馈交卷、文件导航。README.md 采用相同目录说明。旧题面保留在 historical/,起点、用例、评分代码、合同、协议和成绩均未改变。
|
| 42 |
+
|
| 43 |
+
网页以四个独立 ZIP 提供题包。完整快照 InfraOptBench-complete-20261109.zip 保留 133 个当前发布文件及完整目录;结果数据快照日期保留 2026-11-09,题面格式更新日期为 2026-11-10。
|
THIRD_PARTY_NOTICES.md
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 第三方来源
|
| 2 |
+
|
| 3 |
+
这些题目和案例由本项目整理,源代码保留其已有版权和许可声明;未为无许可声明文件额外指定统一授权。
|
| 4 |
+
|
| 5 |
+
- 公共评测机制来自 [Dabingzz/infrabench](https://github.com/Dabingzz/infrabench),固定修订和文件摘要见 provenance/pipeline-source.json。在线 LoRA 窗口、混合请求服务约束及本文对照是本项目扩展。
|
| 6 |
+
- CPU GQA 数学来源参考固定 [PyTorch](https://github.com/pytorch/pytorch) 修订,见 provenance/pytorch-source.json;PyTorch BSD 许可见 LICENSES/PyTorch-LICENSE.txt。本包不复制完整 PyTorch 源码。
|
| 7 |
+
- SGLang 公共请求来自 [openai/gsm8k](https://huggingface.co/datasets/openai/gsm8k) 与 [allenai/qasper](https://huggingface.co/datasets/allenai/qasper),确切修订及来源见 provenance/datasets.json。上游问题、文章和数据集许可归原权利人,本项目不重新授权其文本。
|
| 8 |
+
- 被测服务使用固定 [DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B);模型权重和 tokenizer 文件不随包上传。
|
| 9 |
+
- Torch、NumPy、SGLang、Transformers、CUDA 依赖由使用者按原许可另行安装。HF 仓库不含这些依赖的二进制、访问凭据或运行环境镜像。
|
cpu-ragged-gqa-v1.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4927607867679cd23c2ac9d646d78d76d1e50743493545f0a45213577bd98e54
|
| 3 |
+
size 69750
|
multilora-windowed-v3.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c6f432bcf48b3239e9ecf6e340d38977f977a9d85b948d50d8c23711bc4cdb9b
|
| 3 |
+
size 80887
|
results.jsonl
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"task_slug": "sglang-mature-v9", "cohort": "mature-starter-independent-score", "model": "gpt-6.1-sol", "reasoning_setting": "high", "cumulative_token_budget": null, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 1, "source_modified": true, "correctness_public_passed": 56, "correctness_public_total": 56, "correctness_private_passed": 168, "correctness_private_total": 168, "public_speedup": 1.0036953695191964, "private_speedup": 0.9747015617213634, "combined_speedup": 0.9826824361859144, "public_ci95": null, "private_ci95": null, "pairs_per_split": 1, "measurement_qualified": false, "outcome": "independent_source_service_target_not_met", "new_model_sample_for_this_cohort": true, "notes": ["Reference-relative observation; not five paired repeats.", "Sol original session received a facility-only feedback; evaluated pre-feedback and final sources are equal.", "Author reference qualification remains incomplete."]}
|
| 2 |
+
{"task_slug": "sglang-mature-v9", "cohort": "mature-starter-independent-score", "model": "qwen3.8-max", "reasoning_setting": "high", "cumulative_token_budget": null, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": true, "correctness_public_passed": 56, "correctness_public_total": 56, "correctness_private_passed": 168, "correctness_private_total": 168, "public_speedup": 0.9724643576740517, "private_speedup": 0.9630161063605914, "combined_speedup": 0.9656538535714588, "public_ci95": null, "private_ci95": null, "pairs_per_split": 1, "measurement_qualified": false, "outcome": "independent_source_service_target_not_met", "new_model_sample_for_this_cohort": true, "notes": ["Reference-relative observation; not five paired repeats.", "Fresh zero-feedback native Codex trial.", "Author reference qualification remains incomplete."]}
|
| 3 |
+
{"task_slug": "multilora-windowed-v3", "cohort": "sol-prefeedback-5pairs", "model": "gpt-6.1-sol", "reasoning_setting": "high", "cumulative_token_budget": 1000000, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 1, "source_modified": true, "correctness_public_passed": 6, "correctness_public_total": 6, "correctness_private_passed": 18, "correctness_private_total": 18, "public_speedup": 1.88, "private_speedup": 2.27, "combined_speedup": null, "public_ci95": [1.55, 2.27], "private_ci95": [1.7, 3.05], "pairs_per_split": 5, "measurement_qualified": false, "outcome": "correct_but_performance_not_qualified", "new_model_sample_for_this_cohort": true, "notes": ["Historical pre-first-feedback source; original session knew one feedback was available."]}
|
| 4 |
+
{"task_slug": "multilora-windowed-v3", "cohort": "qwen-1M", "model": "qwen3.8-max", "reasoning_setting": "medium (provider)", "cumulative_token_budget": 1000000, "actual_completed_tokens": 924815, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": false, "correctness_public_passed": 6, "correctness_public_total": 6, "correctness_private_passed": 18, "correctness_private_total": 18, "public_speedup": null, "private_speedup": null, "combined_speedup": null, "public_ci95": null, "private_ci95": null, "pairs_per_split": null, "measurement_qualified": null, "outcome": "no_optimization_delivered", "new_model_sample_for_this_cohort": true, "notes": ["All correctness is the unchanged provided baseline; not credited as model optimization."]}
|
| 5 |
+
{"task_slug": "multilora-windowed-v3", "cohort": "qwen-3M", "model": "qwen3.8-max", "reasoning_setting": "medium (provider)", "cumulative_token_budget": 3000000, "actual_completed_tokens": 3049453, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": true, "correctness_public_passed": 3, "correctness_public_total": 6, "correctness_private_passed": 3, "correctness_private_total": 18, "public_speedup": null, "private_speedup": null, "combined_speedup": null, "public_ci95": null, "private_ci95": null, "pairs_per_split": null, "measurement_qualified": null, "outcome": "correctness_failed", "new_model_sample_for_this_cohort": true, "notes": ["Per-case diagnostic counts; performance was not run.", "2881650 tokens delivered before boundary; provider-completed usage includes the non-delivered final response.", "At least one confirmed issue: batching used the first request scale for other scales."]}
|
| 6 |
+
{"task_slug": "multilora-windowed-v3", "cohort": "qwen-6M-5pairs", "model": "qwen3.8-max", "reasoning_setting": "medium (provider)", "cumulative_token_budget": 6000000, "actual_completed_tokens": 6063803, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": true, "correctness_public_passed": 6, "correctness_public_total": 6, "correctness_private_passed": 18, "correctness_private_total": 18, "public_speedup": 1.16, "private_speedup": 1.98, "combined_speedup": null, "public_ci95": [0.81, 1.64], "private_ci95": [1.29, 3.04], "pairs_per_split": 5, "measurement_qualified": false, "outcome": "correct_but_performance_not_qualified", "new_model_sample_for_this_cohort": true, "notes": ["5832007 tokens delivered before boundary; provider-completed count includes final non-delivered response.", "Correctness and resource gates passed; timing intervals too wide."]}
|
| 7 |
+
{"task_slug": "multilora-windowed-v3", "cohort": "sealed-source-diagnostic-10pairs", "model": "gpt-6.1-sol", "reasoning_setting": "high", "cumulative_token_budget": 1000000, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": null, "source_modified": true, "correctness_public_passed": 6, "correctness_public_total": 6, "correctness_private_passed": 18, "correctness_private_total": 18, "public_speedup": 2.0904542993257835, "private_speedup": 1.9123048736942558, "combined_speedup": null, "public_ci95": [1.5980091814410127, 2.7346521085873765], "private_ci95": [1.5541079473303154, 2.3530604397440555], "pairs_per_split": 10, "measurement_qualified": false, "outcome": "measurement_control_invalid_for_model_ranking", "new_model_sample_for_this_cohort": false, "notes": ["No new model call or feedback; both same-code controls failed qualification.", "All correctness and resource checks passed; speedups are raw diagnostic observations."]}
|
| 8 |
+
{"task_slug": "multilora-windowed-v3", "cohort": "sealed-source-diagnostic-10pairs", "model": "qwen3.8-max", "reasoning_setting": "medium (provider)", "cumulative_token_budget": 6000000, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": null, "source_modified": true, "correctness_public_passed": 6, "correctness_public_total": 6, "correctness_private_passed": 18, "correctness_private_total": 18, "public_speedup": 1.6861207850933633, "private_speedup": 1.4800812135971155, "combined_speedup": null, "public_ci95": [1.1744323271009205, 2.420746803642401], "private_ci95": [1.1855998647023684, 1.8477063502306035], "pairs_per_split": 10, "measurement_qualified": false, "outcome": "measurement_control_invalid_for_model_ranking", "new_model_sample_for_this_cohort": false, "notes": ["No new model call or feedback; both same-code controls failed qualification.", "All correctness and resource checks passed; speedups are raw diagnostic observations."]}
|
| 9 |
+
{"task_slug": "cpu-ragged-gqa-v1", "cohort": "sol-historical-prefeedback", "model": "gpt-6.1-sol", "reasoning_setting": "high", "cumulative_token_budget": 180000, "actual_completed_tokens": 72022, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": true, "correctness_public_passed": 15, "correctness_public_total": 15, "correctness_private_passed": 45, "correctness_private_total": 45, "public_speedup": 13.96, "private_speedup": 13.99, "combined_speedup": null, "public_ci95": [13.91, 14.01], "private_ci95": [13.96, 14.03], "pairs_per_split": 5, "measurement_qualified": true, "outcome": "passed", "new_model_sample_for_this_cohort": true, "notes": ["One evaluator task invoked/charged, but result was not delivered to a next model request.", "Historical API file-tool loop, not a strict matched-protocol fresh comparison."]}
|
| 10 |
+
{"task_slug": "cpu-ragged-gqa-v1", "cohort": "qwen-180K", "model": "qwen3.8-max", "reasoning_setting": "thinking; high mapped to xhigh", "cumulative_token_budget": 180000, "actual_completed_tokens": 161596, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": false, "correctness_public_passed": 15, "correctness_public_total": 15, "correctness_private_passed": 45, "correctness_private_total": 45, "public_speedup": 0.9997, "private_speedup": 0.9955, "combined_speedup": null, "public_ci95": [0.9865, 1.013], "private_ci95": [0.9859, 1.0052], "pairs_per_split": 5, "measurement_qualified": true, "outcome": "no_optimization_delivered", "new_model_sample_for_this_cohort": true, "notes": ["Sealed source equals original baseline; not an incorrect model patch."]}
|
| 11 |
+
{"task_slug": "cpu-ragged-gqa-v1", "cohort": "qwen-1M", "model": "qwen3.8-max", "reasoning_setting": "explicit thinking_budget=32768; provider maps to xhigh", "cumulative_token_budget": 1000000, "actual_completed_tokens": 899177, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": true, "correctness_public_passed": 15, "correctness_public_total": 15, "correctness_private_passed": 45, "correctness_private_total": 45, "public_speedup": 10.377582897262952, "private_speedup": 10.360502783347878, "combined_speedup": null, "public_ci95": [10.306305702878458, 10.449353036315088], "private_ci95": [10.28635914119491, 10.435180849741363], "pairs_per_split": 5, "measurement_qualified": true, "outcome": "passed", "new_model_sample_for_this_cohort": true, "notes": ["Fresh separate trial; five saved code edits; source auto-sealed at budget boundary.", "Different token and output allowances from historical Sol; not a controlled model-only ranking."]}
|
| 12 |
+
{"task_slug": "sglang-n0-v9", "cohort": "sol-prefeedback", "model": "gpt-6.1-sol", "reasoning_setting": "high", "cumulative_token_budget": null, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 1, "source_modified": true, "correctness_public_passed": 56, "correctness_public_total": 56, "correctness_private_passed": 168, "correctness_private_total": 168, "public_speedup": null, "private_speedup": null, "combined_speedup": null, "public_ci95": null, "private_ci95": null, "pairs_per_split": 1, "measurement_qualified": false, "outcome": "independent_source_service_target_not_met", "new_model_sample_for_this_cohort": true, "notes": ["Pre-first-feedback source; public KV-pressure TTFT 0.739s exceeded 0.723s.", "Speedups for the different final source are reported separately.", "Original API timeout record remains a facility limitation, not a model failure."]}
|
| 13 |
+
{"task_slug": "sglang-n0-v9", "cohort": "sol-final-after-feedback", "model": "gpt-6.1-sol", "reasoning_setting": "high", "cumulative_token_budget": null, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 1, "original_session_feedback_rounds": 1, "source_modified": true, "correctness_public_passed": 56, "correctness_public_total": 56, "correctness_private_passed": 167, "correctness_private_total": 168, "public_speedup": 7.028, "private_speedup": 7.938, "combined_speedup": null, "public_ci95": null, "private_ci95": null, "pairs_per_split": 1, "measurement_qualified": false, "outcome": "independent_source_service_target_not_met", "new_model_sample_for_this_cohort": true, "notes": ["Speedup relative to actual N0, paired once per split.", "Public prefix TTFT and one private numeric request missed original gates; not a full formal model-trial result."]}
|
| 14 |
+
{"task_slug": "sglang-n0-v9", "cohort": "qwen-zero-sealed-pending-grade", "model": "qwen3.8-max", "reasoning_setting": "high", "cumulative_token_budget": null, "actual_completed_tokens": null, "evaluated_source_feedback_rounds": 0, "original_session_feedback_rounds": 0, "source_modified": false, "correctness_public_passed": null, "correctness_public_total": 56, "correctness_private_passed": null, "correctness_private_total": 168, "public_speedup": null, "private_speedup": null, "combined_speedup": null, "public_ci95": null, "private_ci95": null, "pairs_per_split": null, "measurement_qualified": null, "outcome": "independent_grade_pending", "new_model_sample_for_this_cohort": true, "notes": ["120-minute budget exhausted; sealed four files equal exact N0.", "Initial public N0 log was available and actually read.", "No measured speedup or passing/failing counts yet; do not fill 1.0 or classify as a behavioral failure."]}
|
sglang-mature-v9.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1ea67199e0ce3d796e8349c24d449ec29f712b919091b8d52de60ed3284a049f
|
| 3 |
+
size 316585
|
sglang-n0-v9.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5d711c5c010af07fa697b01536976b88c2781b87d30a1ab2ad8ecd3721124840
|
| 3 |
+
size 319718
|
tasks.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"task_slug": "sglang-mature-v9",
|
| 4 |
+
"conversation": "性能 (5)",
|
| 5 |
+
"conversation_url": "codex://threads/01a1b15a-68ec-7c92-a674-a4e3e4025c09",
|
| 6 |
+
"archive": "sglang-mature-v9.zip",
|
| 7 |
+
"folder_inside_complete_snapshot": "tasks/sglang-mature-v9"
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"task_slug": "multilora-windowed-v3",
|
| 11 |
+
"conversation": "性能 (3)",
|
| 12 |
+
"conversation_url": "codex://threads/01a1b025-0706-70b0-81da-445f5387c16c",
|
| 13 |
+
"archive": "multilora-windowed-v3.zip",
|
| 14 |
+
"folder_inside_complete_snapshot": "tasks/multilora-windowed-v3"
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"task_slug": "cpu-ragged-gqa-v1",
|
| 18 |
+
"conversation": "继续 Qwen3.8 与 GPT 对比",
|
| 19 |
+
"conversation_url": "codex://threads/01a1bebe-e308-7cc2-b3c9-367541364ebf",
|
| 20 |
+
"archive": "cpu-ragged-gqa-v1.zip",
|
| 21 |
+
"folder_inside_complete_snapshot": "tasks/cpu-ragged-gqa-v1"
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"task_slug": "sglang-n0-v9",
|
| 25 |
+
"conversation": "性能 (2)",
|
| 26 |
+
"conversation_url": "codex://threads/01a19bd9-82da-7a80-b678-484587334375",
|
| 27 |
+
"archive": "sglang-n0-v9.zip",
|
| 28 |
+
"folder_inside_complete_snapshot": "tasks/sglang-n0-v9"
|
| 29 |
+
}
|
| 30 |
+
]
|