sglang 0.4.6.post3__py3-none-any.whl → 0.4.6.post5__py3-none-any.whl

sglang/srt/entrypoints/http_server.py

@@ -47,8 +47,9 @@ from sglang.srt.disaggregation.utils import (
     register_disaggregation_server,
 )
 from sglang.srt.entrypoints.engine import _launch_subprocesses
-from sglang.srt.function_call_parser import FunctionCallParser
+from sglang.srt.function_call.function_call_parser import FunctionCallParser
 from sglang.srt.managers.io_struct import (
+    AbortReq,
     CloseSessionReqInput,
     ConfigureLoggingReq,
     EmbeddingReqInput,
@@ -181,13 +182,14 @@ async def health_generate(request: Request) -> Response:
         async for _ in _global_state.tokenizer_manager.generate_request(gri, request):
             break
 
-    tic = time.time()
+    tic = time.perf_counter()
     task = asyncio.create_task(gen())
-    while time.time() < tic + HEALTH_CHECK_TIMEOUT:
+    while time.perf_counter() < tic + HEALTH_CHECK_TIMEOUT:
         await asyncio.sleep(1)
         if _global_state.tokenizer_manager.last_receive_tstamp > tic:
             task.cancel()
             _global_state.tokenizer_manager.rid_to_state.pop(rid, None)
+            _global_state.tokenizer_manager.health_check_failed = False
             return Response(status_code=200)
 
     task.cancel()
@@ -201,6 +203,7 @@ async def health_generate(request: Request) -> Response:
         f"last_heartbeat time: {last_receive_time}"
     )
     _global_state.tokenizer_manager.rid_to_state.pop(rid, None)
+    _global_state.tokenizer_manager.health_check_failed = True
     return Response(status_code=503)
 
 
@@ -221,7 +224,7 @@ async def get_server_info():
     return {
         **dataclasses.asdict(_global_state.tokenizer_manager.server_args),
         **_global_state.scheduler_info,
-        **internal_states,
+        "internal_states": internal_states,
         "version": __version__,
     }
 
@@ -337,7 +340,11 @@ async def start_profile_async(obj: Optional[ProfileReqInput] = None):
         obj = ProfileReqInput()
 
     await _global_state.tokenizer_manager.start_profile(
-        obj.output_dir, obj.num_steps, obj.activities
+        output_dir=obj.output_dir,
+        num_steps=obj.num_steps,
+        activities=obj.activities,
+        with_stack=obj.with_stack,
+        record_shapes=obj.record_shapes,
     )
     return Response(
         content="Start profiling.\n",
@@ -348,7 +355,7 @@ async def start_profile_async(obj: Optional[ProfileReqInput] = None):
 @app.api_route("/stop_profile", methods=["GET", "POST"])
 async def stop_profile_async():
     """Stop profiling."""
-    _global_state.tokenizer_manager.stop_profile()
+    await _global_state.tokenizer_manager.stop_profile()
     return Response(
         content="Stop profiling. This will take some time.\n",
         status_code=200,
@@ -539,6 +546,16 @@ async def configure_logging(obj: ConfigureLoggingReq, request: Request):
     return Response(status_code=200)
 
 
+@app.post("/abort_request")
+async def abort_request(obj: AbortReq, request: Request):
+    """Abort a request."""
+    try:
+        _global_state.tokenizer_manager.abort_request(rid=obj.rid)
+        return Response(status_code=200)
+    except Exception as e:
+        return _create_error_response(e)
+
+
 @app.post("/parse_function_call")
 async def parse_function_call_request(obj: ParseFunctionCallReq, request: Request):
     """

sglang/srt/entrypoints/http_server_engine.py

@@ -24,10 +24,10 @@ def launch_server_process(server_args: ServerArgs) -> multiprocessing.Process:
 
     base_url = server_args.url()
     timeout = 300.0  # Increased timeout to 5 minutes for downloading large models
-    start_time = time.time()
+    start_time = time.perf_counter()
 
     with requests.Session() as session:
-        while time.time() - start_time < timeout:
+        while time.perf_counter() - start_time < timeout:
             try:
                 headers = {
                     "Content-Type": "application/json; charset=utf-8",
@@ -140,3 +140,6 @@ class HttpServerEngineAdapter(EngineBase):
 
     def resume_memory_occupation(self):
         return self._make_request("resume_memory_occupation")
+
+    def flush_cache(self):
+        return self._make_request("flush_cache")

sglang/srt/function_call/base_format_detector.py

@@ -0,0 +1,250 @@
+import json
+import logging
+from abc import ABC, abstractmethod
+from typing import Any, Dict, List
+
+from partial_json_parser.core.exceptions import MalformedJSON
+from partial_json_parser.core.options import Allow
+
+from sglang.srt.function_call.core_types import (
+    StreamingParseResult,
+    ToolCallItem,
+    _GetInfoFunc,
+)
+from sglang.srt.function_call.utils import (
+    _find_common_prefix,
+    _is_complete_json,
+    _partial_json_loads,
+)
+from sglang.srt.openai_api.protocol import Tool
+
+logger = logging.getLogger(__name__)
+
+
+class BaseFormatDetector(ABC):
+    """Base class providing two sets of interfaces: one-time and streaming incremental."""
+
+    def __init__(self):
+        # initialize properties used for state when parsing tool calls in
+        self._buffer = ""
+        # streaming mode
+        self.prev_tool_call_arr: List[Dict] = []
+        self.current_tool_id: int = -1
+        self.current_tool_name_sent: bool = False
+        self.streamed_args_for_tool: List[str] = (
+            []
+        )  # map what has been streamed for each tool so far to a list
+        self.bot_token = ""
+        self.eot_token = ""
+
+    def parse_base_json(self, action: Any, tools: List[Tool]) -> List[ToolCallItem]:
+        tool_indices = {
+            tool.function.name: i for i, tool in enumerate(tools) if tool.function.name
+        }
+        if not isinstance(action, list):
+            action = [action]
+
+        results = []
+        for act in action:
+            name = act.get("name")
+            if name and name in tool_indices:
+                results.append(
+                    ToolCallItem(
+                        tool_index=tool_indices[name],
+                        name=name,
+                        parameters=json.dumps(
+                            act.get("parameters") or act.get("arguments", {}),
+                            ensure_ascii=False,
+                        ),
+                    )
+                )
+            else:
+                logger.warning(f"Model attempted to call undefined function: {name}")
+
+        return results
+
+    @abstractmethod
+    def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
+        """
+        Parses the text in one go. Returns success=True if the format matches, otherwise False.
+        Note that leftover_text here represents "content that this parser will not consume further".
+        """
+        action = json.loads(text)
+        return StreamingParseResult(calls=self.parse_base_json(action, tools))
+
+    def parse_streaming_increment(
+        self, new_text: str, tools: List[Tool]
+    ) -> StreamingParseResult:
+        """
+        Streaming incremental parsing with tool validation.
+        """
+        # Append new text to buffer
+        self._buffer += new_text
+        current_text = self._buffer
+        if not (self.bot_token in current_text or current_text.startswith("{")):
+            self._buffer = ""
+            if self.eot_token in new_text:
+                new_text = new_text.replace(self.eot_token, "")
+            return StreamingParseResult(normal_text=new_text)
+
+        # Build tool indices if not already built
+        if not hasattr(self, "_tool_indices"):
+            self._tool_indices = {
+                tool.function.name: i
+                for i, tool in enumerate(tools)
+                if tool.function and tool.function.name
+            }
+
+        flags = Allow.ALL if self.current_tool_name_sent else Allow.ALL & ~Allow.STR
+        try:
+            tool_call_arr = []
+            is_complete = []
+            try:
+                start_idx = (
+                    len(self.bot_token)
+                    if current_text.startswith(self.bot_token)
+                    else 0
+                )
+                while start_idx < len(current_text):
+                    (obj, end_idx) = _partial_json_loads(
+                        current_text[start_idx:], flags
+                    )
+                    is_complete.append(
+                        _is_complete_json(current_text[start_idx : start_idx + end_idx])
+                    )
+                    start_idx += end_idx + len("; ")
+
+                    # Validate tool name if present
+                    if "name" in obj and obj["name"] not in self._tool_indices:
+                        # Invalid tool name - reset state
+                        self._buffer = ""
+                        self.current_tool_id = -1
+                        self.current_tool_name_sent = False
+                        if self.streamed_args_for_tool:
+                            self.streamed_args_for_tool.pop()
+                        return StreamingParseResult()
+
+                    # Handle parameters/arguments consistency
+                    if "parameters" in obj:
+                        assert (
+                            "arguments" not in obj
+                        ), "model generated both parameters and arguments"
+                        obj["arguments"] = obj["parameters"]
+                    tool_call_arr.append(obj)
+
+            except MalformedJSON:
+                return StreamingParseResult()
+
+            if len(tool_call_arr) == 0:
+                return StreamingParseResult()
+
+            current_tool_call: Dict = (
+                tool_call_arr[self.current_tool_id] if len(tool_call_arr) > 0 else {}
+            )
+
+            # Handle new tool in array
+            if len(tool_call_arr) > 0 and len(tool_call_arr) > self.current_tool_id + 1:
+                if self.current_tool_id >= 0:
+                    cur_arguments = current_tool_call.get("arguments")
+                    if cur_arguments:
+                        cur_args_json = json.dumps(cur_arguments)
+                        sent = len(self.streamed_args_for_tool[self.current_tool_id])
+                        argument_diff = cur_args_json[sent:]
+
+                        res = StreamingParseResult(
+                            calls=[
+                                ToolCallItem(
+                                    tool_index=self.current_tool_id,
+                                    name="",
+                                    parameters=argument_diff,
+                                )
+                            ],
+                        )
+                        self.streamed_args_for_tool[
+                            self.current_tool_id
+                        ] += argument_diff
+                    else:
+                        res = StreamingParseResult()
+                else:
+                    res = StreamingParseResult()
+
+                self.current_tool_id = len(tool_call_arr) - 1
+                self.current_tool_name_sent = False
+                self.streamed_args_for_tool.append("")
+                return res
+
+            # Handle tool name
+            elif not self.current_tool_name_sent:
+                function_name = current_tool_call.get("name")
+                if function_name and function_name in self._tool_indices:
+                    res = StreamingParseResult(
+                        calls=[
+                            ToolCallItem(
+                                tool_index=self._tool_indices[function_name],
+                                name=function_name,
+                                parameters="",
+                            )
+                        ],
+                    )
+                    self.current_tool_name_sent = True
+                else:
+                    res = StreamingParseResult()
+
+            # Handle streaming arguments
+            else:
+                cur_arguments = current_tool_call.get("arguments")
+                res = StreamingParseResult()
+
+                if cur_arguments:
+                    sent = len(self.streamed_args_for_tool[self.current_tool_id])
+                    cur_args_json = json.dumps(cur_arguments)
+                    prev_arguments = self.prev_tool_call_arr[self.current_tool_id].get(
+                        "arguments"
+                    )
+
+                    argument_diff = None
+                    if is_complete[self.current_tool_id]:
+                        argument_diff = cur_args_json[sent:]
+                        self._buffer = ""
+                        self.prev_tool_call_arr[self.current_tool_id].clear()
+                        self.current_tool_name_sent = False
+                        self.streamed_args_for_tool[self.current_tool_id] = ""
+
+                    elif prev_arguments:
+                        prev_args_json = json.dumps(prev_arguments)
+                        if cur_args_json != prev_args_json:
+                            prefix = _find_common_prefix(prev_args_json, cur_args_json)
+                            argument_diff = prefix[sent:]
+
+                    if argument_diff is not None:
+                        res = StreamingParseResult(
+                            calls=[
+                                ToolCallItem(
+                                    tool_index=self.current_tool_id,
+                                    parameters=argument_diff,
+                                )
+                            ],
+                        )
+                        if not is_complete[self.current_tool_id]:
+                            self.streamed_args_for_tool[
+                                self.current_tool_id
+                            ] += argument_diff
+
+            self.prev_tool_call_arr = tool_call_arr
+            return res
+
+        except Exception as e:
+            logger.error(f"Error in parse_streaming_increment: {e}")
+            return StreamingParseResult()
+
+    @abstractmethod
+    def has_tool_call(self, text: str) -> bool:
+        raise NotImplementedError()
+
+    @abstractmethod
+    def structure_info(self) -> _GetInfoFunc:
+        raise NotImplementedError()
+
+    @abstractmethod
+    def build_ebnf(self, tools: List[Tool]) -> str:
+        raise NotImplementedError()

sglang/srt/function_call/core_types.py

@@ -0,0 +1,34 @@
+from dataclasses import dataclass
+from typing import Callable, List, Optional
+
+from pydantic import BaseModel
+
+
+class ToolCallItem(BaseModel):
+    """Simple encapsulation of the parsed ToolCall result for easier usage in streaming contexts."""
+
+    tool_index: int
+    name: Optional[str] = None
+    parameters: str  # JSON string
+
+
+class StreamingParseResult(BaseModel):
+    """Result of streaming incremental parsing."""
+
+    normal_text: str = ""
+    calls: List[ToolCallItem] = []
+
+
+@dataclass
+class StructureInfo:
+    begin: str
+    end: str
+    trigger: str
+
+
+"""
+Helper alias of function
+Usually it is a function that takes a name string and returns a StructureInfo object,
+which can be used to construct a structural_tag object
+"""
+_GetInfoFunc = Callable[[str], StructureInfo]

sglang/srt/function_call/deepseekv3_detector.py

@@ -0,0 +1,157 @@
+import json
+import logging
+import re
+from typing import List
+
+from sglang.srt.function_call.base_format_detector import BaseFormatDetector
+from sglang.srt.function_call.core_types import (
+    StreamingParseResult,
+    StructureInfo,
+    ToolCallItem,
+    _GetInfoFunc,
+)
+from sglang.srt.function_call.ebnf_composer import EBNFComposer
+from sglang.srt.function_call.utils import _is_complete_json
+from sglang.srt.openai_api.protocol import Tool
+
+logger = logging.getLogger(__name__)
+
+
+class DeepSeekV3Detector(BaseFormatDetector):
+    """
+    Detector for DeepSeek models.
+    Assumes function call format:
+      '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_current_weather\n```json\n{"location": "Tokyo"}\n```<｜tool▁call▁end｜>\n<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_current_weather\n```json\n{"location": "Paris"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜>
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.bot_token = "<｜tool▁calls▁begin｜>"
+        self.eot_token = "<｜tool▁calls▁end｜>"
+        self.func_call_regex = r"<｜tool▁call▁begin｜>.*?<｜tool▁call▁end｜>"
+        self.func_detail_regex = r"<｜tool▁call▁begin｜>(.*)<｜tool▁sep｜>(.*)\n```json\n(.*)\n```<｜tool▁call▁end｜>"
+        self._last_arguments = ""
+
+    def has_tool_call(self, text: str) -> bool:
+        """Check if the text contains a deepseek format tool call."""
+        return self.bot_token in text
+
+    def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
+        """
+        One-time parsing: Detects and parses tool calls in the provided text.
+
+        :param text: The complete text to parse.
+        :param tools: List of available tools.
+        :return: ParseResult indicating success or failure, consumed text, leftover text, and parsed calls.
+        """
+        idx = text.find(self.bot_token)
+        normal_text = text[:idx].strip() if idx != -1 else text
+        if self.bot_token not in text:
+            return StreamingParseResult(normal_text=normal_text, calls=[])
+        match_result_list = re.findall(self.func_call_regex, text, re.DOTALL)
+        calls = []
+        try:
+            for match_result in match_result_list:
+                # Get function name
+                func_detail = re.search(self.func_detail_regex, match_result, re.DOTALL)
+                func_name = func_detail.group(2)
+                func_args = func_detail.group(3)
+                func_args = json.loads(func_args)
+                # construct match_result for parse_base_json
+                match_result = {"name": func_name, "parameters": func_args}
+                calls.extend(self.parse_base_json(match_result, tools))
+            return StreamingParseResult(normal_text=normal_text, calls=calls)
+        except Exception as e:
+            logger.error(f"Error in detect_and_parse: {e}")
+            # return the normal text if parsing fails
+            return StreamingParseResult(normal_text=text)
+
+    def parse_streaming_increment(
+        self, new_text: str, tools: List[Tool]
+    ) -> StreamingParseResult:
+        """
+        Streaming incremental parsing tool calls for DeepSeekV3 format.
+        """
+        self._buffer += new_text
+        current_text = self._buffer
+
+        if self.bot_token not in current_text:
+            self._buffer = ""
+            for e_token in [self.eot_token, "```", "<｜tool▁call▁end｜>"]:
+                if e_token in new_text:
+                    new_text = new_text.replace(e_token, "")
+            return StreamingParseResult(normal_text=new_text)
+
+        if not hasattr(self, "_tool_indices"):
+            self._tool_indices = {
+                tool.function.name: i
+                for i, tool in enumerate(tools)
+                if tool.function and tool.function.name
+            }
+
+        calls: list[ToolCallItem] = []
+        try:
+            partial_match = re.search(
+                pattern=r"<｜tool▁call▁begin｜>(.*)<｜tool▁sep｜>(.*)\n```json\n(.*)",
+                string=current_text,
+                flags=re.DOTALL,
+            )
+            if partial_match:
+                func_name = partial_match.group(2).strip()
+                func_args_raw = partial_match.group(3).strip()
+
+                if not self.current_tool_name_sent:
+                    calls.append(
+                        ToolCallItem(
+                            tool_index=self._tool_indices.get(func_name, 0),
+                            name=func_name,
+                            parameters="",
+                        )
+                    )
+                    self.current_tool_name_sent = True
+                else:
+                    argument_diff = (
+                        func_args_raw[len(self._last_arguments) :]
+                        if func_args_raw.startswith(self._last_arguments)
+                        else func_args_raw
+                    )
+
+                    if argument_diff:
+                        calls.append(
+                            ToolCallItem(
+                                tool_index=self._tool_indices.get(func_name, 0),
+                                name=None,
+                                parameters=argument_diff,
+                            )
+                        )
+                        self._last_arguments += argument_diff
+
+                    if _is_complete_json(func_args_raw):
+                        result = StreamingParseResult(normal_text="", calls=calls)
+                        self._buffer = ""
+                        self._last_arguments = ""
+                        self.current_tool_name_sent = False
+                        return result
+
+            return StreamingParseResult(normal_text="", calls=calls)
+
+        except Exception as e:
+            logger.error(f"Error in parse_streaming_increment: {e}")
+            return StreamingParseResult(normal_text=current_text)
+
+    def structure_info(self) -> _GetInfoFunc:
+        return lambda name: StructureInfo(
+            begin=">" + name + "\n```json\n",
+            end="\n```<",
+            trigger=">" + name + "\n```json\n",
+        )
+
+    def build_ebnf(self, tools: List[Tool]):
+        return EBNFComposer.build_ebnf(
+            tools,
+            bot_token=self.bot_token,
+            eot_token=self.eot_token,
+            tool_call_separator="",
+            call_rule_fmt='"<｜tool▁call▁begin｜>function<｜tool▁sep｜>{name}\\n```json\\n" {arguments_rule} "\\n```<｜tool▁call▁end｜>"',
+            function_format="json",
+        )