[benchmark] Enhance benchmark metrics with ITL aggregation details

[ZhangYulongg]

Added metrics for cleaned ITL aggregation including decode speed and counts for different ITL categories.

Motivation

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Modifications

Usage or Command

Accuracy Tests

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[PaddlePaddle-bot]

🤖 Paddle-CI-Agent | pr_review | 2026-06-17 16:38:18

📋 Review 摘要

PR 概述：为 benchmark_serving.py 增加清洗后的 ITL 全局聚合解码速度及样本计数输出。
变更范围：benchmarks/benchmark_serving.py
影响面 Tag：[Benchmark]

问题

级别	文件	概述
🔴 Bug	benchmarks/benchmark_serving.py:304	clean ITL 聚合与主统计使用不同有效样本口径，可能把 output_tokens == 0 的响应计入新增指标
🔴 安全	benchmarks/benchmark_serving.py:471	warmup 成功路径无条件打印完整响应对象，可能泄露生成内容和 token ids

📝 PR 规范检查

标题 Tag 大小写不符合官方 Tag，PR 描述保留了模板占位内容且关键章节未填写。

标题建议（可直接复制）：

• [Benchmark] Add cleaned ITL aggregation metrics

PR 描述建议（点击展开，可直接复制）

## Motivation
Add benchmark metrics that report cleaned global ITL aggregation, so decode speed can be inspected after filtering burst and preemption intervals.

## Modifications
- Updated `benchmarks/benchmark_serving.py` to collect global ITL samples, count burst/preemption/clean intervals, and report clean decode speed.
- Added the clean ITL aggregate fields to benchmark JSON results and console output.

## Usage or Command
N/A

## Accuracy Tests
N/A

## Checklist

- [x] Add at least a tag in the PR title.
  - Tag list: [`[FDConfig]`,`[APIServer]`,`[Engine]`, `[Scheduler]`, `[PD Disaggregation]`, `[Executor]`, `[Graph Optimization]`, `[Speculative Decoding]`, `[RL]`, `[Models]`, `[Quantization]`, `[Loader]`, `[OP]`, `[KVCache]`, `[DataProcessor]`, `[BugFix]`, `[Docs]`, `[CI]`, `[Optimization]`, `[Feature]`, `[Benchmark]`, `[Others]`, `[XPU]`, `[HPU]`, `[GCU]`, `[DCU]`, `[Iluvatar]`, `[Metax]`]
  - You can add new tags based on the PR content, but the semantics must be clear.
- [ ] Format your code, run `pre-commit` before commit.
- [ ] Add unit tests. Please write the reason in this PR if no unit tests.
- [ ] Provide accuracy results.
- [ ] If the current PR is submitting to the `release` branch, make sure the PR has been submitted to the `develop` branch, then cherry-pick it to the `release` branch with the `[Cherry-Pick]` PR tag.

总体评价

新增指标方向合理，但当前实现存在统计口径不一致和日志暴露风险，建议修复后再合入。

[PaddlePaddle-bot]

🔴 Bug 新增 clean ITL 聚合只按 o.success 收集样本，会把主统计已经跳过的 output_tokens == 0 响应重新计入。

calculate_metrics() 前面的主循环在 output_len = outputs[i].output_tokens 后会对空输出直接 continue，因此这些请求不会进入 completed、total_output、普通 itl/s_decode 等指标。这里重新遍历 outputs 并只判断 o.success，会让 s_decode_clean、n_itls_* 与同一份结果中的主指标使用不同样本集；对某些流式后端，成功返回文本但没有 usage/completion_tokens 时，已有代码正是走这个跳过路径。

建议修复方式：复用主循环已经确认有效的 itls 样本，例如直接基于 itls 构造 _arr，或把 clean 聚合移动到 if not output_len: continue 之后，保证与 completed/total_output/普通 ITL 指标口径一致。

[PaddlePaddle-bot]

🔴 安全 这里把 warmup 成功响应也无条件打印到日志，会泄露完整 RequestFuncOutput。

RequestFuncOutput 的 repr 包含 generated_text、reasoning_content、output_ids、tool_calls 等字段。此前只在 warmup 失败时打印，主要用于排错；现在成功请求也会进入 stdout，在 benchmark 日志被采集或共享时会暴露生成内容和 token ids。

建议修复方式：恢复为仅失败时打印完整 test_output；如成功路径确实需要调试信息，请放到 if debug: 或只打印 request id、latency、output_tokens 等非内容字段。