feat: log request when e2e latency exceeds the specified value (#15759)

Co-authored-by: qiuxuan.lzw <qiuxuan.lzw@alibaba-inc.com>
Co-authored-by: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com>

python/sglang/srt/environ.py

@@ -139,6 +139,7 @@ class Envs:
     SGLANG_LOG_FORWARD_ITERS = EnvBool(False)
     SGLANG_LOG_MS = EnvBool(False)
     SGLANG_DISABLE_REQUEST_LOGGING = EnvBool(False)
+    SGLANG_LOG_REQUEST_EXCEEDED_MS = EnvInt(-1)
 
     # SGLang CI
     SGLANG_IS_IN_CI = EnvBool(False)

python/sglang/srt/utils/request_logger.py

@@ -20,6 +20,7 @@ from datetime import datetime
 from functools import lru_cache
 from typing import TYPE_CHECKING, Any, Optional, Set, Tuple, Union
 
+from sglang.srt.environ import envs
 from sglang.srt.utils.common import get_bool_env_var
 
 if TYPE_CHECKING:
@@ -47,6 +48,7 @@ class RequestLogger:
         self.metadata: Tuple[Optional[int], Optional[Set[str]], Optional[Set[str]]] = (
             self._compute_metadata()
         )
+        self.log_exceeded_ms = envs.SGLANG_LOG_REQUEST_EXCEEDED_MS.get()
 
     def configure(
         self,
@@ -100,6 +102,10 @@ class RequestLogger:
         if not self.log_requests:
             return
 
+        e2e_latency_ms = out["meta_info"]["e2e_latency"] * 1000
+        if self.log_exceeded_ms > 0 and e2e_latency_ms < self.log_exceeded_ms:
+            return
+
         max_length, skip_names, out_skip_names = self.metadata
         if self.log_requests_format == "json":
             log_data = {