sglang 0.4.4.post1__py3-none-any.whl → 0.4.4.post2__py3-none-any.whl

sglang/srt/entrypoints/engine.py

@@ -27,12 +27,16 @@ import signal
 import threading
 from typing import AsyncIterator, Dict, Iterator, List, Optional, Tuple, Union
 
+import zmq
+import zmq.asyncio
+
 # Fix a bug of Python threading
 setattr(threading, "_register_atexit", lambda *args, **kwargs: None)
 
 import torch
 import uvloop
 
+from sglang.srt.code_completion_parser import load_completion_template_for_openai_api
 from sglang.srt.managers.data_parallel_controller import (
     run_data_parallel_controller_process,
 )
@@ -44,6 +48,8 @@ from sglang.srt.managers.io_struct import (
     InitWeightsUpdateGroupReqInput,
     ReleaseMemoryOccupationReqInput,
     ResumeMemoryOccupationReqInput,
+    RpcReqInput,
+    RpcReqOutput,
     UpdateWeightFromDiskReqInput,
     UpdateWeightsFromDistributedReqInput,
     UpdateWeightsFromTensorReqInput,
@@ -57,6 +63,7 @@ from sglang.srt.utils import (
     MultiprocessingSerializer,
     assert_pkg_version,
     configure_logger,
+    get_zmq_socket,
     kill_process_tree,
     launch_dummy_health_check_server,
     maybe_set_triton_cache_manager,
@@ -102,15 +109,25 @@ class Engine:
         # Shutdown the subprocesses automatically when the program exits
         atexit.register(self.shutdown)
 
+        # Allocate ports for inter-process communications
+        port_args = PortArgs.init_new(server_args)
+        logger.info(f"{server_args=}")
+
         # Launch subprocesses
         tokenizer_manager, scheduler_info = _launch_subprocesses(
-            server_args=server_args
+            server_args=server_args,
+            port_args=port_args,
         )
 
         self.server_args = server_args
         self.tokenizer_manager = tokenizer_manager
         self.scheduler_info = scheduler_info
 
+        context = zmq.Context(2)
+        self.send_to_rpc = get_zmq_socket(
+            context, zmq.DEALER, port_args.rpc_ipc_name, True
+        )
+
     def generate(
         self,
         # The input prompt. It can be a single prompt or a batch of prompts.
@@ -232,6 +249,13 @@ class Engine:
         """Shutdown the engine"""
         kill_process_tree(os.getpid(), include_parent=False)
 
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        self.shutdown()
+        return False
+
     def start_profile(self):
         loop = asyncio.get_event_loop()
         loop.run_until_complete(self.tokenizer_manager.start_profile())
@@ -296,7 +320,10 @@ class Engine:
         """Update weights from distributed source. If there are going to be more updates, set `flush_cache` to be true
         to avoid duplicated operations such as clearing cache."""
         obj = UpdateWeightsFromTensorReqInput(
-            serialized_named_tensors=MultiprocessingSerializer.serialize(named_tensors),
+            serialized_named_tensors=[
+                MultiprocessingSerializer.serialize(named_tensors)
+                for _ in range(self.server_args.tp_size)
+            ],
             load_format=load_format,
             flush_cache=flush_cache,
         )
@@ -350,6 +377,23 @@ class Engine:
             self.tokenizer_manager.resume_memory_occupation(obj, None)
         )
 
+    """
+    Execute an RPC call on all scheduler processes.
+    """
+
+    def collective_rpc(self, method: str, **kwargs):
+        obj = RpcReqInput(method=method, parameters=kwargs)
+        self.send_to_rpc.send_pyobj(obj)
+        recv_req = self.send_to_rpc.recv_pyobj(zmq.BLOCKY)
+        assert isinstance(recv_req, RpcReqOutput)
+        assert recv_req.success, recv_req.message
+
+    def save_remote_model(self, **kwargs):
+        self.collective_rpc("save_remote_model", **kwargs)
+
+    def save_sharded_model(self, **kwargs):
+        self.collective_rpc("save_sharded_model", **kwargs)
+
 
 def _set_envs_and_config(server_args: ServerArgs):
     # Set global environments
@@ -408,7 +452,9 @@ def _set_envs_and_config(server_args: ServerArgs):
     mp.set_start_method("spawn", force=True)
 
 
-def _launch_subprocesses(server_args: ServerArgs) -> Tuple[TokenizerManager, Dict]:
+def _launch_subprocesses(
+    server_args: ServerArgs, port_args: Optional[PortArgs] = None
+) -> Tuple[TokenizerManager, Dict]:
     """
     Launch the TokenizerManager in the main process, the Scheduler in a subprocess, and the DetokenizerManager in another subprocess.
     """
@@ -418,8 +464,9 @@ def _launch_subprocesses(server_args: ServerArgs) -> Tuple[TokenizerManager, Dic
     _set_envs_and_config(server_args)
 
     # Allocate ports for inter-process communications
-    port_args = PortArgs.init_new(server_args)
-    logger.info(f"{server_args=}")
+    if port_args is None:
+        port_args = PortArgs.init_new(server_args)
+        logger.info(f"{server_args=}")
 
     # If using model from www.modelscope.cn, first download the model.
     server_args.model_path, server_args.tokenizer_path = prepare_model_and_tokenizer(
@@ -502,6 +549,9 @@ def _launch_subprocesses(server_args: ServerArgs) -> Tuple[TokenizerManager, Dic
             tokenizer_manager, server_args.chat_template, server_args.model_path
         )
 
+    if server_args.completion_template:
+        load_completion_template_for_openai_api(server_args.completion_template)
+
     # Wait for the model to finish loading
     scheduler_infos = []
     for i in range(len(scheduler_pipe_readers)):

sglang/srt/entrypoints/http_server.py

@@ -14,11 +14,12 @@
 """
 The entry point of inference server. (SRT = SGLang Runtime)
 
-This file implements HTTP APIs for the inferenc engine via fastapi.
+This file implements HTTP APIs for the inference engine via fastapi.
 """
 
 import asyncio
 import dataclasses
+import json
 import logging
 import multiprocessing as multiprocessing
 import os
@@ -259,6 +260,29 @@ async def generate_request(obj: GenerateReqInput, request: Request):
             return _create_error_response(e)
 
 
+@app.api_route("/generate_from_file", methods=["POST"])
+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"))
+
+    obj = GenerateReqInput(
+        input_embeds=input_embeds,
+        sampling_params={
+            "repetition_penalty": 1.2,
+            "temperature": 0.2,
+            "max_new_tokens": 512,
+        },
+    )
+
+    try:
+        ret = await _global_state.generate_request(obj, request).__anext__()
+        return ret
+    except ValueError as e:
+        logger.error(f"Error: {e}")
+        return _create_error_response(e)
+
+
 @app.api_route("/encode", methods=["POST", "PUT"])
 async def encode_request(obj: EmbeddingReqInput, request: Request):
     """Handle an embedding request."""
@@ -283,7 +307,7 @@ async def classify_request(obj: EmbeddingReqInput, request: Request):
         return _create_error_response(e)
 
 
-@app.post("/flush_cache")
+@app.api_route("/flush_cache", methods=["GET", "POST"])
 async def flush_cache():
     """Flush the radix cache."""
     _global_state.tokenizer_manager.flush_cache()
@@ -319,6 +343,36 @@ async def stop_profile_async():
     )
 
 
+@app.api_route("/start_expert_distribution_record", methods=["GET", "POST"])
+async def start_expert_distribution_record_async():
+    """Start recording the expert distribution. Clear the previous record if any."""
+    await _global_state.tokenizer_manager.start_expert_distribution_record()
+    return Response(
+        content="Start recording the expert distribution.\n",
+        status_code=200,
+    )
+
+
+@app.api_route("/stop_expert_distribution_record", methods=["GET", "POST"])
+async def stop_expert_distribution_record_async():
+    """Stop recording the expert distribution."""
+    await _global_state.tokenizer_manager.stop_expert_distribution_record()
+    return Response(
+        content="Stop recording the expert distribution.\n",
+        status_code=200,
+    )
+
+
+@app.api_route("/dump_expert_distribution_record", methods=["GET", "POST"])
+async def dump_expert_distribution_record_async():
+    """Dump expert distribution record."""
+    await _global_state.tokenizer_manager.dump_expert_distribution_record()
+    return Response(
+        content="Dump expert distribution record.\n",
+        status_code=200,
+    )
+
+
 @app.post("/update_weights_from_disk")
 async def update_weights_from_disk(obj: UpdateWeightFromDiskReqInput, request: Request):
     """Update the weights from disk inplace without re-launching the server."""
@@ -706,9 +760,15 @@ def _wait_and_warmup(
         },
     }
     if server_args.skip_tokenizer_init:
-        json_data["input_ids"] = [10, 11, 12]
+        json_data["input_ids"] = [[10, 11, 12] for _ in range(server_args.dp_size)]
+        # TODO Workaround the bug that embedding errors for list of size 1
+        if server_args.dp_size == 1:
+            json_data["input_ids"] = json_data["input_ids"][0]
     else:
-        json_data["text"] = "The capital city of France is"
+        json_data["text"] = ["The capital city of France is"] * server_args.dp_size
+        # TODO Workaround the bug that embedding errors for list of size 1
+        if server_args.dp_size == 1:
+            json_data["text"] = json_data["text"][0]
 
     # Debug dumping
     if server_args.debug_tensor_dump_input_file:
@@ -719,14 +779,13 @@ def _wait_and_warmup(
         json_data["sampling_params"]["max_new_tokens"] = 0
 
     try:
-        for i in range(server_args.dp_size):
-            res = requests.post(
-                url + request_name,
-                json=json_data,
-                headers=headers,
-                timeout=600,
-            )
-            assert res.status_code == 200, f"{res}"
+        res = requests.post(
+            url + request_name,
+            json=json_data,
+            headers=headers,
+            timeout=600,
+        )
+        assert res.status_code == 200, f"{res}"
     except Exception:
         last_traceback = get_exception_traceback()
         if pipe_finish_writer is not None:

sglang/srt/function_call_parser.py

@@ -1,12 +1,21 @@
 import json
 import logging
 import re
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
 from json import JSONDecodeError, JSONDecoder
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Type
 
 import partial_json_parser
+from partial_json_parser.core.exceptions import MalformedJSON
 from partial_json_parser.core.options import Allow
-from pydantic import BaseModel, Field
+from pydantic import BaseModel
+
+from sglang.srt.openai_api.protocol import (
+    StructuralTagResponseFormat,
+    StructuresResponseFormat,
+    Tool,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -19,14 +28,6 @@ TOOLS_TAG_LIST = [
 ]
 
 
-class Function(BaseModel):
-    """Function Tool Template."""
-
-    description: Optional[str] = Field(default=None, examples=[None])
-    name: Optional[str] = None
-    parameters: Optional[object] = None
-
-
 class ToolCallItem(BaseModel):
     """Simple encapsulation of the parsed ToolCall result for easier usage in streaming contexts."""
 
@@ -74,7 +75,22 @@ class StreamingParseResult:
         self.calls = calls or []
 
 
-class BaseFormatDetector:
+@dataclass
+class StructureInfo:
+    begin: str
+    end: str
+    trigger: str
+
+
+_GetInfoFunc = Callable[[str], StructureInfo]
+"""
+helper alias of function
+ususally it is a function that takes a name string and returns a StructureInfo object,
+which can be used to construct a structural_tag object
+"""
+
+
+class BaseFormatDetector(ABC):
     """Base class providing two sets of interfaces: one-time and streaming incremental."""
 
     def __init__(self):
@@ -90,26 +106,12 @@ class BaseFormatDetector:
         self.bot_token = ""
         self.eot_token = ""
 
-    def parse_base_json(self, action: Any, tools: List[Function]) -> List[ToolCallItem]:
+    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):
-            name = action.get("name")
-            if not name or name not in tool_indices:
-                logger.warning(f"Model attempted to call undefined function: {name}")
-                return []
-
-            return [
-                ToolCallItem(
-                    tool_index=tool_indices[name],
-                    name=name,
-                    parameters=json.dumps(
-                        action.get("parameters") or action.get("arguments", {}),
-                        ensure_ascii=False,
-                    ),
-                )
-            ]
+            action = [action]
 
         results = []
         for act in action:
@@ -125,19 +127,22 @@ class BaseFormatDetector:
                         ),
                     )
                 )
+            else:
+                logger.warning(f"Model attempted to call undefined function: {name}")
 
         return results
 
-    def detect_and_parse(self, text: str, tools: List[Function]) -> List[ToolCallItem]:
+    @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 self.parse_base_json(action, tools)
+        return StreamingParseResult(calls=self.parse_base_json(action, tools))
 
     def parse_streaming_increment(
-        self, new_text: str, tools: List[Function]
+        self, new_text: str, tools: List[Tool]
     ) -> StreamingParseResult:
         """
         Streaming incremental parsing with tool validation.
@@ -196,7 +201,7 @@ class BaseFormatDetector:
                         obj["arguments"] = obj["parameters"]
                     tool_call_arr.append(obj)
 
-            except partial_json_parser.core.exceptions.MalformedJSON:
+            except MalformedJSON:
                 return StreamingParseResult()
 
             if len(tool_call_arr) == 0:
@@ -302,6 +307,14 @@ class BaseFormatDetector:
             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()
+
 
 class Qwen25Detector(BaseFormatDetector):
     """
@@ -322,7 +335,7 @@ class Qwen25Detector(BaseFormatDetector):
         """Check if the text contains a Qwen 2.5 format tool call."""
         return self.bot_token in text
 
-    def detect_and_parse(self, text: str, tools: List[Function]) -> List[ToolCallItem]:
+    def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
         """
         One-time parsing: Detects and parses tool calls in the provided text.
 
@@ -330,15 +343,24 @@ class Qwen25Detector(BaseFormatDetector):
         :param tools: List of available tools.
         :return: ParseResult indicating success or failure, consumed text, leftover text, and parsed calls.
         """
-        if "<tool_call>" not in text:
-            return []
-        pattern = r"<tool_call>(.*?)</tool_call>"
+        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=[])
+        pattern = rf"{self.bot_token}(.*?){self.eot_token}"
         match_result_list = re.findall(pattern, text, re.DOTALL)
         calls = []
         for match_result in match_result_list:
             match_result = json.loads(match_result)
             calls.extend(self.parse_base_json(match_result, tools))
-        return calls
+        return StreamingParseResult(normal_text=normal_text, calls=calls)
+
+    def structure_info(self) -> _GetInfoFunc:
+        return lambda name: StructureInfo(
+            begin='<tool_call>{"name":"' + name + '", "arguments":',
+            end="}</tool_call>",
+            trigger="<tool_call>",
+        )
 
 
 class MistralDetector(BaseFormatDetector):
@@ -374,7 +396,7 @@ class MistralDetector(BaseFormatDetector):
         else:
             return ""
 
-    def detect_and_parse(self, text: str, tools: List[Function]) -> List[ToolCallItem]:
+    def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
         """
         One-time parsing: Detects and parses tool calls in the provided text.
 
@@ -382,6 +404,8 @@ class MistralDetector(BaseFormatDetector):
         :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
         text = self._clean_text(text)
         tool_content = text.replace("[TOOL_CALLS]", "").strip()
         raw_tool_calls = self.tool_call_regex.findall(tool_content)
@@ -391,7 +415,14 @@ class MistralDetector(BaseFormatDetector):
             function_call_arr = json.loads(raw_tool_call)
             for match_result in function_call_arr:
                 calls.extend(self.parse_base_json(match_result, tools))
-        return calls
+        return StreamingParseResult(normal_text=normal_text, calls=calls)
+
+    def structure_info(self) -> _GetInfoFunc:
+        return lambda name: StructureInfo(
+            begin='[TOOL_CALLS] [{"name":"' + name + '", "arguments":',
+            end="}]",
+            trigger="[TOOL_CALLS]",
+        )
 
 
 class Llama32Detector(BaseFormatDetector):
@@ -411,19 +442,18 @@ class Llama32Detector(BaseFormatDetector):
         # prefix the output with the <|python_tag|> token
         return "<|python_tag|>" in text or text.startswith("{")
 
-    def detect_and_parse(self, text: str, tools: List[Function]) -> List[ToolCallItem]:
+    def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
         """Parse function calls from text, handling multiple JSON objects."""
         if "<|python_tag|>" not in text and not text.startswith("{"):
-            return []
+            return StreamingParseResult(normal_text=text, calls=[])
 
         if "<|python_tag|>" in text:
-            _, action_text = text.split("<|python_tag|>")
+            normal_text, action_text = text.split("<|python_tag|>")
         else:
-            action_text = text
+            normal_text, action_text = "", text
 
         # Split by semicolon and process each part
         json_parts = [part.strip() for part in action_text.split(";") if part.strip()]
-
         all_actions = []
         for part in json_parts:
             try:
@@ -434,12 +464,18 @@ class Llama32Detector(BaseFormatDetector):
                 logger.warning(f"Failed to parse JSON part: {part}")
                 logger.warning(f"JSON parse error: {str(e)}")
                 continue
-
+        calls = []
         # Only process if we found valid JSON objects
         if all_actions:
-            return self.parse_base_json(all_actions, tools)
-
-        return []
+            calls = self.parse_base_json(all_actions, tools)
+        return StreamingParseResult(normal_text=normal_text, calls=calls)
+
+    def structure_info(self) -> _GetInfoFunc:
+        return lambda name: StructureInfo(
+            begin='<|python_tag|>{"name":"' + name + '", "arguments":',
+            end="}",
+            trigger="<|python_tag|>",
+        )
 
 
 class MultiFormatParser:
@@ -449,7 +485,9 @@ class MultiFormatParser:
         """
         self.detectors = detectors
 
-    def parse_once(self, text: str, tools: List[Function]):
+    def parse_once(
+        self, text: str, tools: List[Tool]
+    ) -> Tuple[str, list[ToolCallItem]]:
         """
         One-time parsing: Loop through detectors until there are no new matches or text is exhausted
         Return: (final_text, all_calls)
@@ -459,15 +497,19 @@ class MultiFormatParser:
         final_calls = []
         final_normal_text = text
         for detector in self.detectors:
-            tool_call_list = detector.detect_and_parse(text, tools)
+            parsed_result = detector.detect_and_parse(text, tools)
+            tool_call_list = parsed_result.calls
             if len(tool_call_list) > 0:  # parsed successfully
                 final_calls = tool_call_list
+                final_normal_text = parsed_result.normal_text
                 break
 
         # leftover_text is the normal text not consumed by any Detector
         return final_normal_text, final_calls
 
-    def parse_streaming_increment(self, new_text: str, tools: List[Function]):
+    def parse_streaming_increment(
+        self, new_text: str, tools: List[Tool]
+    ) -> Tuple[str, list[ToolCallItem]]:
         """
         Streaming incremental parsing: Feed new_text to each detector's parse_streaming_increment
         and merge their produced normal_text/calls to return.
@@ -498,13 +540,13 @@ class FunctionCallParser:
     and returns the resulting normal_text and calls to the upper layer (or SSE).
     """
 
-    ToolCallParserEnum: Dict[str, BaseFormatDetector] = {
+    ToolCallParserEnum: Dict[str, Type[BaseFormatDetector]] = {
         "llama3": Llama32Detector,
         "qwen25": Qwen25Detector,
         "mistral": MistralDetector,
     }
 
-    def __init__(self, tools: List[Function], tool_call_parser: str = None):
+    def __init__(self, tools: List[Tool], tool_call_parser: str):
         detectors = []
         if tool_call_parser:
             detector_class = self.ToolCallParserEnum.get(tool_call_parser)
@@ -532,7 +574,7 @@ class FunctionCallParser:
                 return True
         return False
 
-    def parse_non_stream(self, full_text: str):
+    def parse_non_stream(self, full_text: str) -> Tuple[str, list[ToolCallItem]]:
         """
         Non-streaming call: one-time parsing
         """
@@ -541,7 +583,7 @@ class FunctionCallParser:
         )
         return full_normal_text, calls
 
-    def parse_stream_chunk(self, chunk_text: str):
+    def parse_stream_chunk(self, chunk_text: str) -> Tuple[str, list[ToolCallItem]]:
         """
         Streaming call: incremental parsing
         """
@@ -549,3 +591,40 @@ class FunctionCallParser:
             chunk_text, self.tools
         )
         return normal_text, calls
+
+    def structure_infos(self) -> List[_GetInfoFunc]:
+        """
+        Returns a list of structure_info functions for each detector
+        """
+        return [
+            detector.structure_info() for detector in self.multi_format_parser.detectors
+        ]
+
+    def get_structure_tag(self) -> StructuralTagResponseFormat:
+        tool_structures: List[StructuresResponseFormat] = list()
+        tool_trigger_set: Set[str] = set()
+
+        for wrapper in self.structure_infos():
+            for tool in self.tools:
+                function = tool.function
+                name = function.name
+                assert name is not None
+                info = wrapper(name)
+
+                # accept all if not strict, otherwise only accept the schema
+                schema = function.parameters if function.strict else {}
+
+                tool_structures.append(
+                    StructuresResponseFormat(
+                        begin=info.begin,
+                        schema=schema,  # type: ignore
+                        end=info.end,
+                    )
+                )
+                tool_trigger_set.add(info.trigger)
+
+        return StructuralTagResponseFormat(
+            type="structural_tag",
+            structures=tool_structures,
+            triggers=list(tool_trigger_set),
+        )