v2.8.0 清空历史真正物理删除 + 请求超时可配修复慢接口被掐断

1. 清空历史=真正清空（不再只是删行留痕）:
   - 清空前先停止所有运行中测试线程，避免其继续写入孤儿数据
   - DELETE 后重置自增序列(sqlite_sequence)，新测试 ID 从 1 开始
   - WAL checkpoint(TRUNCATE) + VACUUM 物理回收文件空间，数据不可恢复
   - 单条删除删光时同样重置自增序列

2. 慢接口被掐断修复（预处理/首字耗时过长 Read timed out）:
   - 读取超时默认 300s→1800s(30分钟)，连接 60s
   - 「速度测试配置」新增 请求超时(秒) 输入框(默认1800)，测试连接+速度采样共用，按次可调(如3600)
   - 超时错误改为友好提示：等待响应超时(超过N秒未收到数据)，引导调大请求超时，不再裸抛 Read timed out
   - 测试启动日志打印当前超时配置
   - 修复案例: 18008 Qwen3.8-FP8 131072 长上下文预填充被 300s 掐断

app.py

@@ -3,6 +3,7 @@
 import io
 import json
 import subprocess
+import time
 from datetime import datetime
 
 import requests
@@ -112,13 +113,17 @@ def list_models_api():
 
 @app.route("/api/configs/test", methods=["POST"])
 def test_config():
-    cfg = _fill_defaults(request.get_json(force=True) or {})
+    body = request.get_json(force=True) or {}
+    cfg = _fill_defaults(body)
     if not cfg.get("api_key"):
         return jsonify({"ok": False, "error": "请填写 API Key"}), 400
     try:
         # 连接测试：只要流式请求成功返回（哪怕正文为空/只有思维链）都算连通
+        # 慢接口长预处理可能长时间等不到首字，这里直接用请求里带的超时（默认 30 分钟），避免误判接口不通
         m = call_stream(cfg, "你好，请简要回答：1+1=？",
-                        {"max_tokens": 32, "avoid_cache": False})
+                        {"max_tokens": 32, "avoid_cache": False,
+                         "read_timeout": body.get("read_timeout"),
+                         "connect_timeout": body.get("connect_timeout")})
         note = ""
         if not (m.get("output_tokens") or m.get("output_chars")):
             note = "（连接正常，但本次未返回正文内容，可能为推理型模型）"
@@ -188,7 +193,20 @@ def list_tests():
 
 @app.route("/api/tests", methods=["DELETE"])
 def clear_tests():
-    """一键清空全部测试历史（含采样指标与日志）"""
+    """一键清空全部测试历史（含采样指标与日志）。
+    先停止所有运行中的测试线程，避免清空后它们继续写入孤儿数据；
+    再由 db.clear_tests 做物理删除（重置自增 + VACUUM 瘦身），真正清空、不可恢复。"""
+    # 1) 停止所有运行中的测试，并等待其退出
+    alive = [r for r in RUNNERS.values() if r.is_alive()]
+    for r in alive:
+        r.request_cancel()
+    if alive:
+        deadline = time.time() + 5
+        for r in alive:
+            if r.is_alive():
+                r.join(timeout=max(0.1, deadline - time.time()))
+    RUNNERS.clear()
+    # 2) 物理清空数据库
     db.clear_tests()
     return jsonify({"ok": True})
 

config.py

@@ -10,9 +10,11 @@ DATA_DIR = os.path.join(BASE_DIR, "data")
 LOG_DIR = os.path.join(BASE_DIR, "logs")
 DB_PATH = os.path.join(DATA_DIR, "llm_speed_tester.db")
 
-# 流式请求超时：连接 60s，两次数据包间隔最长 300s（推理型/长上文模型也够用）
+# 流式请求超时：连接 60s，两次数据包间隔最长 1800s（30 分钟）
+# 注意：慢接口长上下文预处理（预填充）期间服务端可能长时间不返回任何字节，
+# 等待首字受 STREAM_READ_TIMEOUT 限制；如接口更慢可在「速度测试配置」里按次调大请求超时。
 CONNECT_TIMEOUT = 60
-STREAM_READ_TIMEOUT = 300
+STREAM_READ_TIMEOUT = 1800
 
 # data-chart-tool 图表服务地址（用它的 /api/chart 画折线图：预填充左轴虚线 / 解码右轴实线）
 CHART_API_BASE = "http://127.0.0.1:16016"

database.py

@@ -277,14 +277,21 @@ def query_tests(q="", status="", provider="", sort="id", order="desc",
 
 
 def clear_tests():
-    """一键清空全部历史（测试记录 + 采样指标 + 日志）"""
+    """一键清空全部历史（测试记录 + 采样指标 + 日志）。
+    真正清空：删除全部行 + 重置自增序列 + WAL 检查点 + VACUUM 物理回收文件空间，
+    数据不可恢复（不是只从列表里移除，文件里也不再残留）。"""
     with _lock:
         conn = _connect()
         try:
             conn.execute("DELETE FROM test_runs")
             conn.execute("DELETE FROM logs")
             conn.execute("DELETE FROM tests")
+            # 重置自增序列：清空后新测试 ID 从 1 重新开始
+            conn.execute("DELETE FROM sqlite_sequence WHERE name IN ('tests','test_runs','logs')")
             conn.commit()
+            # WAL 检查点截断 + VACUUM：把文件里已删除的数据页物理抹掉、文件瘦身
+            conn.execute("PRAGMA wal_checkpoint(TRUNCATE)")
+            conn.execute("VACUUM")
         finally:
             conn.close()
 
@@ -296,6 +303,10 @@ def delete_test(tid: int):
             conn.execute("DELETE FROM tests WHERE id=?", (tid,))
             conn.execute("DELETE FROM test_runs WHERE test_id=?", (tid,))
             conn.execute("DELETE FROM logs WHERE test_id=?", (tid,))
+            # 全部删光时顺带重置自增序列
+            n = conn.execute("SELECT COUNT(*) FROM tests").fetchone()[0]
+            if n == 0:
+                conn.execute("DELETE FROM sqlite_sequence WHERE name IN ('tests','test_runs','logs')")
             conn.commit()
         finally:
             conn.close()

llm_providers.py

@@ -116,7 +116,7 @@ def stream_openai(cfg, prompt, gen, log, should_stop=None):
             if should_stop and should_stop():
                 raise StopRequested()
             resp = requests.post(url, json=build_payload(), headers=headers, stream=True,
-                                 timeout=(config.CONNECT_TIMEOUT, config.STREAM_READ_TIMEOUT))
+                                 timeout=_timeouts(gen))
             if resp.status_code == 200:
                 break
             err = resp.text[:400]
@@ -152,6 +152,10 @@ def stream_openai(cfg, prompt, gen, log, should_stop=None):
                     cached_tokens = details.get("cached_tokens") or 0
     except StopRequested:
         raise
+    except requests.exceptions.ReadTimeout:
+        raise ProviderError(_read_timeout_err(gen))
+    except requests.exceptions.ConnectTimeout:
+        raise ProviderError(_connect_timeout_err(gen))
     except Exception as e:
         raise ProviderError("流式请求异常: %s" % e)
     finally:
@@ -171,6 +175,30 @@ def _bad_stream_options(err: str):
             or "additional properties" in err)
 
 
+def _timeouts(gen):
+    """从测试参数 gen 里取超时配置（秒），未配置则用全局默认。
+    返回 (connect, read) 元组，供 requests timeout 使用。"""
+    try:
+        connect = float(gen.get("connect_timeout") or config.CONNECT_TIMEOUT)
+    except (TypeError, ValueError):
+        connect = config.CONNECT_TIMEOUT
+    try:
+        read = float(gen.get("read_timeout") or config.STREAM_READ_TIMEOUT)
+    except (TypeError, ValueError):
+        read = config.STREAM_READ_TIMEOUT
+    return (max(connect, 5), max(read, 10))
+
+
+def _read_timeout_err(gen):
+    """读超时的友好报错：提示这是等待数据超时，可调大请求超时"""
+    return ("等待响应超时：超过 %.0f 秒未收到数据（接口预处理/首字耗时过长或网络过慢），"
+            "请在「速度测试配置」中调大“请求超时（秒）”" % _timeouts(gen)[1])
+
+
+def _connect_timeout_err(gen):
+    return "连接超时：超过 %.0f 秒未建立连接（请检查地址/网络）" % _timeouts(gen)[0]
+
+
 def _iter_json(resp):
     """解析 SSE data: 行，逐个返回 JSON 对象"""
     for raw in resp.iter_lines(decode_unicode=True):
@@ -206,7 +234,7 @@ def stream_anthropic(cfg, prompt, gen, log, should_stop=None):
         if should_stop and should_stop():
             raise StopRequested()
         resp = requests.post(url, json=payload, headers=headers, stream=True,
-                             timeout=(config.CONNECT_TIMEOUT, config.STREAM_READ_TIMEOUT))
+                             timeout=_timeouts(gen))
         if resp.status_code != 200:
             err = resp.text[:400]
             resp.close()
@@ -233,6 +261,10 @@ def stream_anthropic(cfg, prompt, gen, log, should_stop=None):
                 output_tokens = usage.get("output_tokens") or output_tokens
     except StopRequested:
         raise
+    except requests.exceptions.ReadTimeout:
+        raise ProviderError(_read_timeout_err(gen))
+    except requests.exceptions.ConnectTimeout:
+        raise ProviderError(_connect_timeout_err(gen))
     except Exception as e:
         raise ProviderError("流式请求异常: %s" % e)
     finally:
@@ -272,7 +304,7 @@ def stream_google(cfg, prompt, gen, log, should_stop=None):
         if should_stop and should_stop():
             raise StopRequested()
         resp = requests.post(url, params=params, json=payload, headers=headers, stream=True,
-                             timeout=(config.CONNECT_TIMEOUT, config.STREAM_READ_TIMEOUT))
+                             timeout=_timeouts(gen))
         if resp.status_code != 200:
             err = resp.text[:400]
             resp.close()
@@ -299,6 +331,10 @@ def stream_google(cfg, prompt, gen, log, should_stop=None):
                 cached_tokens = um.get("cachedContentTokenCount") or 0
     except StopRequested:
         raise
+    except requests.exceptions.ReadTimeout:
+        raise ProviderError(_read_timeout_err(gen))
+    except requests.exceptions.ConnectTimeout:
+        raise ProviderError(_connect_timeout_err(gen))
     except Exception as e:
         raise ProviderError("流式请求异常: %s" % e)
     finally:

static/index.html

@@ -125,6 +125,11 @@
             <span class="slider"></span>
           </label>
         </div>
+        <div class="field">
+          <label>请求超时（秒）<span class="hint-inline">等待首字/预处理上限，接口慢就调大</span></label>
+          <div class="icon-input"><span class="icon">⏱</span><input id="gen-timeout" type="number" min="30" step="30" value="1800"></div>
+          <div class="hint">连接测试与速度采样共用：超过此时长仍未收到数据才判定超时（默认 1800 秒 = 30 分钟）。测慢接口/超长上下文时建议调大，如 3600。</div>
+        </div>
         <div class="btn-group">
           <button class="btn primary block" id="btn-start">▶ 开始测试</button>
           <button class="btn danger block" id="btn-cancel" disabled>■ 停止</button>

static/js/app.js

@@ -69,6 +69,7 @@ function currentGen() {
     concurrency_levels: [...concurrencyLevelsActive].sort((a, b) => a - b),
     avoid_cache: $("#gen-avoid-cache").checked,
     warmup: $("#gen-warmup").checked,
+    read_timeout: parseInt($("#gen-timeout").value) || 1800,
   };
 }
 
@@ -201,9 +202,11 @@ function loadSelectedConfig() {
 async function testConnection() {
   const cfg = currentConfig();
   if (!cfg.api_key) { showConn(false, "请先填写 API Key"); return; }
+  // 慢接口长预处理可能等首字很久，连接测试也带上请求超时，避免误判接口不通
+  cfg.read_timeout = parseInt($("#gen-timeout").value) || 1800;
   const box = $("#conn-result");
   box.hidden = false; box.className = "conn-result";
-  box.textContent = "⏳ 正在测试连接...";
+  box.textContent = "⏳ 正在测试连接（等待首字，最长约 " + cfg.read_timeout + " 秒）...";
   try {
     const r = await api("/api/configs/test", "POST", cfg);
     if (r.ok) {

tester.py

@@ -77,6 +77,9 @@ class TestRunner(threading.Thread):
                  % (" / ".join(str(x) for x in lengths), max_tokens,
                     " / ".join(str(x) for x in concurrency_levels), n,
                     "开" if warmup else "关", "开" if avoid_cache else "关"))
+        to = self.gen.get("read_timeout") or "默认(1800)"
+        self.log("INFO", "请求超时: 连接 %s s | 等待首字/预处理 %s s（接口慢可在左侧调大）"
+                 % (self.gen.get("connect_timeout") or 60, to))
 
         ratio = self._calibrate()
         self.ratio = ratio
@@ -157,7 +160,9 @@ class TestRunner(threading.Thread):
         """
         from concurrent.futures import ThreadPoolExecutor
 
-        gen_opt = {"max_tokens": max_tokens, "avoid_cache": avoid_cache}
+        gen_opt = {"max_tokens": max_tokens, "avoid_cache": avoid_cache,
+                   "connect_timeout": self.gen.get("connect_timeout"),
+                   "read_timeout": self.gen.get("read_timeout")}
 
         def worker(idx):
             prompt = self._finalize_prompt(base_prompt)  # 每流独立随机前缀，避免共享缓存
@@ -240,7 +245,9 @@ class TestRunner(threading.Thread):
         self.log("INFO", "正在校准 token/字符 比例（发送小探测请求）...")
         try:
             m = lp.call_stream(self.cfg, probe,
-                               {"max_tokens": 8, "avoid_cache": False},
+                               {"max_tokens": 8, "avoid_cache": False,
+                                "connect_timeout": self.gen.get("connect_timeout"),
+                                "read_timeout": self.gen.get("read_timeout")},
                                log=lambda lv, msg: self.log(lv, msg),
                                should_stop=self.should_stop)
             pt = m.get("prompt_tokens") or 0