chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 14 Packages Projects Releases Wiki Activity

Support JSON format request logging for easier parsing (#15743)

Browse Source
This commit is contained in:
fzyzcjy
2025-12-24 19:52:43 +08:00
committed by GitHub
parent c6a6ba4368
commit e245cac0c5
5 changed files with 140 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/io_struct.py
+1
View File
@@ -1485,6 +1485,7 @@ class FreezeGCReq(BaseReq):
class ConfigureLoggingReq(BaseReq):
log_requests: Optional[bool] = None
log_requests_level: Optional[int] = None
log_requests_format: Optional[str] = None
dump_requests_folder: Optional[str] = None
dump_requests_threshold: Optional[int] = None
crash_dump_folder: Optional[str] = None
python/sglang/srt/managers/tokenizer_manager.py
+2
View File
@@ -185,6 +185,7 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
self.request_logger = RequestLogger(
log_requests=server_args.log_requests,
log_requests_level=server_args.log_requests_level,
log_requests_format=server_args.log_requests_format,
)
self.preferred_sampling_params = server_args.preferred_sampling_params
self.crash_dump_folder = server_args.crash_dump_folder
@@ -1314,6 +1315,7 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
self.request_logger.configure(
log_requests=obj.log_requests,
log_requests_level=obj.log_requests_level,
log_requests_format=obj.log_requests_format,
)
if obj.dump_requests_folder is not None:
self.dump_requests_folder = obj.dump_requests_folder
python/sglang/srt/server_args.py
+8
View File
@@ -338,6 +338,7 @@ class ServerArgs:
log_level_http: Optional[str] = None
log_requests: bool = False
log_requests_level: int = 2
log_requests_format: str = "text"
crash_dump_folder: Optional[str] = None
show_time_cost: bool = False
enable_metrics: bool = False
@@ -2908,6 +2909,13 @@ class ServerArgs:
help="0: Log metadata (no sampling parameters). 1: Log metadata and sampling parameters. 2: Log metadata, sampling parameters and partial input/output. 3: Log every input/output.",
choices=[0, 1, 2, 3],
)
parser.add_argument(
"--log-requests-format",
type=str,
default=ServerArgs.log_requests_format,
choices=["text", "json"],
help="Format for request logging: 'text' (human-readable) or 'json' (structured)",
)
parser.add_argument(
"--crash-dump-folder",
type=str,
python/sglang/srt/utils/request_logger.py
+88 -8
View File
@@ -14,7 +14,9 @@
from __future__ import annotations
import dataclasses
import json
import logging
from datetime import datetime
from functools import lru_cache
from typing import TYPE_CHECKING, Any, Optional, Set, Tuple, Union
@@ -25,11 +27,23 @@ if TYPE_CHECKING:
logger = logging.getLogger(__name__)
_json_logger = logging.getLogger(__name__ + ".json")
_json_logger.propagate = False
_json_handler = logging.StreamHandler()
_json_handler.setFormatter(logging.Formatter("%(message)s"))
_json_logger.addHandler(_json_handler)
class RequestLogger:
def __init__(self, log_requests: bool, log_requests_level: int):
def __init__(
self,
log_requests: bool,
log_requests_level: int,
log_requests_format: str,
):
self.log_requests = log_requests
self.log_requests_level = log_requests_level
self.log_requests_format = log_requests_format
self.metadata: Tuple[Optional[int], Optional[Set[str]], Optional[Set[str]]] = (
self._compute_metadata()
)
@@ -38,11 +52,14 @@ class RequestLogger:
self,
log_requests: Optional[bool] = None,
log_requests_level: Optional[int] = None,
log_requests_format: Optional[str] = None,
) -> None:
if log_requests is not None:
self.log_requests = log_requests
if log_requests_level is not None:
self.log_requests_level = log_requests_level
if log_requests_format is not None:
self.log_requests_format = log_requests_format
self.metadata = self._compute_metadata()
def log_received_request(
@@ -52,9 +69,16 @@ class RequestLogger:
return
max_length, skip_names, _ = self.metadata
logger.info(
f"Receive: obj={_dataclass_to_string_truncated(obj, max_length, skip_names=skip_names)}"
)
if self.log_requests_format == "json":
log_data = {
"rid": obj.rid,
"obj": _transform_data_for_logging(obj, max_length, skip_names),
}
_log_json("request.received", log_data)
else:
logger.info(
f"Receive: obj={_dataclass_to_string_truncated(obj, max_length, skip_names=skip_names)}"
)
# FIXME: This is a temporary fix to get the text from the input ids.
# We should remove this once we have a proper way.
@@ -77,11 +101,22 @@ class RequestLogger:
return
max_length, skip_names, out_skip_names = self.metadata
if is_multimodal_gen:
msg = f"Finish: obj={_dataclass_to_string_truncated(obj, max_length, skip_names=skip_names)}"
if self.log_requests_format == "json":
log_data = {
"rid": obj.rid,
"obj": _transform_data_for_logging(obj, max_length, skip_names),
}
if not is_multimodal_gen:
log_data["out"] = _transform_data_for_logging(
out, max_length, out_skip_names
)
_log_json("request.finished", log_data)
else:
msg = f"Finish: obj={_dataclass_to_string_truncated(obj, max_length, skip_names=skip_names)}, out={_dataclass_to_string_truncated(out, max_length, skip_names=out_skip_names)}"
logger.info(msg)
if is_multimodal_gen:
msg = f"Finish: obj={_dataclass_to_string_truncated(obj, max_length, skip_names=skip_names)}"
else:
msg = f"Finish: obj={_dataclass_to_string_truncated(obj, max_length, skip_names=skip_names)}, out={_dataclass_to_string_truncated(out, max_length, skip_names=out_skip_names)}"
logger.info(msg)
def _compute_metadata(
self,
@@ -130,6 +165,17 @@ def disable_request_logging() -> bool:
return get_bool_env_var("SGLANG_DISABLE_REQUEST_LOGGING")
# TODO unify logging, e.g. allow normal logs to be JSON as well
def _log_json(event: str, data: dict) -> None:
log_data = {
"timestamp": datetime.now().isoformat(),
"event": event,
**data,
}
_json_logger.info(json.dumps(log_data, ensure_ascii=False))
# TODO unify this w/ `_transform_data_for_logging` if we find performance enough
def _dataclass_to_string_truncated(
data: Any, max_length: int = 2048, skip_names: Optional[Set[str]] = None
) -> str:
@@ -170,3 +216,37 @@ def _dataclass_to_string_truncated(
)
else:
return str(data)
def _transform_data_for_logging(
data: Any, max_length: int = 2048, skip_names: Optional[Set[str]] = None
) -> Any:
if skip_names is None:
skip_names = set()
if isinstance(data, str):
if len(data) > max_length:
half_length = max_length // 2
return data[:half_length] + "..." + data[-half_length:]
return data
elif isinstance(data, (list, tuple)):
if len(data) > max_length:
half_length = max_length // 2
return list(data[:half_length]) + ["..."] + list(data[-half_length:])
return [_transform_data_for_logging(v, max_length) for v in data]
elif isinstance(data, dict):
return {
k: _transform_data_for_logging(v, max_length)
for k, v in data.items()
if k not in skip_names
}
elif dataclasses.is_dataclass(data):
fields = dataclasses.fields(data)
return {
f.name: _transform_data_for_logging(getattr(data, f.name), max_length)
for f in fields
if f.name not in skip_names
}
elif isinstance(data, (int, float, bool, type(None))):
return data
else:
return str(data)
test/registered/debug_utils/test_request_logger.py
+41 -5
View File
@@ -1,4 +1,5 @@
import io
import json
import unittest
import requests
@@ -12,15 +13,16 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=60, suite="nightly-1-gpu", nightly=True)
register_cuda_ci(est_time=120, suite="nightly-1-gpu", nightly=True)
class TestRequestLogger(CustomTestCase):
class BaseTestRequestLogger:
log_requests_format = None
@classmethod
def setUpClass(cls):
cls.stdout = io.StringIO()
cls.stderr = io.StringIO()
cls.process = popen_launch_server(
"Qwen/Qwen3-0.6B",
DEFAULT_URL_FOR_TEST,
@@ -29,6 +31,8 @@ class TestRequestLogger(CustomTestCase):
"--log-requests",
"--log-requests-level",
"2",
"--log-requests-format",
cls.log_requests_format,
"--skip-server-warmup",
],
return_stdout_stderr=(cls.stdout, cls.stderr),
@@ -40,7 +44,7 @@ class TestRequestLogger(CustomTestCase):
cls.stdout.close()
cls.stderr.close()
def test_request_logging(self):
def _send_request(self):
response = requests.post(
DEFAULT_URL_FOR_TEST + "/generate",
json={
@@ -50,11 +54,43 @@ class TestRequestLogger(CustomTestCase):
timeout=30,
)
self.assertEqual(response.status_code, 200)
return self.stdout.getvalue() + self.stderr.getvalue()
combined_output = self.stdout.getvalue() + self.stderr.getvalue()
class TestRequestLoggerText(BaseTestRequestLogger, CustomTestCase):
log_requests_format = "text"
def test_text_format_logging(self):
combined_output = self._send_request()
self.assertIn("Receive:", combined_output)
self.assertIn("Finish:", combined_output)
class TestRequestLoggerJson(BaseTestRequestLogger, CustomTestCase):
log_requests_format = "json"
def test_json_format_logging(self):
combined_output = self._send_request()
received_found = False
finished_found = False
for line in combined_output.splitlines():
if not line.startswith("{"):
continue
data = json.loads(line)
if data.get("event") == "request.received":
self.assertIn("rid", data)
self.assertIn("obj", data)
received_found = True
elif data.get("event") == "request.finished":
self.assertIn("rid", data)
self.assertIn("obj", data)
self.assertIn("out", data)
finished_found = True
self.assertTrue(received_found, "request.received event not found in logs")
self.assertTrue(finished_found, "request.finished event not found in logs")
if __name__ == "__main__":
unittest.main()