Fix the request loggings to make it fully able to be easily replayed (#2973)

python/sglang/srt/managers/configure_logging.py

@@ -27,6 +27,7 @@ import requests
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("--url", type=str, default="http://localhost:30000")
+    parser.add_argument("--log-requests", action="store_true")
     parser.add_argument(
         "--dump-requests-folder", type=str, default="/tmp/sglang_request_dump"
     )
@@ -36,6 +37,8 @@ if __name__ == "__main__":
     response = requests.post(
         args.url + "/configure_logging",
         json={
+            "log_requests": args.log_requests,
+            "log_requests_level": 1,  # Log full requests
             "dump_requests_folder": args.dump_requests_folder,
             "dump_requests_threshold": args.dump_requests_threshold,
         },

python/sglang/srt/managers/io_struct.py

@@ -495,6 +495,7 @@ class ProfileReq(Enum):
 @dataclass
 class ConfigureLoggingReq:
     log_requests: Optional[bool] = None
+    log_requests_level: Optional[int] = None
     dump_requests_folder: Optional[str] = None
     dump_requests_threshold: Optional[int] = None
 

python/sglang/srt/managers/tokenizer_manager.py

@@ -117,6 +117,7 @@ class TokenizerManager:
         self.server_args = server_args
         self.enable_metrics = server_args.enable_metrics
         self.log_requests = server_args.log_requests
+        self.log_requests_level = 0
 
         # Init inter-process communication
         context = zmq.asyncio.Context(2)
@@ -276,7 +277,10 @@ class TokenizerManager:
         obj.normalize_batch_and_arguments()
 
         if self.log_requests:
-            logger.info(f"Receive: obj={dataclass_to_string_truncated(obj)}")
+            max_length = 2048 if self.log_requests_level == 0 else 1 << 30
+            logger.info(
+                f"Receive: obj={dataclass_to_string_truncated(obj, max_length)}"
+            )
 
         async with self.model_update_lock.reader_lock:
             is_single = obj.is_single
@@ -419,7 +423,8 @@ class TokenizerManager:
             state.out_list = []
             if state.finished:
                 if self.log_requests:
-                    msg = f"Finish: obj={dataclass_to_string_truncated(obj)}, out={dataclass_to_string_truncated(out)}"
+                    max_length = 2048 if self.log_requests_level == 0 else 1 << 30
+                    msg = f"Finish: obj={dataclass_to_string_truncated(obj, max_length)}, out={dataclass_to_string_truncated(out, max_length)}"
                     logger.info(msg)
                 del self.rid_to_state[obj.rid]
 
@@ -682,6 +687,8 @@ class TokenizerManager:
     def configure_logging(self, obj: ConfigureLoggingReq):
         if obj.log_requests is not None:
             self.log_requests = obj.log_requests
+        if obj.log_requests_level is not None:
+            self.log_requests_level = obj.log_requests_level
         if obj.dump_requests_folder is not None:
             self.dump_requests_folder = obj.dump_requests_folder
         if obj.dump_requests_threshold is not None:

python/sglang/srt/utils.py

@@ -1262,9 +1262,9 @@ def dataclass_to_string_truncated(data, max_length=2048):
     if isinstance(data, str):
         if len(data) > max_length:
             half_length = max_length // 2
-            return f'"{data[:half_length]} ... {data[-half_length:]}"'
+            return f"{repr(data[:half_length])} ... {repr(data[-half_length:])}"
         else:
-            return f'"{data}"'
+            return f"{repr(data)}"
     elif isinstance(data, (list, tuple)):
         if len(data) > max_length:
             half_length = max_length // 2
@@ -1275,7 +1275,7 @@ def dataclass_to_string_truncated(data, max_length=2048):
         return (
             "{"
             + ", ".join(
-                f"{k}: {dataclass_to_string_truncated(v, max_length)}"
+                f"'{k}': {dataclass_to_string_truncated(v, max_length)}"
                 for k, v in data.items()
             )
             + "}"