Fix /generate JSON serialization for non-finite top_logprobs (#20714)
This commit is contained in:
@@ -42,7 +42,6 @@ setattr(threading, "_register_atexit", lambda *args, **kwargs: None)
|
||||
|
||||
|
||||
import numpy as np
|
||||
import orjson
|
||||
import requests
|
||||
import uvicorn
|
||||
import uvloop
|
||||
@@ -173,6 +172,11 @@ from sglang.srt.utils import (
|
||||
set_uvicorn_logging_configs,
|
||||
)
|
||||
from sglang.srt.utils.auth import AuthLevel, app_has_admin_force_endpoints, auth_level
|
||||
from sglang.srt.utils.json_response import (
|
||||
SGLangORJSONResponse,
|
||||
dumps_json,
|
||||
orjson_response,
|
||||
)
|
||||
from sglang.utils import get_exception_traceback
|
||||
from sglang.version import __version__
|
||||
|
||||
@@ -666,7 +670,11 @@ if os.environ.get("DUMPER_SERVER_PORT") == "reuse":
|
||||
|
||||
|
||||
# fastapi implicitly converts json in the request to obj (dataclass)
|
||||
@app.api_route("/generate", methods=["POST", "PUT"])
|
||||
@app.api_route(
|
||||
"/generate",
|
||||
methods=["POST", "PUT"],
|
||||
response_class=SGLangORJSONResponse,
|
||||
)
|
||||
async def generate_request(obj: GenerateReqInput, request: Request):
|
||||
"""Handle a generate request."""
|
||||
if obj.stream:
|
||||
@@ -676,15 +684,11 @@ async def generate_request(obj: GenerateReqInput, request: Request):
|
||||
async for out in _global_state.tokenizer_manager.generate_request(
|
||||
obj, request
|
||||
):
|
||||
yield b"data: " + orjson.dumps(
|
||||
out, option=orjson.OPT_NON_STR_KEYS | orjson.OPT_SERIALIZE_NUMPY
|
||||
) + b"\n\n"
|
||||
yield b"data: " + dumps_json(out) + b"\n\n"
|
||||
except ValueError as e:
|
||||
out = {"error": {"message": str(e)}}
|
||||
logger.error(f"[http_server] Error: {e}")
|
||||
yield b"data: " + orjson.dumps(
|
||||
out, option=orjson.OPT_NON_STR_KEYS | orjson.OPT_SERIALIZE_NUMPY
|
||||
) + b"\n\n"
|
||||
yield b"data: " + dumps_json(out) + b"\n\n"
|
||||
yield b"data: [DONE]\n\n"
|
||||
|
||||
return StreamingResponse(
|
||||
@@ -697,12 +701,7 @@ async def generate_request(obj: GenerateReqInput, request: Request):
|
||||
ret = await _global_state.tokenizer_manager.generate_request(
|
||||
obj, request
|
||||
).__anext__()
|
||||
return Response(
|
||||
content=orjson.dumps(
|
||||
ret, option=orjson.OPT_NON_STR_KEYS | orjson.OPT_SERIALIZE_NUMPY
|
||||
),
|
||||
media_type="application/json",
|
||||
)
|
||||
return orjson_response(ret)
|
||||
except ValueError as e:
|
||||
logger.error(f"[http_server] Error: {e}")
|
||||
return _create_error_response(e)
|
||||
|
||||
32
python/sglang/srt/utils/json_response.py
Normal file
32
python/sglang/srt/utils/json_response.py
Normal file
@@ -0,0 +1,32 @@
|
||||
"""Utilities for JSON serialization in HTTP responses."""
|
||||
|
||||
from typing import Any
|
||||
|
||||
import orjson
|
||||
from fastapi.responses import ORJSONResponse, Response
|
||||
|
||||
# Keep response serialization behavior consistent across endpoints:
|
||||
# - Support non-string dictionary keys used in some metadata payloads.
|
||||
# - Support numpy scalars/arrays without pre-conversion.
|
||||
ORJSON_RESPONSE_OPTIONS = orjson.OPT_NON_STR_KEYS | orjson.OPT_SERIALIZE_NUMPY
|
||||
|
||||
|
||||
def dumps_json(content: Any) -> bytes:
|
||||
"""Serialize content to JSON bytes using SGLang's ORJSON options."""
|
||||
return orjson.dumps(content, option=ORJSON_RESPONSE_OPTIONS)
|
||||
|
||||
|
||||
class SGLangORJSONResponse(ORJSONResponse):
|
||||
"""ORJSON response with SGLang-specific serialization options."""
|
||||
|
||||
def render(self, content: Any) -> bytes:
|
||||
return dumps_json(content)
|
||||
|
||||
|
||||
def orjson_response(content: Any, status_code: int = 200) -> Response:
|
||||
"""Create a JSON response with stable ORJSON serialization options."""
|
||||
return Response(
|
||||
content=dumps_json(content),
|
||||
media_type="application/json",
|
||||
status_code=status_code,
|
||||
)
|
||||
Reference in New Issue
Block a user