fix: compact 504 超时修复

- 降低 auto_compact 阈值: 1.2MB → 350KB (426KB 请求也会导致上游 504)
- 预截断: compact 前每条消息限制 500 字, 大幅减少 token 数
- 降低批次大小: 800KB → 350KB 每批
- 429 限流自动重试: 最多 2 次, 递增等待 (5s, 10s)
- 504 超时自动重试: 1 次, 降级为 auto_compact
- compact 内部 429/504 重试: 最多 3 次
- 全局请求日志: before_request 捕获所有请求

ai/huawei_gateway.py

@@ -45,9 +45,12 @@ TARGET_HOST = 'tokenhub.developer.huaweicloud.com'
 
 # ================= 请求体限制配置 =================
 APIG_BODY_LIMIT = 1200 * 1024    # APIG 请求体限制 ~1.2MB (实测边界1260KB)
+COMPACT_THRESHOLD = 350 * 1024   # 自动压缩阈值：超过350KB就压缩（避免上游504超时）
 GZIP_THRESHOLD = 200 * 1024      # 超过 200KB 时启用 gzip 压缩转发
 UPSTREAM_TIMEOUT_MIN = 60        # 小请求超时 60s
 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（之后再降级compact）
 
 # ================= 日志 =================
 logging.basicConfig(
@@ -299,8 +302,19 @@ def auto_compact(raw_body, real_token, orig_request):
     if not convo_msgs:
         return {"error": {"message": "无对话内容可压缩", "type": "invalid_request_error"}}, 400
 
+    # 预截断：每条消息内容限制 500 字，大幅减少 compact 请求的 token 数，避免上游 504
+    MAX_MSG_CHARS = 500
+    truncated_msgs = []
+    for m in convo_msgs:
+        content = m.get('content', '')
+        if len(content) > MAX_MSG_CHARS:
+            truncated_msgs.append({**m, 'content': content[:MAX_MSG_CHARS] + '...[截断]'})
+        else:
+            truncated_msgs.append(m)
+    convo_msgs = truncated_msgs
+
     # 按批次分割对话：每批控制在安全大小内
-    SAFE_BATCH_BYTES = 800 * 1024  # 每批 800KB
+    SAFE_BATCH_BYTES = 350 * 1024  # 每批 350KB（避免上游504超时）
     batches = []
     current_batch = []
     current_size = 0
@@ -452,6 +466,15 @@ def auto_compact(raw_body, real_token, orig_request):
         }
 
 
+# ================= 全局请求日志（捕获所有请求，包括404） =================
+@app.before_request
+def log_every_request():
+    body_preview = ""
+    if request.method in ('POST', 'PUT', 'PATCH') and request.content_length and request.content_length < 2048:
+        body_preview = request.get_data()[:200].decode('utf-8', errors='replace')
+    logger.info(f">>> {request.method} {request.full_path} | body={request.content_length or 0}bytes | from={request.remote_addr} | {body_preview}")
+
+
 @app.route('/health')
 def health():
     """健康检查端点"""
@@ -583,7 +606,7 @@ def compact_endpoint():
         return {"error": {"message": "无对话内容可压缩", "type": "invalid_request_error"}}, 400
 
     # 按批次分割对话：每批控制在安全大小内
-    SAFE_BATCH_BYTES = 800 * 1024  # 每批 800KB（留余量给 JSON 开销）
+    SAFE_BATCH_BYTES = 350 * 1024  # 每批 350KB（避免上游504超时）
     batches = []
     current_batch = []
     current_size = 0
@@ -636,6 +659,20 @@ def compact_endpoint():
                 data=batch_body, allow_redirects=False,
                 timeout=UPSTREAM_TIMEOUT_MAX, stream=False
             )
+
+            # compact 内部 429/504 重试
+            for _retry in range(3):
+                if resp.status_code not in (429, 504):
+                    break
+                retry_wait = (_retry + 1) * 5
+                logger.warning(f"compact: 第{i+1}批收到 {resp.status_code}，等待{retry_wait}s重试...")
+                time.sleep(retry_wait)
+                resp = http_session.request(
+                    method='POST', url=target_url, headers=headers,
+                    data=batch_body, allow_redirects=False,
+                    timeout=UPSTREAM_TIMEOUT_MAX, stream=False
+                )
+
             data = resp.json()
 
             if resp.status_code == 200 and 'choices' in data:
@@ -791,9 +828,12 @@ def proxy(subpath):
     else:
         upstream_timeout = UPSTREAM_TIMEOUT_MIN
 
-    # ============ 超限自动分批压缩（仅对 chat/completions） ============
+    # ============ 自动压缩（仅对 chat/completions） ============
-    if body_size > APIG_BODY_LIMIT and subpath in ('chat/completions', 'chat/completions/'):
+    # 1. 超过 APIG 限制 → 必须压缩（否则请求会被截断）
-        logger.info(f"chat/completions 请求体超限: {body_size//1024}KB > {APIG_BODY_LIMIT//1024}KB, 自动触发分批压缩")
+    # 2. 超过 COMPACT_THRESHOLD → 主动压缩（避免上游504超时）
+    if body_size > COMPACT_THRESHOLD and subpath in ('chat/completions', 'chat/completions/'):
+        reason = "超APIG限制" if body_size > APIG_BODY_LIMIT else "可能超时"
+        logger.info(f"chat/completions 请求体 {body_size//1024}KB > {COMPACT_THRESHOLD//1024}KB ({reason}), 自动触发分批压缩")
         return auto_compact(raw_body, real_token, request)
 
     # 非 chat/completions 超限，返回清晰错误
@@ -840,6 +880,55 @@ def proxy(subpath):
                     stream=True
                 )
 
+        # ============ 429 限流重试 ============
+        if resp.status_code == 429:
+            for retry_i in range(1, RETRY_ON_429 + 1):
+                resp.close()
+                wait = retry_i * 5  # 5s, 10s
+                logger.warning(f"上游 429 限流，等待{wait}s后重试({retry_i}/{RETRY_ON_429})...")
+                time.sleep(wait)
+                resp = http_session.request(
+                    method=request.method, url=target_url, headers=headers,
+                    data=raw_body, cookies=request.cookies,
+                    allow_redirects=False, timeout=upstream_timeout, stream=True
+                )
+                if resp.status_code != 429:
+                    break
+                logger.warning(f"重试仍返回 429")
+
+        # ============ 504 超时重试 + 降级compact ============
+        if resp.status_code == 504 and subpath in ('chat/completions', 'chat/completions/'):
+            for retry_i in range(1, RETRY_ON_504 + 1):
+                resp.close()
+                wait = retry_i * 3
+                logger.warning(f"上游 504 超时，等待{wait}s后重试({retry_i}/{RETRY_ON_504})...")
+                time.sleep(wait)
+                resp = http_session.request(
+                    method=request.method, url=target_url, headers=headers,
+                    data=raw_body, cookies=request.cookies,
+                    allow_redirects=False, timeout=upstream_timeout, stream=True
+                )
+                if resp.status_code != 504:
+                    break
+
+            # 重试仍504 → 降级为compact压缩后重试
+            if resp.status_code == 504:
+                resp.close()
+                logger.warning(f"504重试仍失败，降级为auto_compact压缩后重试...")
+                try:
+                    return auto_compact(raw_body, real_token, request)
+                except Exception as e:
+                    logger.error(f"降级compact也失败: {e}")
+                    # compact也失败，返回友好错误
+                    return {
+                        "error": {
+                            "message": "模型推理超时，已尝试压缩上下文但仍失败。请缩短对话后重试。",
+                            "type": "server_error",
+                            "code": "model_timeout",
+                            "param": None
+                        }
+                    }, 504
+
         # 过滤 hop-by-hop 头和压缩编码头
         skip_headers = {'transfer-encoding', 'content-encoding', 'content-length',
                         'connection', 'keep-alive', 'upgrade'}