refactor(ai): 优化华为网关连接池配置与 Token 缓存查找逻辑

ai/huawei_gateway.py

@@ -43,6 +43,12 @@ UPSTREAM_TIMEOUT_MAX = 300       # 大请求超时 300s
 RETRY_ON_429 = 2                 # 429限流重试次数
 RETRY_ON_504 = 1                 # 504超时重试次数
 
+# ================= 并发配置 =================
+POOL_CONNECTIONS = 64            # 连接池大小（须 ≥ Waitress 线程数）
+POOL_MAXSIZE = 64               # 单主机最大连接数
+WAITRESS_THREADS = 64           # Waitress 工作线程数
+SSE_CHUNK_SIZE = 4096           # SSE 流式转发块大小（越小首 token 延迟越低）
+
 # ================= 日志 =================
 logging.basicConfig(
     level=logging.INFO,
@@ -184,11 +190,14 @@ def scan_pid_mem(pid):
 
 def find_token_in_memory():
     """在所有进程中扫描 Token"""
+    # 快速路径：缓存有效直接返回（避免每次请求都查环境变量/文件）
+    cached = cache.get()
+    if cached:
+        return cached
+
     # HUAWEI_TOKEN 环境变量优先级最高
     env_token = os.environ.get('HUAWEI_TOKEN', '').strip()
     if env_token and len(env_token) > 200 and not cache.is_blacklisted(env_token):
-        cached = cache.get()
-        if cached != env_token:
         cache.set(env_token)
         logger.info("Token 从 HUAWEI_TOKEN 环境变量加载")
         return env_token
@@ -203,8 +212,6 @@ def find_token_in_memory():
                     if line.startswith('HUAWEI_TOKEN='):
                         file_token = line.split('=', 1)[1].strip().strip('"').strip("'")
                         if file_token and len(file_token) > 200 and not cache.is_blacklisted(file_token):
-                            cached = cache.get()
-                            if cached != file_token:
                             cache.set(file_token)
                             logger.info("Token 从持久化文件加载")
                             return file_token
@@ -212,7 +219,7 @@ def find_token_in_memory():
         except (OSError, IOError):
             pass
 
-    if cache.is_scan_cooldown() and cache.get():
+    if cache.is_scan_cooldown():
         return cache.get()
 
     try:
@@ -257,11 +264,12 @@ def find_token_in_memory():
 
 
 # ================= HTTP 会话池 =================
+# max_retries=0：禁用 urllib3 自动重试，由 proxy 手动控制重试逻辑（避免双重重试）
 http_session = requests.Session()
 adapter = requests.adapters.HTTPAdapter(
-    pool_connections=20,
-    pool_maxsize=20,
-    max_retries=3
+    pool_connections=POOL_CONNECTIONS,
+    pool_maxsize=POOL_MAXSIZE,
+    max_retries=0
 )
 http_session.mount('https://', adapter)
 http_session.mount('http://', adapter)
@@ -338,7 +346,7 @@ def _build_response(resp):
     if 'text/event-stream' in content_type or resp.headers.get('Transfer-Encoding', '') == 'chunked':
         def sse_stream():
             try:
-                for chunk in resp.iter_content(chunk_size=16384):
+                for chunk in resp.iter_content(chunk_size=SSE_CHUNK_SIZE):
                     if chunk:
                         yield chunk
             finally:
@@ -355,6 +363,13 @@ def _build_response(resp):
         content = resp.content
         resp.close()
 
+        # 记录上游非 200 响应（在读取 content 之后，避免提前消费流）
+        if resp.status_code != 200:
+            try:
+                logger.warning(f"上游返回 {resp.status_code}: {content[:500].decode('utf-8', errors='replace')}")
+            except Exception:
+                logger.warning(f"上游返回 {resp.status_code}")
+
         # 华为云 ModelArts 错误 → OpenAI 标准格式
         if resp.status_code >= 400:
             try:
@@ -404,6 +419,7 @@ def proxy(subpath):
 
     headers['Authorization'] = f'Bearer {real_token}'
     headers['Host'] = TARGET_HOST
+    headers['Connection'] = 'keep-alive'
 
     target_url = f'https://{TARGET_HOST}/v2/{subpath}'
 
@@ -485,14 +501,6 @@ def proxy(subpath):
                 if resp.status_code != 504:
                     break
 
-        # 记录上游非200响应
-        if resp.status_code != 200:
-            try:
-                _err_peek = resp.content[:500]
-                logger.warning(f"上游返回 {resp.status_code}: {_err_peek.decode('utf-8', errors='replace')}")
-            except Exception:
-                logger.warning(f"上游返回 {resp.status_code}")
-
         return _build_response(resp)
 
     except requests.exceptions.Timeout:
@@ -530,7 +538,7 @@ def main():
     try:
         import waitress
         logger.info(f"使用 Waitress 启动网关 ({host}:{port})")
-        waitress.serve(app, host=host, port=port, threads=32)
+        waitress.serve(app, host=host, port=port, threads=WAITRESS_THREADS)
     except ImportError:
         try:
             import gunicorn.app.wsgiapp