跳转至

Multi-dimensional Model Benchmarking / 多维度模型性能基准 / 实验 6-8

English

This directory contains two layers. demo.py is the short interactive sampler; campaign.py is the complete, resumable Experiment 6-8 campaign. The latter supports OpenAI-compatible, native Anthropic, and native Gemini APIs and records every real request in SQLite.

The complete campaign supports: - the 8K / 32K / 128K input × 512 / 2048 output workload matrix; - at least 100 requests per provider/model/cell; - exact TTFT, end-to-end latency, thinking TTFT, reported reasoning length, token usage, cache hits, and error classification; - a 168-hour hourly availability monitor with outage duration, MTTR, and longest continuous availability analysis; - a measured concurrency ramp with RPM and input/output TPM saturation; - cached/input/output pricing and a six-round Agent cost trace; - an explicit same-model/different-provider comparison group (DeepSeek V4 Flash on the official DeepSeek API and SiliconFlow); - resumability through unique request cells in SQLite and a strict completion audit.

The quick sampler still supports: - Concurrency stress testing: sweep concurrency to identify rate limits and observe metric curves. - Offline mock mode (--mock): synthetic data pipeline for verifying aggregation logic without API keys/network.

Synthetic mode is never accepted by campaign.py and cannot populate the official evidence database.

Full campaign

Review campaign_config.json before a cost-sensitive run. Pricing fields are deliberately explicit; analysis.py refuses to declare the campaign complete while any cached/input/output rate is missing.

# Real API integration smoke (small scope is visibly labelled)
python campaign.py workload --smoke --requests 1 \
  --context-tokens 256 --output-tokens 128 \
  --provider 'Ark/doubao-seed-1.6' \
  --campaign-id integration-smoke --db results/integration-smoke.sqlite3

# Official standardized workload: defaults are 3 contexts × 2 outputs × N=100
python campaign.py workload --campaign-id release-2026-07

# Actual RPM/TPM ramp
python campaign.py rate-limit --campaign-id release-2026-07

# Six-turn stable-prefix Agent cost/cache trace
python campaign.py agent-cost --campaign-id release-2026-07

# Keep this process alive for one week; every hourly cell is resumable
python campaign.py availability --duration-hours 168 --interval-seconds 3600 \
  --campaign-id release-2026-07

python analysis.py --campaign-id release-2026-07

analysis.py writes JSON and Markdown reports and prints Official completion: True only after every manuscript requirement has direct database evidence. A smoke run is useful validation, but it can never satisfy the 100-request or 168-hour gates.

The completion audit requires every configured provider—not merely one working provider—to have the complete workload, 169 hourly boundary probes spanning 168 hours, rate ramp, and six-round cost trace. Pricing is accepted only when input, cached-input, and output rates are all pinned together with an authoritative source_url and as_of date. Prices remain in the provider's published native currency. A non-USD price also requires usd_per_currency_unit, fx_source_url, and fx_as_of before the cross-provider USD cost gate can pass; CNY values are never copied into USD-labelled fields. A null rate cannot pass simply because a smoke run reported zero tokens in that category. The same-model provider group must also have successful official workload cells on both endpoints. The configured identifiers are deepseek-v4-flash on the official API and deepseek-ai/DeepSeek-V4-Flash on SiliconFlow. Ark's deepseek-v4-flash-260425 remains a third independent deployment in the wider provider matrix, but it is not substituted for either comparison arm.

Metric definitions

Metric Meaning How measured
Success rate Availability successful request count / total
TTFT Time to first token stream first non-empty chunk - request start
End-to-end latency complete response time request start -> final chunk
Throughput (tokens/s) generation speed output token count / (end-to-end - TTFT)
p50 / p95 / p99 latency percentiles interpolation over successful requests
std standard deviation per-provider latency dispersion
aggregate throughput / RPS batch throughput metrics aggregated output token rate and request rate

If usage.completion_tokens is unavailable, token count falls back to chunk-count approximation with a documented caveat.

Run

# From the repository root: use the shared Chapter 6 environment
uv sync --locked --python 3.12 --extra ch6

# Activate it before changing directories:
# macOS/Linux:
source .venv/bin/activate
# Windows PowerShell: .\.venv\Scripts\Activate.ps1
# Windows cmd: .venv\Scripts\activate.bat

# pip fallback when uv is not installed:
# python -m pip install -e ".[ch6]"

cd chapter6/model-benchmark

# Single-project compatibility path, still supported during migration:
# python -m pip install -r requirements.txt

cp env.example .env
# or export OPENAI_API_KEY=... MOONSHOT_API_KEY=... ARK_API_KEY=...

python demo.py

Common parameters

python demo.py --list
python demo.py --num-requests 20 --concurrency 5
python demo.py --serial
python demo.py --max-tokens 256

Specify custom endpoint/model

Use --base-url, --model, and --api-key-env to test a new provider without changing DEFAULT_PROVIDERS.

python demo.py --base-url https://api.deepseek.com --model deepseek-chat \
  --api-key-env DEEPSEEK_API_KEY --name "DeepSeek官方/deepseek-chat"

Concurrency sweep

python demo.py --model gpt-5.6-luna --concurrency-sweep 1,2,4,8,16 --num-requests 100

As concurrency increases, p95/p99/std generally get worse, and success rate may drop due to rate limits; aggregate throughput/RPS usually rises then plateaus.

Metrics and export

python demo.py --metrics ttft,throughput
python demo.py --output result.json

Offline mock validation

python demo.py --mock
python demo.py --mock --concurrency-sweep 1,2,4,8,16

This validates full aggregation logic with synthetic numbers labelled [SYNTHETIC].

Default providers

DEFAULT_PROVIDERS includes the keys that are present in environment:

  • OpenAI-compatible entries (gpt-5.6-luna)
  • Moonshot / doubao (explicit base_url + key)

OpenRouter fallback behavior: - If OPENAI_API_KEY is missing, OpenAI-style entries can still run via OpenRouter (OPENROUTER_API_KEY), with model id mapping. - For gpt-5.x, OpenRouter is preferred when OPENROUTER_API_KEY exists.

Files

File Purpose
benchmark.py core benchmark core: provider config, streaming measure, concurrency scheduling, aggregation
demo.py CLI, parameter parsing, reporting and mock mode
campaign.py full provider adapters, exact workloads, scheduler, rate ramp, cache trace, SQLite persistence
analysis.py p50/p95/p99/std, outages/MTTR, RPM/TPM, cost, and completion audit
campaign_config.json provider/model/workload/pricing inputs; no credentials
test_campaign.py deterministic native/OpenAI adapter, persistence, and analysis tests
requirements.txt dependencies
env.example env templates

Operational boundaries

  • demo.py retains low-cost defaults; it is not the full experiment.
  • The official campaign is intentionally expensive and takes at least seven days.
  • Provider pricing changes over time. Pin the rates used for a decision in campaign_config.json; missing prices remain visible as an incomplete gate.
  • TTFT depends heavily on geography/network.
  • Offline mock is for method validation only, not production decisions.

中文

多维度模型性能基准测试(实验 6-8 配套代码)

对多个 OpenAI 兼容的 LLM API 提供商做横向基准测试,一条命令跑出 TTFT / 端到端延迟 / 吞吐 / 标准差 / p50 / p95 / p99 / 成功率 的多维度对比表, 为模型选型提供实测依据。还支持并发压测(逐档加压找限流点,看指标随并发的变化) 与离线自检--mock 合成数据,无需 key/网络即可验证指标聚合)。

对应《深入理解 AI Agent》第 6 章 实验 6-8:多维度模型性能基准测试

目的

本目录现在分为两层:demo.py 保留几分钟即可运行的低成本抽样; campaign.py 则完整实现正文要求的长期实验。完整路径包含 8K/32K/128K × 512/2048、每格至少 100 次请求、一周逐小时探测、 故障分组与 MTTR、并发爬坡实测 RPM/TPM、思考长度/延迟、缓存/输入/输出 三类价格与典型多轮 Agent 成本。每次真实请求写入 SQLite,进程中断后可继续。 analysis.py 会逐项审计证据;只跑小样本或留下未填写的价格时不会误报完成。

完整实验命令

# 小规模真实 API 集成验证(不会被当成正式结果)
python campaign.py workload --smoke --requests 1 \
  --context-tokens 256 --output-tokens 128 \
  --provider 'Ark/doubao-seed-1.6' \
  --campaign-id integration-smoke --db results/integration-smoke.sqlite3

# 正式负载矩阵(默认每格 N=100)
python campaign.py workload --campaign-id release-2026-07

# 逐级并发实测 RPM / TPM 上限
python campaign.py rate-limit --campaign-id release-2026-07

# 多轮 Agent 缓存与成本轨迹
python campaign.py agent-cost --campaign-id release-2026-07

# 正式一周可用性监控
python campaign.py availability --duration-hours 168 --interval-seconds 3600 \
  --campaign-id release-2026-07

python analysis.py --campaign-id release-2026-07

正式运行前必须在 campaign_config.json 中固定本次决策采用的公开价格。 价格保留提供商发布的原始币种;非美元价格还必须固定带日期和来源的汇率,才能进入 跨提供商美元成本比较。程序不会把人民币数字直接写入美元字段。任何 input / cached input / output 单价为空,完成审计都会明确失败,而不会用猜测价格填补。

指标定义

指标 含义 怎么测的
成功率(可用性) 成功请求数 / 总请求数 单次请求任何异常(超时/限流/网络错误/空响应)都计为失败,不中断整表
TTFT 首个 token 到达延迟 流式读取,记录第一个"有内容" chunk 到达的时刻 − 请求发出时刻
端到端延迟 请求发出到响应结束的总耗时 最后一个 chunk 时刻 − 请求发出时刻
吞吐(tokens/s) 生成阶段的输出速度 输出 token 数 / (端到端 − TTFT),剥离首 token 等待,反映纯解码速度
p50 / p95 / p99 延迟的中位数 / 95 / 99 分位 对同一 (provider, model) 的多次成功请求排序后线性插值;p95、p99 高说明长尾重、体验不稳
标准差(std) 延迟的离散程度 样本标准差;书中强调"高延迟方差意味着用户体验不稳定"
聚合吞吐 / RPS 整批的总吞吐 并发压测时:全部成功请求的输出 token 总数 / 整批墙钟耗时(RPS 为成功请求数 / 墙钟);随并发上升先增后趋平,触及服务端上限即触顶

输出 token 数优先取服务端回传的精确 usage.completion_tokens; 若服务不返回 usage,则以流式 chunk 数近似计数(会略微偏高,已在代码注释标明)。

运行

# 在仓库根目录使用统一的第 6 章环境
uv sync --locked --python 3.12 --extra ch6

# 切换目录前先激活环境:
# macOS/Linux:
source .venv/bin/activate
# Windows PowerShell:.\.venv\Scripts\Activate.ps1
# Windows cmd:.venv\Scripts\activate.bat

# 未安装 uv 时可用 pip 兜底:
# python -m pip install -e ".[ch6]"

cd chapter6/model-benchmark

# 迁移期间仍支持单项目兼容路径:
# python -m pip install -r requirements.txt

# 配置 key:只需填手上有的,未设置的提供商会自动跳过
cp env.example .env        # 然后编辑 .env
# 或直接 export OPENAI_API_KEY=... MOONSHOT_API_KEY=... ARK_API_KEY=...

python demo.py             # 一条命令跑出对比表

常用参数:

python demo.py --list                          # 仅列出将测试的提供商
python demo.py --num-requests 20 --concurrency 5   # 加大样本与并发
python demo.py --serial                        # 串行发送(并发=1,看无竞争下的基线延迟)
python demo.py --max-tokens 256                # 生成更长响应,更充分地测吞吐

默认参数(N=10/家, 并发=3, max_tokens=64)单次全跑成本约几分钱。 要接近书中"每配置 ≥100 次请求"的统计口径,把 --num-requests 调到 100 即可 (注意成本与限流会同步上升)。

指定任意 OpenAI 兼容端点(不改代码测新模型/新提供商)

书中要求"对同一模型测试不同 API 提供商(如 DeepSeek 官方 vs SiliconFlow)"。 用 --base-url / --model / --api-key-env 即可直接指定单个端点,无需改 DEFAULT_PROVIDERS

python demo.py --base-url https://api.deepseek.com --model deepseek-chat \
               --api-key-env DEEPSEEK_API_KEY --name "DeepSeek官方/deepseek-chat"
# 换个 base_url、保持同一 model,即可对比"同模型不同提供商"

并发压测:逐步加压找限流点

书中实验 6-8 要求"通过逐步提升并发量来找到限流点,记录 RPM/TPM 上限"。 --concurrency-sweep 对同一模型逐档加压,产出一张随并发变化的指标表 (p50/p95/p99/std/成功率/RPS/聚合吞吐):

python demo.py --model gpt-5.6-luna --concurrency-sweep 1,2,4,8,16 --num-requests 100

随着并发上升,单请求延迟长尾(p95/p99/std)通常变差、可用性可能因限流下降, 而聚合吞吐(tokens/s)与 RPS 先升后趋平——趋平点即服务端的实际吞吐上限。

选择要显示的指标 / 导出结果

python demo.py --metrics ttft,throughput      # 主表只看 TTFT 与吞吐(成功率始终显示)
python demo.py --output result.json           # 完整结果(含 p50/p95/p99/std)写入 JSON

离线自检(--mock,无需 key/网络)

合成(synthetic)数据跑通整条指标聚合链路,便于在没有 API key 或无网络时 验证 p50/p95/p99/std/可用性/聚合吞吐的计算是否正确。输出数字全部为伪随机合成, 带 [SYNTHETIC] 标注,绝非真实基准,切勿用于选型。

python demo.py --mock                                   # 合成横向对比表
python demo.py --mock --concurrency-sweep 1,2,4,8,16     # 合成并发压测表

一次合成并发压测的输出(--mock --concurrency-sweep 1,2,4,8,16 --num-requests 100数字为合成,仅演示趋势):

并发 | 成功率         | TTFT_p50 | TTFT_p95 | 端到端p50 | 端到端p95 | 端到端p99 | 端到端std | RPS  | 聚合吞吐
-----+----------------+----------+----------+-----------+-----------+-----------+-----------+------+----------
1    | 99/100 (99%)   | 301ms    | 514ms    | 0.73s     | 1.04s     | 1.16s     | 0.13s     | 1.3  | 49.8 t/s
2    | 100/100 (100%) | 335ms    | 570ms    | 0.79s     | 1.07s     | 1.19s     | 0.15s     | 2.5  | 94.4 t/s
4    | 98/100 (98%)   | 381ms    | 617ms    | 0.83s     | 1.11s     | 1.19s     | 0.16s     | 4.7  | 180.0 t/s
8    | 92/100 (92%)   | 523ms    | 932ms    | 0.96s     | 1.53s     | 1.67s     | 0.25s     | 8.0  | 305.3 t/s
16   | 97/100 (97%)   | 878ms    | 1487ms   | 1.30s     | 1.97s     | 2.37s     | 0.35s     | 11.9 | 441.0 t/s

可见随并发上升:端到端 p95/p99 与 std 走高(长尾变差),聚合吞吐持续增长(尚未触顶)。 真实端点上这条曲线会在某个并发处趋平并伴随可用性下降——那就是限流点。

默认测试的提供商

代码里 DEFAULT_PROVIDERS 默认只跑手上有有效 key的提供商(OpenAI 一个 key 测多个模型):

展示名 模型 base_url key 环境变量
OpenAI/gpt-5.6-luna gpt-5.6-luna (官方默认,可回退 OpenRouter) OPENAI_API_KEY
Moonshot/moonshot-v1-8k moonshot-v1-8k https://api.moonshot.cn/v1 MOONSHOT_API_KEY
Doubao/doubao-1.5-pro-32k doubao-1-5-pro-32k-250115 https://ark.cn-beijing.volces.com/api/v3 ARK_API_KEY

OpenRouter 回退OpenAI/* 这几条(base_url 为空的 OpenAI 原生条目)在未设置 OPENAI_API_KEY 时会自动改走 OpenRouterOPENROUTER_API_KEY,模型名映射为 openai/*)。gpt-5.x 直连 OpenAI 需组织实名认证,因此只要设置了 OPENROUTER_API_KEY 就优先走 OpenRouter。带专属 base_url 的条目(Kimi/豆包)不参与回退。

提供商列表是可配置的:在 benchmark.pyDEFAULT_PROVIDERS 里追加 ProviderConfig(...) 即可扩展。所有提供商都走同一套 OpenAI 兼容协议, 只是 base_urlmodel 不同——这正是可以"同一模型对比不同提供商" (如书中提到的 DeepSeek 官方 vs SiliconFlow)的原因。

真实运行结果(示例)

以下是一次真实运行的输出(python demo.py --num-requests 10 --concurrency 3, 测试机在中国大陆网络环境,2026-07)。数字为真实测得,非虚构; 不同网络/时段会有波动,请以自己跑出的结果为准。

Provider/Model            | 成功率       | TTFT均值 | TTFT_p95 | 端到端均值 | 端到端p95 | 吞吐      | 输出tok
--------------------------+--------------+----------+----------+------------+-----------+-----------+--------
OpenAI/gpt-5.6-luna       | 10/10 (100%) | 1360ms   | 2334ms   | 1.73s      | 2.54s     | 174.9 t/s | 26
Moonshot/moonshot-v1-8k   | 10/10 (100%) | 530ms    | 671ms    | 0.89s      | 1.07s     | 92.1 t/s  | 32
Doubao/doubao-1.5-pro-32k | 10/10 (100%) | 1097ms   | 1409ms   | 2.32s      | 2.91s     | 36.2 t/s  | 44

结论(基于上面这次运行)

  • 可用性:本次三家全部 10/10(100%)成功。可用性差异往往要在更大样本、 更高并发或更长时间窗口下才暴露——这正是书中强调"一周每小时探测"的原因。 代码已把单点失败设计成"记为可用性下降、不中断整表",便于长时间挂机采样。
  • 首 token 延迟(TTFT):本测试机在国内网络下,Kimi 的 TTFT(~530ms)明显低于 跨境访问的 OpenAI/gpt-5.6-luna(~1.36s);豆包 TTFT(~1.1s)略低于 OpenAI 但端到端更长。 TTFT 强依赖网络位置——同一份代码在美国机房跑,OpenAI 的 TTFT 会大幅下降。
  • 吞吐:本次 gpt-5.6-luna(175 t/s)> Kimi(92 t/s)> 豆包(36 t/s)。 吞吐决定长响应的等待时间,与 TTFT 是两个独立维度。
  • 稳定性(p95):看 p95 与均值的差距。gpt-5.6-luna 跨境访问,TTFT p95(2.33s)/均值(1.36s) 拉开较大,长尾更重;Kimi 的 p95 与均值最接近,本次最稳。
  • 选型启示:不存在"全面最优"的一家——延迟、吞吐、可用性、价格是多维权衡。 面向国内用户的实时交互场景,低 TTFT 的本地化服务体验更好; 批处理/长文本生成则更看重吞吐与单价。务必在你自己的部署网络环境下实测, 不要直接照搬第三方监测平台(如 Artificial Analysis)的数字。

文件说明

文件 作用
benchmark.py 核心:提供商配置、单次流式测量、并发调度、指标聚合(含 p99/std/聚合吞吐)、并发扫描 sweep_concurrency、合成数据 synthetic_summary
demo.py 命令行入口:解析参数、跑测试(含并发压测 / --mock 离线自检)、打印对比表、导出 JSON
requirements.txt 依赖(openai SDK + 可选 python-dotenv)
env.example key 配置模板

注意事项

  • 成本控制:默认 max_tokens=64N=10,全跑成本极低。调大参数前请留意计费。
  • 限流:把并发或 N 调很大时可能触发提供商 RPM/TPM 限流,届时会以失败形式 计入可用性下降——这本身也是一种"实测限流阈值"的方式(书中实验 6-8 的一环)。
  • TTFT 与网络强相关:跨境访问的服务 TTFT 会显著偏高,结论需结合部署地点解读。
  • OpenRouter 回退:未设置 OPENAI_API_KEY 时,OpenAI/* 条目自动经 OpenRouter 路由 (需 OPENROUTER_API_KEYgpt-* 映射为 openai/*);gpt-5.x 只要有 OPENROUTER_API_KEY 即优先走 OpenRouter(直连需实名认证)。其它提供商(DEEPSEEK / SILICONFLOW 等)如需启用, 在 DEFAULT_PROVIDERS 中补充配置并设置对应环境变量即可。