[ perf ] Replace json-> orjson in hot path (#11221)

Signed-off-by: vincentzed <207368749+vincentzed@users.noreply.github.com>

python/sglang/srt/entrypoints/context.py

@@ -5,6 +5,8 @@ import logging
 from abc import ABC, abstractmethod
 from typing import Union
 
+import orjson
+
 logger = logging.getLogger(__name__)
 
 try:
@@ -148,7 +150,7 @@ class HarmonyContext(ConversationContext):
         if isinstance(tool_session, Tool):
             return await tool_session.get_result(self)
         tool_name = last_msg.recipient.split(".")[1]
-        args = json.loads(last_msg.content[0].text)
+        args = orjson.loads(last_msg.content[0].text)
         result = await tool_session.call_tool(tool_name, args)
         result_str = result.content[0].text
         content = TextContent(text=result_str)

python/sglang/srt/entrypoints/harmony_utils.py

@@ -7,6 +7,7 @@ import json
 from collections.abc import Iterable
 from typing import Literal, Optional, Union
 
+import orjson
 from openai.types.responses import (
     ResponseOutputItem,
     ResponseOutputMessage,
@@ -228,7 +229,7 @@ def parse_output_message(message: Message):
         if len(message.content) != 1:
             raise ValueError("Invalid number of contents in browser message")
         content = message.content[0]
-        browser_call = json.loads(content.text)
+        browser_call = orjson.loads(content.text)
         # TODO: translate to url properly!
         if recipient == "browser.search":
             action = ActionSearch(

python/sglang/srt/entrypoints/http_server.py

@@ -555,7 +555,7 @@ async def generate_request(obj: GenerateReqInput, request: Request):
 async def generate_from_file_request(file: UploadFile, request: Request):
     """Handle a generate request, this is purely to work with input_embeds."""
     content = await file.read()
-    input_embeds = json.loads(content.decode("utf-8"))
+    input_embeds = orjson.loads(content.decode("utf-8"))
 
     obj = GenerateReqInput(
         input_embeds=input_embeds,

python/sglang/srt/entrypoints/openai/serving_base.py

@@ -6,6 +6,7 @@ import uuid
 from abc import ABC, abstractmethod
 from typing import TYPE_CHECKING, Any, Optional, Union
 
+import orjson
 from fastapi import HTTPException, Request
 from fastapi.responses import ORJSONResponse, StreamingResponse
 
@@ -197,7 +198,7 @@ class OpenAIServingBase(ABC):
         )
         try:
             raw_labels = (
-                json.loads(raw_request.headers.get(header))
+                orjson.loads(raw_request.headers.get(header))
                 if raw_request and raw_request.headers.get(header)
                 else None
             )

python/sglang/srt/entrypoints/openai/serving_chat.py

@@ -7,6 +7,7 @@ import time
 import uuid
 from typing import TYPE_CHECKING, Any, AsyncGenerator, Dict, List, Optional, Union
 
+import orjson
 from fastapi import Request
 from fastapi.responses import ORJSONResponse, StreamingResponse
 from jsonschema import Draft202012Validator, SchemaError
@@ -285,7 +286,7 @@ class OpenAIServingChat(OpenAIServingBase):
                     if "arguments" in item["function"] and isinstance(
                         item["function"]["arguments"], str
                     ):
-                        item["function"]["arguments"] = json.loads(
+                        item["function"]["arguments"] = orjson.loads(
                             item["function"]["arguments"]
                         )
 
@@ -860,7 +861,7 @@ class OpenAIServingChat(OpenAIServingBase):
                 finish_reason["matched"] = None
             try:
                 # For required tool choice, we expect a JSON array of tool calls
-                tool_call_data = json.loads(text)
+                tool_call_data = orjson.loads(text)
                 tool_calls = []
                 for i, tool in enumerate(tool_call_data):
                     # Create a ToolCallItem from the JSON data

python/sglang/srt/entrypoints/openai/serving_responses.py

@@ -5,7 +5,6 @@ from __future__ import annotations
 
 import asyncio
 import copy
-import json
 import logging
 import time
 from contextlib import AsyncExitStack
@@ -14,6 +13,7 @@ from typing import TYPE_CHECKING, Any, AsyncGenerator, AsyncIterator, Optional,
 
 import jinja2
 import openai.types.responses as openai_responses_types
+import orjson
 from fastapi import Request
 from fastapi.responses import ORJSONResponse
 from openai.types.responses import (
@@ -1061,7 +1061,7 @@ class OpenAIServingResponses(OpenAIServingChat):
                 ):
                     function_name = previous_item.recipient[len("browser.") :]
                     action = None
-                    parsed_args = json.loads(previous_item.content[0].text)
+                    parsed_args = ororjson.loads(previous_item.content[0].text)
                     if function_name == "search":
                         action = openai_responses_types.response_function_web_search.ActionSearch(
                             type="search",