task_slug string | cohort string | model string | reasoning_setting string | cumulative_token_budget int64 | actual_completed_tokens int64 | evaluated_source_feedback_rounds int64 | original_session_feedback_rounds int64 | source_modified bool | correctness_public_passed int64 | correctness_public_total int64 | correctness_private_passed int64 | correctness_private_total int64 | public_speedup float64 | private_speedup float64 | combined_speedup float64 | public_ci95 list | private_ci95 list | pairs_per_split int64 | measurement_qualified bool | outcome string | new_model_sample_for_this_cohort bool | notes list |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
sglang-mature-v9 | mature-starter-independent-score | gpt-6.1-sol | high | null | null | 0 | 1 | true | 56 | 56 | 168 | 168 | 1.003695 | 0.974702 | 0.982682 | null | null | 1 | false | independent_source_service_target_not_met | true | [
"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."
] |
sglang-mature-v9 | mature-starter-independent-score | qwen3.8-max | high | null | null | 0 | 0 | true | 56 | 56 | 168 | 168 | 0.972464 | 0.963016 | 0.965654 | null | null | 1 | false | independent_source_service_target_not_met | true | [
"Reference-relative observation; not five paired repeats.",
"Fresh zero-feedback native Codex trial.",
"Author reference qualification remains incomplete."
] |
multilora-windowed-v3 | sol-prefeedback-5pairs | gpt-6.1-sol | high | 1,000,000 | null | 0 | 1 | true | 6 | 6 | 18 | 18 | 1.88 | 2.27 | null | [
1.55,
2.27
] | [
1.7,
3.05
] | 5 | false | correct_but_performance_not_qualified | true | [
"Historical pre-first-feedback source; original session knew one feedback was available."
] |
multilora-windowed-v3 | qwen-1M | qwen3.8-max | medium (provider) | 1,000,000 | 924,815 | 0 | 0 | false | 6 | 6 | 18 | 18 | null | null | null | null | null | null | null | no_optimization_delivered | true | [
"All correctness is the unchanged provided baseline; not credited as model optimization."
] |
multilora-windowed-v3 | qwen-3M | qwen3.8-max | medium (provider) | 3,000,000 | 3,049,453 | 0 | 0 | true | 3 | 6 | 3 | 18 | null | null | null | null | null | null | null | correctness_failed | true | [
"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."
] |
multilora-windowed-v3 | qwen-6M-5pairs | qwen3.8-max | medium (provider) | 6,000,000 | 6,063,803 | 0 | 0 | true | 6 | 6 | 18 | 18 | 1.16 | 1.98 | null | [
0.81,
1.64
] | [
1.29,
3.04
] | 5 | false | correct_but_performance_not_qualified | true | [
"5832007 tokens delivered before boundary; provider-completed count includes final non-delivered response.",
"Correctness and resource gates passed; timing intervals too wide."
] |
multilora-windowed-v3 | sealed-source-diagnostic-10pairs | gpt-6.1-sol | high | 1,000,000 | null | 0 | null | true | 6 | 6 | 18 | 18 | 2.090454 | 1.912305 | null | [
1.5980091814410127,
2.7346521085873765
] | [
1.5541079473303154,
2.3530604397440555
] | 10 | false | measurement_control_invalid_for_model_ranking | false | [
"No new model call or feedback; both same-code controls failed qualification.",
"All correctness and resource checks passed; speedups are raw diagnostic observations."
] |
multilora-windowed-v3 | sealed-source-diagnostic-10pairs | qwen3.8-max | medium (provider) | 6,000,000 | null | 0 | null | true | 6 | 6 | 18 | 18 | 1.686121 | 1.480081 | null | [
1.1744323271009205,
2.420746803642401
] | [
1.1855998647023684,
1.8477063502306035
] | 10 | false | measurement_control_invalid_for_model_ranking | false | [
"No new model call or feedback; both same-code controls failed qualification.",
"All correctness and resource checks passed; speedups are raw diagnostic observations."
] |
cpu-ragged-gqa-v1 | sol-historical-prefeedback | gpt-6.1-sol | high | 180,000 | 72,022 | 0 | 0 | true | 15 | 15 | 45 | 45 | 13.96 | 13.99 | null | [
13.91,
14.01
] | [
13.96,
14.03
] | 5 | true | passed | true | [
"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."
] |
cpu-ragged-gqa-v1 | qwen-180K | qwen3.8-max | thinking; high mapped to xhigh | 180,000 | 161,596 | 0 | 0 | false | 15 | 15 | 45 | 45 | 0.9997 | 0.9955 | null | [
0.9865,
1.013
] | [
0.9859,
1.0052
] | 5 | true | no_optimization_delivered | true | [
"Sealed source equals original baseline; not an incorrect model patch."
] |
cpu-ragged-gqa-v1 | qwen-1M | qwen3.8-max | explicit thinking_budget=32768; provider maps to xhigh | 1,000,000 | 899,177 | 0 | 0 | true | 15 | 15 | 45 | 45 | 10.377583 | 10.360503 | null | [
10.306305702878458,
10.449353036315088
] | [
10.28635914119491,
10.435180849741363
] | 5 | true | passed | true | [
"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."
] |
sglang-n0-v9 | sol-prefeedback | gpt-6.1-sol | high | null | null | 0 | 1 | true | 56 | 56 | 168 | 168 | null | null | null | null | null | 1 | false | independent_source_service_target_not_met | true | [
"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."
] |
sglang-n0-v9 | sol-final-after-feedback | gpt-6.1-sol | high | null | null | 1 | 1 | true | 56 | 56 | 167 | 168 | 7.028 | 7.938 | null | null | null | 1 | false | independent_source_service_target_not_met | true | [
"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."
] |
sglang-n0-v9 | qwen-zero-sealed-pending-grade | qwen3.8-max | high | null | null | 0 | 0 | false | null | 56 | null | 168 | null | null | null | null | null | null | null | independent_grade_pending | true | [
"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."
] |
InfraOptBench：四个会话的性能优化题与模型对照
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。
| 来源会话 | 题包 | 当前结果范围 |
|---|---|---|
| 性能 (5) | SGLang 成熟起点 v9 | 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。 |
| 性能 (3) | 在线多 LoRA 窗口 v3 | 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。 |
| 继续 Qwen3.8 与 GPT 对比 | CPU Ragged-GQA | 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。 |
| 性能 (2) | SGLang N0 起点 v9 | 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。 |
详细报告：COMPARISON.zh-CN.md。结构化结果：results.jsonl。会话映射：tasks.json。原来源和许可说明：THIRD_PARTY_NOTICES.md。
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。
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。
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。
四题统一格式
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。
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。
- Downloads last month
- 63