diff --git a/README.md b/README.md new file mode 100644 index 0000000..ec43dc6 --- /dev/null +++ b/README.md @@ -0,0 +1,151 @@ +# ⚡ LLM 速度测试台 + +大模型推理性能基准测试工具，用于测量不同上下文长度下的 **预填充速度（prefill）**、**首字延迟（TTFT）** 与 **解码吞吐（decode）**，并以网页 + Excel 形式输出结果。 + +- **访问地址：** `http://:16097/` +- **技术栈：** Python 3 + Flask + SQLite（纯 REST，无额外依赖） +- **版本：** v2.0.0 + +--- + +## 功能特性 + +### 🔌 大模型接口配置 +- 支持 **OpenAI 兼容**（OpenAI / DeepSeek / 火山方舟 / 任意兼容网关）、**Anthropic（Claude）**、**Google Gemini** 三类提供商 +- 配置项：配置名称、Base URL（留空自动使用官方默认）、API Key（可显示/隐藏）、模型名称、温度 +- 配置可**保存/加载/删除**，方便多模型对比 + +### 🚀 速度测试配置 +- **多上下文长度测试**：默认为 `512 / 2048 / 8192 / 32768 / 131072` tokens，可通过标签点击启用/禁用，并支持手动添加任意自定义长度（≥16） +- **解码输出长度（max tokens）**：默认为 `128`，可手动自定义 +- **每个长度采样次数**：默认为 `2`，可手动自定义 +- **测试前预热（空转）**：默认开启，先发一次不计速度的空转请求，避免冷启动/首请求偏慢污染真实采样数据 +- **避免缓存**：默认开启，为每次采样追加随机前缀，测量真实预填充性能 + +### 📊 指标与结果 +- 实时指标卡：首字延迟、预填充速度、解码速度、上文/输出 tokens、总耗时 +- 实时控制台日志：校准、预热、每次采样明细全程可追溯 +- **每次完整测试**支持： + - **网页点击查看**：历史记录「查看」按钮弹出详情（整体平均 + 按上下文长度分组 + 每次采样明细 + 完整日志） + - **文件下载 Excel（xlsx）**：历史记录「Excel」按钮或详情弹窗「导出 Excel」，包含 汇总 / 采样明细 / 日志 三个 Sheet +- 测试历史留存，可随时刷新、查看、导出、删除 + +--- + +## 快速开始 + +### 环境要求 +- Python 3.10+（推荐使用 openclaw conda 环境） +- 依赖见 `requirements.txt` + +### 安装 +```bash +pip install -r requirements.txt +# 或使用 conda 环境 +/home/hz1/miniconda3/envs/openclaw/bin/pip install -r requirements.txt +``` + +### 启动 +```bash +./start.sh # 启动（默认端口 16097） +./start.sh stop # 停止 +./start.sh restart # 重启 +``` + +启动成功后访问 `http://:16097/`。 + +--- + +## 使用说明 + +1. **配置接口**：选择提供商 → 填写配置名称 / Base URL（可留空）/ API Key / 模型名称 → 点「保存」可留存，或直接点「🔍 测试连接」验证连通性 +2. **配置测试参数**：勾选要测试的上下文长度（默认 5 档），设置解码输出长度与采样次数，按需开关预热/避免缓存 +3. **开始测试**：点「▶ 开始测试」，右侧实时展示指标与日志；可随时「■ 停止」 +4. **查看与导出**：测试完成后，在「测试历史」中点「查看」看完整详情，点「Excel」或详情内「导出 Excel」下载 xlsx 报告 + +### 指标含义 +| 指标 | 含义 | +|------|------| +| 首字延迟 TTFT (ms) | 从请求发出到收到第一个 token 的时间（含预填充） | +| 预填充速度 (tok/s) | prompt tokens / 首字延迟，衡量上文处理吞吐 | +| 解码速度 (tok/s) | 输出 tokens / 解码阶段耗时，衡量逐 token 生成吞吐 | +| 上下文/输出 tokens | 实际发送的提示词 token 数与模型返回的 token 数 | + +--- + +## API + +| 方法 | 路径 | 说明 | +|------|------|------| +| GET | `/api/health` | 健康检查 | +| GET/POST | `/api/configs` | 配置列表 / 新增配置 | +| GET/DELETE | `/api/configs/` | 单个配置 / 删除 | +| POST | `/api/configs/test` | 测试连接 | +| GET/POST | `/api/tests` | 测试历史 / 启动测试 | +| GET | `/api/tests/` | 测试详情（含 runs / logs / summary） | +| GET | `/api/tests//logs?after=` | 增量日志（前端轮询用） | +| POST | `/api/tests//cancel` | 停止测试 | +| DELETE | `/api/tests/` | 删除测试 | +| GET | `/api/tests//export.xlsx` | 导出 Excel 报告 | + +### 启动测试请求示例 +```json +POST /api/tests +{ + "config": { + "provider": "openai", + "name": "DeepSeek-V4", + "base_url": "https://api.deepseek.com/v1", + "api_key": "sk-xxx", + "model": "deepseek-chat", + "temperature": 0.7 + }, + "gen": { + "context_lengths": [512, 2048, 8192, 32768, 131072], + "max_tokens": 128, + "samples": 2, + "warmup": true, + "avoid_cache": true + } +} +``` + +--- + +## 项目结构 + +``` +llm-speed-tester/ +├── app.py # Flask 主应用（路由 + Excel 导出） +├── config.py # 全局配置（端口/超时/路径） +├── database.py # SQLite 存储（配置/测试/采样/日志） +├── tester.py # 测试执行器（校准 → 预热 → 多长度采样 → 汇总） +├── llm_providers.py # 提供商适配器（OpenAI / Anthropic / Gemini） +├── requirements.txt +├── start.sh # 启动/停止脚本 +├── static/ # 前端（index.html / js/app.js / css/style.css） +├── data/ # SQLite 数据库（llm_speed_tester.db） +└── logs/ # 运行日志 + PID +``` + +### 测试执行流程 +1. **校准**：发送一个短探测请求，估算 `token/字符` 比例 +2. **逐上下文长度**：对每个选中的长度构造对应 token 量的基准提示词 +3. **预热（空转，不计速度）**：每个长度先发一次请求，规避冷启动偏差 +4. **采样**：每个长度按采样次数重复请求，记录每次指标 +5. **汇总**：按长度分组 + 整体平均，写入测试记录 + +--- + +## 常见问题 + +- **无 openpyxl**：`pip install openpyxl`（已加入 requirements.txt） +- **老版本数据库**：程序启动时自动迁移，为 `test_runs` 表补充 `context_length` 列，无需手动处理 +- **慢模型超时**：连接超时 30s、两次数据包间隔 120s，足够覆盖大多数慢模型；超长文（131072）生成慢属正常，请耐心等待 + +--- + +## Git + +- **仓库：** `hz4th_coder/llm-speed-tester` +- **版本：** v2.0.0（新增多上下文长度测试 + 预热 + Excel 导出 + 界面优化） diff --git a/app.py b/app.py index c1c1ebd..96cece5 100644 --- a/app.py +++ b/app.py @@ -1,8 +1,9 @@ # -*- coding: utf-8 -*- """LLM 速度测试台 - Flask 主应用""" +import io import json -from flask import Flask, jsonify, request, send_from_directory +from flask import Flask, jsonify, request, send_file, send_from_directory import config import database as db @@ -139,5 +140,150 @@ def del_test(tid): return jsonify({"ok": True}) +# ───────────────────────── Excel 导出 ───────────────────────── + +@app.route("/api/tests//export.xlsx") +def export_xlsx(tid): + t = db.get_test(tid) + if not t: + return jsonify({"ok": False, "error": "测试不存在"}), 404 + t["runs"] = db.get_runs(tid) + t["logs"] = db.get_logs(tid) + try: + data = _build_xlsx(t) + except Exception as e: + return jsonify({"ok": False, "error": "导出失败: %s" % e}), 500 + return send_file(data, as_attachment=True, + download_name="llm_speed_test_%d.xlsx" % tid, + mimetype="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet") + + +def _build_xlsx(t): + from openpyxl import Workbook + from openpyxl.styles import Alignment, Font, PatternFill + + s = t.get("summary") or {} + g = t.get("gen") or {} + cfg = t.get("config") or {} + by_length = s.get("by_length") or {} + runs = t.get("runs") or [] + logs = t.get("logs") or [] + + wb = Workbook() + head_fill = PatternFill("solid", fgColor="2A3550") + head_font = Font(color="FFFFFF", bold=True) + title_font = Font(bold=True, size=12) + + def style_header(ws, row, ncol): + for c in range(1, ncol + 1): + cell = ws.cell(row=row, column=c) + cell.fill = head_fill + cell.font = head_font + cell.alignment = Alignment(horizontal="center", vertical="center") + + # ── Sheet1 汇总 ── + ws = wb.active + ws.title = "汇总" + ws.append(["LLM 速度测试报告"]) + ws.cell(1, 1).font = Font(bold=True, size=14) + info = [ + ["测试编号", "#%d" % t["id"]], + ["创建时间", t.get("created_at", "")], + ["状态", t.get("status", "")], + ["提供商", t.get("provider", "")], + ["模型", t.get("model", "")], + ["Base URL", cfg.get("base_url") or "（默认）"], + ["上下文长度列表", " / ".join(str(x) for x in (g.get("context_lengths") or []))], + ["生成长度(max tokens)", g.get("max_tokens", 128)], + ["每个长度采样次数", g.get("samples", 2)], + ["预热(空转)", "开" if g.get("warmup", True) else "关"], + ["避免缓存", "开" if g.get("avoid_cache") else "关"], + ["采样(成功/总数)", "%s / %s" % (s.get("samples_ok"), s.get("samples_total"))], + ["校准 字符/token", s.get("calibration_chars_per_token") or "—"], + ["错误信息", t.get("error") or ""], + ] + for row in info: + ws.append(row) + ws.cell(14, 1).font = title_font + r0 = len(info) + 2 + overall = [ + ["平均首字延迟(ms)", s.get("avg_ttft_ms")], + ["最佳首字延迟(ms)", s.get("best_ttft_ms")], + ["平均预填充速度(tok/s)", s.get("avg_prefill_speed")], + ["平均解码速度(tok/s)", s.get("avg_decode_speed")], + ["平均提示词(tok)", s.get("avg_prompt_tokens")], + ["平均输出(tok)", s.get("avg_output_tokens")], + ["平均总耗时(ms)", s.get("avg_total_ms")], + ] + ws.cell(r0, 1, "整体平均指标").font = title_font + for i, row in enumerate(overall, start=r0 + 1): + ws.append([]) + for j, v in enumerate(row, start=1): + ws.cell(row=i, column=j, value=v) + + # 按上下文长度分组 + r1 = r0 + len(overall) + 2 + ws.cell(r1, 1, "按上下文长度分组").font = title_font + cols = ["上下文长度(tok)", "采样(成功/总数)", "首字ms", "预填充tok/s", "解码tok/s", "提示词tok", "输出tok", "总耗时ms"] + ws.append([]) + for j, c in enumerate(cols, start=1): + ws.cell(row=r1 + 1, column=j, value=c) + style_header(ws, r1 + 1, len(cols)) + if by_length: + rr = r1 + 2 + for L in sorted(int(k) for k in by_length): + bl = by_length[str(L)] if str(L) in by_length else by_length[L] + ws.cell(row=rr, column=1, value=L) + ws.cell(row=rr, column=2, value="%s / %s" % (bl.get("samples_ok"), bl.get("samples_total"))) + ws.cell(row=rr, column=3, value=bl.get("avg_ttft_ms")) + ws.cell(row=rr, column=4, value=bl.get("avg_prefill_speed")) + ws.cell(row=rr, column=5, value=bl.get("avg_decode_speed")) + ws.cell(row=rr, column=6, value=bl.get("avg_prompt_tokens")) + ws.cell(row=rr, column=7, value=bl.get("avg_output_tokens")) + ws.cell(row=rr, column=8, value=bl.get("avg_total_ms")) + rr += 1 + else: + ws.cell(row=r1 + 2, column=1, value="（无成功采样数据）") + for col, w in zip("ABCDEFGH", [22, 20, 12, 14, 14, 12, 12, 14]): + ws.column_dimensions[col].width = w + + # ── Sheet2 采样明细 ── + ws2 = wb.create_sheet("采样明细") + h2 = ["序号", "上下文长度tok", "提示词tok", "缓存tok", "首字ms", "预填充tok/s", + "输出tok", "解码tok/s", "总耗时ms", "备注"] + ws2.append(h2) + style_header(ws2, 1, len(h2)) + for i, r in enumerate(runs, start=1): + m = r.get("metrics") or {} + ws2.append([ + i, + r.get("context_length") or m.get("context_length") or "", + m.get("prompt_tokens") or "", + m.get("cached_tokens") if m.get("cached_tokens") else "", + m.get("ttft_ms"), + m.get("prefill_speed"), + m.get("output_tokens"), + m.get("decode_speed"), + m.get("total_ms"), + r.get("error") or "OK", + ]) + for col, w in zip("ABCDEFGHIJ", [8, 14, 12, 10, 12, 14, 12, 14, 12, 30]): + ws2.column_dimensions[col].width = w + + # ── Sheet3 日志 ── + ws3 = wb.create_sheet("日志") + ws3.append(["相对时间(s)", "级别", "内容"]) + style_header(ws3, 1, 3) + for l in logs: + ws3.append([l.get("rel", 0), l.get("level", ""), l.get("msg", "")]) + for col, w in zip("ABC", [14, 10, 90]): + ws3.column_dimensions[col].width = w + + bio = io.BytesIO() + wb.save(bio) + bio.seek(0) + return bio + + if __name__ == "__main__": app.run(host=config.HOST, port=config.PORT, threaded=True, debug=False) diff --git a/database.py b/database.py index 966c87f..d3d663f 100644 --- a/database.py +++ b/database.py @@ -39,6 +39,7 @@ CREATE TABLE IF NOT EXISTS test_runs( id INTEGER PRIMARY KEY AUTOINCREMENT, test_id INTEGER NOT NULL, run_index INTEGER DEFAULT 0, + context_length INTEGER DEFAULT 0, metrics_json TEXT DEFAULT '{}', error TEXT DEFAULT '' ); @@ -65,11 +66,20 @@ def _connect(): return conn +def _migrate(conn): + """老库升级：为 test_runs 补 context_length 列""" + cur = conn.execute("PRAGMA table_info(test_runs)") + cols = [r[1] for r in cur.fetchall()] + if "context_length" not in cols: + conn.execute("ALTER TABLE test_runs ADD COLUMN context_length INTEGER DEFAULT 0") + + def init_db(): with _lock: conn = _connect() try: conn.executescript(SCHEMA) + _migrate(conn) conn.commit() finally: conn.close() @@ -202,13 +212,13 @@ def delete_test(tid: int): # ───────────────────────── 采样指标 ───────────────────────── -def add_run(tid: int, run_index: int, metrics: dict, error: str = ""): +def add_run(tid: int, run_index: int, metrics: dict, error: str = "", context_length: int = 0): with _lock: conn = _connect() try: conn.execute( - "INSERT INTO test_runs(test_id,run_index,metrics_json,error) VALUES(?,?,?,?)", - (tid, run_index, json.dumps(metrics, ensure_ascii=False), error)) + "INSERT INTO test_runs(test_id,run_index,context_length,metrics_json,error) VALUES(?,?,?,?,?)", + (tid, run_index, context_length, json.dumps(metrics, ensure_ascii=False), error)) conn.commit() finally: conn.close() @@ -219,7 +229,7 @@ def get_runs(tid: int): conn = _connect() try: rows = conn.execute( - "SELECT run_index,metrics_json,error FROM test_runs " + "SELECT run_index,context_length,metrics_json,error FROM test_runs " "WHERE test_id=? ORDER BY run_index", (tid,)).fetchall() out = [] for r in rows: diff --git a/requirements.txt b/requirements.txt index a0d407c..c0c822d 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,2 +1,3 @@ flask>=3.0 requests>=2.31 +openpyxl>=3.1 diff --git a/static/css/style.css b/static/css/style.css index 0a60b52..8caef87 100644 --- a/static/css/style.css +++ b/static/css/style.css @@ -71,6 +71,26 @@ body { .row > select { flex: 1; } .row > input { flex: 1; } +/* 带图标前缀的输入框（配置名称 / Base URL / 模型名称） */ +.icon-input { display: flex; align-items: center; gap: 8px; background: var(--panel2); border: 1px solid var(--border); border-radius: 8px; padding: 0 10px; transition: border-color .15s; } +.icon-input:focus-within { border-color: var(--accent); } +.icon-input .icon { font-size: 14px; flex-shrink: 0; opacity: .9; } +.icon-input input { flex: 1; background: transparent; border: none; outline: none; color: var(--text); font-size: 13px; padding: 8px 0; } + +/* 上下文长度 chips */ +.chips { display: flex; flex-wrap: wrap; gap: 6px; min-height: 30px; } +.chips-empty { color: var(--muted); font-size: 12px; align-self: center; } +.chip { + display: inline-flex; align-items: center; gap: 6px; + padding: 4px 8px 4px 10px; border-radius: 16px; cursor: pointer; user-select: none; + background: rgba(79,140,255,.16); border: 1px solid var(--accent); color: var(--accent); + font-family: var(--mono); font-size: 12px; transition: .15s; +} +.chip:hover { filter: brightness(1.15); } +.chip.off { background: var(--panel2); border-color: var(--border); color: var(--muted); text-decoration: line-through; } +.chip .chip-x { font-size: 11px; line-height: 1; border-radius: 50%; padding: 1px 3px; } +.chip .chip-x:hover { background: rgba(255,92,108,.25); color: var(--danger); } + /* 开关 */ .switch-field { display: flex; align-items: center; justify-content: space-between; } .switch { position: relative; display: inline-block; width: 46px; height: 26px; flex-shrink: 0; } diff --git a/static/index.html b/static/index.html index 8b8967d..6de9541 100644 --- a/static/index.html +++ b/static/index.html @@ -40,11 +40,11 @@
- +
📛
- +
🌐
@@ -56,7 +56,7 @@
- +
🤖
@@ -69,16 +69,27 @@

🚀 速度测试配置

- - + +
+
+ + +
- - + +
- - + + +
+
+ +
@@ -141,6 +152,7 @@

测试详情 #

+
diff --git a/static/js/app.js b/static/js/app.js index bd390c8..99bf13d 100644 --- a/static/js/app.js +++ b/static/js/app.js @@ -21,6 +21,11 @@ let pollTimer = null; let lastLogId = 0; let consoleLogs = []; // 当前测试已加载日志 [{id,level,msg,rel}] +// 上下文长度：chips 列表 + 启用集合（默认 512/2048/8192/32768/131072） +const DEFAULT_CONTEXT_LENGTHS = [512, 2048, 8192, 32768, 131072]; +let contextLengths = [...DEFAULT_CONTEXT_LENGTHS]; +let contextLengthsActive = new Set(contextLengths); + const esc = (s) => String(s ?? "").replace(/[&<>"']/g, (c) => ({ "&": "&", "<": "<", ">": ">", '"': """, "'": "'" }[c])); const fmt = (v, d = "—") => (v === null || v === undefined || isNaN(v) ? d : v); @@ -50,14 +55,58 @@ function currentConfig() { } function currentGen() { + const lens = contextLengths.filter((l) => contextLengthsActive.has(l)); return { - prompt_tokens: parseInt($("#gen-prompt-tokens").value) || 2048, - max_tokens: parseInt($("#gen-max-tokens").value) || 256, - samples: parseInt($("#gen-samples").value) || 3, + context_lengths: lens.length ? lens : [2048], + max_tokens: parseInt($("#gen-max-tokens").value) || 128, + samples: parseInt($("#gen-samples").value) || 2, avoid_cache: $("#gen-avoid-cache").checked, + warmup: $("#gen-warmup").checked, }; } +/* ───────────────────────── 上下文长度 chips ───────────────────────── */ + +function renderChips() { + const box = $("#gen-contexts"); + box.innerHTML = ""; + if (!contextLengths.length) { + box.innerHTML = '暂无长度，点击下方“添加”自定义'; + return; + } + for (const len of contextLengths) { + const active = contextLengthsActive.has(len); + const c = document.createElement("span"); + c.className = "chip" + (active ? "" : " off"); + c.title = "点击启用/禁用"; + c.innerHTML = `${len}✕`; + c.addEventListener("click", (e) => { + if (e.target.closest(".chip-x")) { + contextLengths = contextLengths.filter((x) => x !== len); + contextLengthsActive.delete(len); + renderChips(); + } else { + if (contextLengthsActive.has(len)) contextLengthsActive.delete(len); + else contextLengthsActive.add(len); + renderChips(); + } + }); + box.appendChild(c); + } +} + +function addContextLength() { + const v = parseInt($("#gen-context-add").value); + if (!v || v < 16) { toast("请输入有效长度（≥16）"); return; } + if (!contextLengths.includes(v)) { + contextLengths.push(v); + contextLengthsActive.add(v); + contextLengths.sort((a, b) => a - b); + renderChips(); + } + $("#gen-context-add").value = ""; +} + function updateDefaultUrlHint() { const p = $("#cfg-provider").value; $("#cfg-default-url").textContent = "默认地址：" + PROVIDER_DEFAULT_URL[p]; @@ -269,6 +318,7 @@ async function loadHistory() { ${STATUS_LABEL[t.status] || t.status} + `; tb.appendChild(tr); @@ -284,17 +334,43 @@ async function viewDetail(id) { const cfg = t.config || {}; const runs = t.runs || []; const logs = t.logs || []; + const byLength = s.by_length || {}; + + let byLengthHtml; + const lens = Object.keys(byLength).sort((a, b) => a - b); + if (lens.length) { + byLengthHtml = ` + + + ` + + lens.map((L) => { + const bl = byLength[L] || {}; + return ` + + + + + + + + + `; + }).join("") + `
上下文长度tok采样(成功/总数)首字ms预填充tok/s解码tok/s提示词tok输出tok总耗时ms
${L}${fmt(bl.samples_ok)}/${fmt(bl.samples_total)}${fmt(bl.avg_ttft_ms)}${fmt(bl.avg_prefill_speed)}${fmt(bl.avg_decode_speed)}${fmt(bl.avg_prompt_tokens)}${fmt(bl.avg_output_tokens)}${fmt(bl.avg_total_ms)}
`; + } else { + byLengthHtml = '
无成功采样数据
'; + } let runsHtml; if (runs.length) { runsHtml = ` - + ` + runs.map((r, i) => { const m = r.metrics || {}; return ` + @@ -331,11 +407,15 @@ async function viewDetail(id) {
平均总耗时
${fmt(s.avg_total_ms)} ms
+

📏 按上下文长度汇总

+ ${byLengthHtml} +

⚙️ 测试参数

-
上文长度
${g.prompt_tokens ?? "—"} tok
-
生成长度
${g.max_tokens ?? "—"} tok
-
采样次数
${g.samples ?? "—"}
+
上下文长度
${(g.context_lengths || []).join(" / ") || "—"} tok
+
解码输出长度
${g.max_tokens ?? "—"} tok
+
每个长度采样
${g.samples ?? "—"}
+
预热(空转)
${g.warmup === false ? "关" : "开"}
避免缓存
${g.avoid_cache ? "开" : "关"}
温度
${fmt(cfg.temperature)}
Base URL
${esc(cfg.base_url || "（默认）")}
@@ -362,6 +442,25 @@ function exportDetail() { download(JSON.stringify(window.__detail, null, 2), `test_${window.__detail.id}.json`, "application/json"); } +async function exportXlsx(id) { + try { + const resp = await fetch(`/api/tests/${id}/export.xlsx`); + if (!resp.ok) { + const j = await resp.json().catch(() => ({})); + throw new Error(j.error || "导出失败"); + } + const blob = await resp.blob(); + const a = document.createElement("a"); + a.href = URL.createObjectURL(blob); + a.download = `llm_speed_test_${id}.xlsx`; + document.body.appendChild(a); + a.click(); + setTimeout(() => { URL.revokeObjectURL(a.href); a.remove(); }, 100); + } catch (e) { + toast("导出失败：" + e.message); + } +} + /* ───────────────────────── 导出 / 下载 ───────────────────────── */ function download(text, filename, type = "text/plain") { @@ -424,6 +523,8 @@ function bind() { $("#btn-test-conn").addEventListener("click", testConnection); $("#btn-start").addEventListener("click", startTest); $("#btn-cancel").addEventListener("click", stopTest); + $("#btn-context-add").addEventListener("click", addContextLength); + $("#gen-context-add").addEventListener("keydown", (e) => { if (e.key === "Enter") addContextLength(); }); $("#btn-clear-console").addEventListener("click", clearConsole); $("#btn-export-log").addEventListener("click", exportCurrentLog); @@ -431,8 +532,10 @@ function bind() { $("#history tbody").addEventListener("click", (e) => { const v = e.target.closest("[data-view]"); + const x = e.target.closest("[data-xlsx]"); const d = e.target.closest("[data-del]"); if (v) viewDetail(Number(v.dataset.view)); + if (x) exportXlsx(Number(x.dataset.xlsx)); if (d) { const id = Number(d.dataset.del); if (confirm(`确定删除测试 #${id} 及其全部日志？`)) { @@ -444,6 +547,7 @@ function bind() { $("#dt-close").addEventListener("click", closeDetail); $("#detail-mask").addEventListener("click", (e) => { if (e.target === $("#detail-mask")) closeDetail(); }); $("#dt-export").addEventListener("click", exportDetail); + $("#dt-export-xlsx").addEventListener("click", () => { if (window.__detail) exportXlsx(window.__detail.id); }); document.addEventListener("keydown", (e) => { if (e.key === "Escape") closeDetail(); }); } @@ -452,6 +556,7 @@ function bind() { (async function init() { bind(); + renderChips(); updateDefaultUrlHint(); loadConfigs(); loadHistory(); diff --git a/tester.py b/tester.py index ba5247a..9f564d2 100644 --- a/tester.py +++ b/tester.py @@ -48,43 +48,59 @@ class TestRunner(threading.Thread): def _run(self): provider = self.cfg.get("provider", "openai") model = self.cfg.get("model", "") - n = max(1, int(self.gen.get("samples", 3))) - target_tokens = max(16, int(self.gen.get("prompt_tokens", 2048))) - max_tokens = max(1, int(self.gen.get("max_tokens", 256))) - avoid_cache = bool(self.gen.get("avoid_cache")) + gen = self.gen + + # 上下文长度列表（支持手动自定义，默认 512/2048/8192/32768/131072） + raw_lengths = gen.get("context_lengths") or [] + if not raw_lengths: + # 兼容旧版单值配置 + raw_lengths = [int(gen.get("prompt_tokens", 2048))] + lengths = sorted(set(int(x) for x in raw_lengths if int(x) >= 16)) or [2048] + n = max(1, int(gen.get("samples", 2))) # 每个长度采样次数 + max_tokens = max(1, int(gen.get("max_tokens", 128))) # 解码输出长度 + avoid_cache = bool(gen.get("avoid_cache")) + warmup = bool(gen.get("warmup", True)) # 测试前空转预热 self.log("INFO", "═══ 开始速度测试 ═══") self.log("INFO", "提供商: %s | 模型: %s" % (lp.PROVIDER_LABELS.get(provider, provider), model)) - self.log("INFO", "目标上文: %d tokens | 生成长度: %d tokens | 采样: %d 次 | 避免缓存: %s" - % (target_tokens, max_tokens, n, "开" if avoid_cache else "关")) + self.log("INFO", "上下文长度: %s tokens | 生成长度: %d tokens | 每个长度采样: %d 次 | 预热: %s | 避免缓存: %s" + % (" / ".join(str(x) for x in lengths), max_tokens, n, + "开" if warmup else "关", "开" if avoid_cache else "关")) ratio = self._calibrate() self.ratio = ratio - base_prompt = self._build_prompt(target_tokens, ratio) - self.log("INFO", "构造基准提示词完成，目标约 %d tokens" % target_tokens) + self.log("INFO", "校准完成: %.3f tok/字符（%.2f 字符/token）" % (ratio, 1.0 / ratio)) - for i in range(1, n + 1): + for L in lengths: if self.should_stop(): raise StopRequested() - prompt = self._finalize_prompt(base_prompt) - self.log("INFO", "── 采样 %d/%d 开始 ──" % (i, n)) - try: - m = lp.call_stream( - self.cfg, prompt, - {"max_tokens": max_tokens, "avoid_cache": avoid_cache}, - log=lambda lv, msg: self.log(lv, msg), - should_stop=self.should_stop) - m["run_index"] = i - self.samples.append({"run_index": i, "ok": True, "metrics": m}) - db.add_run(self.test_id, i, m) - self.log("METRIC", self._fmt_metric(i, n, m)) - except StopRequested: - raise - except ProviderError as e: - self.log("ERROR", "采样 %d/%d 失败: %s" % (i, n, e)) - self.samples.append({"run_index": i, "ok": False, "error": str(e)}) - db.add_run(self.test_id, i, {}, str(e)) - raise e + base_prompt = self._build_prompt(L, ratio) + self.log("INFO", "▸▸ 上下文长度 %d tokens（基准提示词构造完成）" % L) + if warmup: + self._warmup(base_prompt) + for i in range(1, n + 1): + if self.should_stop(): + raise StopRequested() + prompt = self._finalize_prompt(base_prompt) + self.log("INFO", "── [%d tok] 采样 %d/%d 开始 ──" % (L, i, n)) + try: + m = lp.call_stream( + self.cfg, prompt, + {"max_tokens": max_tokens, "avoid_cache": avoid_cache}, + log=lambda lv, msg: self.log(lv, msg), + should_stop=self.should_stop) + m["run_index"] = i + m["context_length"] = L + self.samples.append({"run_index": i, "context_length": L, "ok": True, "metrics": m}) + db.add_run(self.test_id, i, m, context_length=L) + self.log("METRIC", self._fmt_metric(L, i, n, m)) + except StopRequested: + raise + except ProviderError as e: + self.log("ERROR", "[%d tok] 采样 %d/%d 失败: %s" % (L, i, n, e)) + self.samples.append({"run_index": i, "context_length": L, "ok": False, "error": str(e)}) + db.add_run(self.test_id, i, {}, str(e), context_length=L) + raise e summary = self._make_summary() db.update_status(self.test_id, "done", summary=summary) @@ -94,6 +110,20 @@ class TestRunner(threading.Thread): summary.get("avg_prefill_speed") or 0, summary.get("avg_decode_speed") or 0)) + def _warmup(self, base_prompt): + """空转预热：不计入任何速度统计，用于避免冷启动/首次请求偏慢影响采样""" + self.log("INFO", "预热（空转，不计速度）...") + try: + lp.call_stream(self.cfg, base_prompt, + {"max_tokens": 8, "avoid_cache": False}, + log=lambda lv, msg: self.log(lv, msg), + should_stop=self.should_stop) + self.log("INFO", "预热完成（不纳入统计）") + except StopRequested: + raise + except Exception as e: + self.log("WARN", "预热失败（继续测试）: %s" % e) + # ───────────────────────── 工具方法 ───────────────────────── def _calibrate(self): @@ -129,15 +159,15 @@ class TestRunner(threading.Thread): return "[cache-bust %s]\n%s" % (uuid.uuid4().hex, base) return base - def _fmt_metric(self, i, n, m): - return ("采样 %d/%d 完成 | 提示词 %d tok | 缓存 %d tok | 首字 %s ms | 预填充 %s tok/s" + def _fmt_metric(self, L, i, n, m): + return ("[%d tok] 采样 %d/%d 完成 | 提示词 %d tok | 缓存 %d tok | 首字 %s ms | 预填充 %s tok/s" " | 输出 %d tok | 解码 %s tok/s | 总耗时 %s ms" - % (i, n, m.get("prompt_tokens") or 0, m.get("cached_tokens") or 0, + % (L, i, n, m.get("prompt_tokens") or 0, m.get("cached_tokens") or 0, m.get("ttft_ms"), m.get("prefill_speed"), m.get("output_tokens") or 0, m.get("decode_speed"), m.get("total_ms"))) def _make_summary(self): - ok = [s["metrics"] for s in self.samples if s.get("ok")] + ok = [s for s in self.samples if s.get("ok")] base = { "provider": self.cfg.get("provider"), "model": self.cfg.get("model"), @@ -149,19 +179,36 @@ class TestRunner(threading.Thread): if not ok: return base - def avg(k): - vals = [m[k] for m in ok if m.get(k) is not None] + def avg(ms, k): + vals = [m[k] for m in ms if m.get(k) is not None] return round(statistics.mean(vals), 1) if vals else None + # 按上下文长度分组汇总 + by_length = {} + for L in sorted(set(s["context_length"] for s in ok)): + group = [s["metrics"] for s in ok if s["context_length"] == L] + by_length[L] = { + "samples_total": sum(1 for s in self.samples if s["context_length"] == L), + "samples_ok": len(group), + "avg_ttft_ms": avg(group, "ttft_ms"), + "avg_prefill_speed": avg(group, "prefill_speed"), + "avg_decode_speed": avg(group, "decode_speed"), + "avg_prompt_tokens": avg(group, "prompt_tokens"), + "avg_output_tokens": avg(group, "output_tokens"), + "avg_total_ms": avg(group, "total_ms"), + } + + okm = [s["metrics"] for s in ok] summary = dict(base) summary.update({ - "avg_ttft_ms": avg("ttft_ms"), - "avg_prefill_speed": avg("prefill_speed"), - "avg_decode_speed": avg("decode_speed"), - "avg_prompt_tokens": avg("prompt_tokens"), - "avg_output_tokens": avg("output_tokens"), - "avg_cached_tokens": avg("cached_tokens"), - "avg_total_ms": avg("total_ms"), - "best_ttft_ms": min([m["ttft_ms"] for m in ok if m.get("ttft_ms") is not None], default=None), + "by_length": by_length, + "avg_ttft_ms": avg(okm, "ttft_ms"), + "avg_prefill_speed": avg(okm, "prefill_speed"), + "avg_decode_speed": avg(okm, "decode_speed"), + "avg_prompt_tokens": avg(okm, "prompt_tokens"), + "avg_output_tokens": avg(okm, "output_tokens"), + "avg_cached_tokens": avg(okm, "cached_tokens"), + "avg_total_ms": avg(okm, "total_ms"), + "best_ttft_ms": min([m["ttft_ms"] for m in okm if m.get("ttft_ms") is not None], default=None), }) return summary
采样提示词tok缓存tok首字ms预填充tok/s采样上下文tok提示词tok缓存tok首字ms预填充tok/s 输出tok解码tok/s总耗时ms备注
${i + 1}${r.context_length || m.context_length || "—"} ${fmt(m.prompt_tokens)} ${fmt(m.cached_tokens, 0)} ${fmt(m.ttft_ms)}