sglang 0.4.10.post2__py3-none-any.whl → 0.5.0rc1__py3-none-any.whl

sglang/srt/entrypoints/openai/protocol.py

@@ -14,9 +14,18 @@
 """Pydantic models for OpenAI API protocol"""
 
 import time
+import uuid
 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Dict, List, Optional, TypeAlias, Union
 
+from openai.types.responses import (
+    ResponseFunctionToolCall,
+    ResponseInputItemParam,
+    ResponseOutputItem,
+    ResponseReasoningItem,
+)
+from openai.types.responses.response import ToolChoice
+from openai.types.responses.tool import Tool
 from pydantic import (
     BaseModel,
     Field,
@@ -84,6 +93,7 @@ class UsageInfo(BaseModel):
     completion_tokens: Optional[int] = 0
     # only used to return cached tokens when --enable-cache-report is set
     prompt_tokens_details: Optional[Dict[str, int]] = None
+    reasoning_tokens: Optional[int] = 0
 
 
 class StreamOptions(BaseModel):
@@ -428,6 +438,13 @@ class ChatCompletionRequest(BaseModel):
         default="auto", examples=["none"]
     )  # noqa
     return_hidden_states: bool = False
+    reasoning_effort: Optional[Literal["low", "medium", "high"]] = Field(
+        default="medium",
+        description="Constrains effort on reasoning for reasoning models. "
+        "'low' is the least effort, 'high' is the most effort. Reducing reasoning effort can "
+        "result in faster responses and fewer tokens used on reasoning in a response. "
+        "Currently only supported for OpenAI models.",
+    )
 
     @model_validator(mode="before")
     @classmethod
@@ -619,6 +636,196 @@ OpenAIServingRequest = Union[
 ]
 
 
+# Response API protocol definitions
+class ResponseReasoningParam(BaseModel):
+    """Reasoning parameters for responses."""
+
+    effort: Optional[Literal["low", "medium", "high"]] = Field(
+        default="medium",
+        description="Constrains effort on reasoning for reasoning models.",
+    )
+
+
+class ResponseTool(BaseModel):
+    """Tool definition for responses."""
+
+    type: Literal["web_search_preview", "code_interpreter"] = Field(
+        description="Type of tool to enable"
+    )
+
+
+ResponseInputOutputItem: TypeAlias = Union[
+    ResponseInputItemParam,
+    "ResponseReasoningItem",
+    ResponseFunctionToolCall,
+]
+
+
+class ResponsesRequest(BaseModel):
+    """Request body for v1/responses endpoint."""
+
+    # Core OpenAI API fields (ordered by official documentation)
+    background: Optional[bool] = False
+    include: Optional[
+        List[
+            Literal[
+                "code_interpreter_call.outputs",
+                "computer_call_output.output.image_url",
+                "file_search_call.results",
+                "message.input_image.image_url",
+                "message.output_text.logprobs",
+                "reasoning.encrypted_content",
+            ]
+        ]
+    ] = None
+    input: Union[str, List[ResponseInputOutputItem]]
+    instructions: Optional[str] = None
+    max_output_tokens: Optional[int] = None
+    max_tool_calls: Optional[int] = None
+    metadata: Optional[Dict[str, Any]] = None
+    model: Optional[str] = None  # Made optional to match vLLM
+    parallel_tool_calls: Optional[bool] = True
+    previous_response_id: Optional[str] = None
+    reasoning: Optional[ResponseReasoningParam] = None
+    service_tier: Literal["auto", "default", "flex", "scale", "priority"] = "auto"
+    store: Optional[bool] = True
+    stream: Optional[bool] = False
+    temperature: Optional[float] = None
+    tool_choice: Literal["auto", "required", "none"] = "auto"
+    tools: List[ResponseTool] = Field(default_factory=list)
+    top_logprobs: Optional[int] = 0
+    top_p: Optional[float] = None
+    truncation: Optional[Literal["auto", "disabled"]] = "disabled"
+    user: Optional[str] = None
+
+    # Extra SGLang parameters
+    request_id: str = Field(
+        default_factory=lambda: f"resp_{uuid.uuid4().hex}",
+        description="The request_id related to this request. If the caller does not set it, a random uuid will be generated.",
+    )
+    priority: int = Field(default=0, description="Request priority")
+
+    # SGLang-specific sampling parameters
+    frequency_penalty: float = 0.0
+    presence_penalty: float = 0.0
+    stop: Optional[Union[str, List[str]]] = None
+    top_k: int = -1
+    min_p: float = 0.0
+    repetition_penalty: float = 1.0
+
+    # Default sampling parameters
+    _DEFAULT_SAMPLING_PARAMS = {
+        "temperature": 0.7,
+        "top_p": 1.0,
+        "top_k": -1,
+        "min_p": 0.0,
+        "repetition_penalty": 1.0,
+    }
+
+    def to_sampling_params(
+        self, default_max_tokens: int, default_params: Optional[Dict] = None
+    ) -> Dict[str, Any]:
+        """Convert to sampling parameters for generation."""
+        if default_params is None:
+            default_params = {}
+
+        # Use max_output_tokens if available, otherwise use max_tokens for backwards compatibility
+        if self.max_output_tokens is not None:
+            max_tokens = min(self.max_output_tokens, default_max_tokens)
+        else:
+            max_tokens = default_max_tokens
+
+        # Avoid exceed the context length by minus 1 token
+        max_tokens -= 1
+
+        # Get parameters with defaults
+        temperature = self.temperature
+        if temperature is None:
+            temperature = default_params.get(
+                "temperature", self._DEFAULT_SAMPLING_PARAMS["temperature"]
+            )
+
+        top_p = self.top_p
+        if top_p is None:
+            top_p = default_params.get("top_p", self._DEFAULT_SAMPLING_PARAMS["top_p"])
+
+        params = {
+            "max_new_tokens": max_tokens,
+            "temperature": temperature,
+            "top_p": top_p,
+            "frequency_penalty": self.frequency_penalty,
+            "presence_penalty": self.presence_penalty,
+            "stop": self.stop,
+            "top_k": self.top_k,
+            "min_p": self.min_p,
+            "repetition_penalty": self.repetition_penalty,
+        }
+
+        # Apply any additional default parameters
+        for key, value in default_params.items():
+            if key not in params or params[key] is None:
+                params[key] = value
+
+        return params
+
+
+class PromptTokenUsageInfo(BaseModel):
+    """Prompt token usage details."""
+
+    cached_tokens: int = 0
+
+
+class ResponsesResponse(BaseModel):
+    """Response body for v1/responses endpoint."""
+
+    id: str = Field(default_factory=lambda: f"resp_{time.time()}")
+    object: Literal["response"] = "response"
+    created_at: int = Field(default_factory=lambda: int(time.time()))
+    model: str
+
+    output: List[
+        Union[ResponseOutputItem, ResponseReasoningItem, ResponseFunctionToolCall]
+    ] = Field(default_factory=list)
+    status: Literal["queued", "in_progress", "completed", "failed", "cancelled"]
+    usage: Optional[UsageInfo] = None
+    parallel_tool_calls: bool = True
+    tool_choice: str = "auto"
+    tools: List[ResponseTool] = Field(default_factory=list)
+
+    @classmethod
+    def from_request(
+        cls,
+        request: ResponsesRequest,
+        sampling_params: Any,
+        model_name: str,
+        created_time: int,
+        output: List[
+            Union[ResponseOutputItem, ResponseReasoningItem, ResponseFunctionToolCall]
+        ],
+        status: str,
+        usage: Optional[UsageInfo],
+    ) -> "ResponsesResponse":
+        """Create a response from a request."""
+        return cls(
+            id=request.request_id,
+            created_at=created_time,
+            model=model_name,
+            output=output,
+            status=status,
+            usage=usage,
+            parallel_tool_calls=request.parallel_tool_calls or True,
+            tool_choice=request.tool_choice,
+            tools=request.tools,
+        )
+
+
+class RequestResponseMetadata(BaseModel):
+    """Metadata for request/response tracking."""
+
+    request_id: str
+    final_usage_info: Optional[UsageInfo] = None
+
+
 @dataclass
 class MessageProcessingResult:
     """Result of processing chat messages and applying templates.
@@ -645,3 +852,13 @@ class MessageProcessingResult:
     modalities: List[str]
     stop: List[str]
     tool_call_constraint: Optional[Any] = None
+
+
+class ResponseReasoningTextContent(BaseModel):
+    text: str
+    type: Literal["reasoning_text"] = "reasoning_text"
+
+
+ResponseInputOutputItem: TypeAlias = Union[
+    ResponseInputItemParam, "ResponseReasoningItem", ResponseFunctionToolCall
+]

sglang/srt/entrypoints/openai/serving_chat.py

@@ -47,7 +47,9 @@ class OpenAIServingChat(OpenAIServingBase):
     """Handler for /v1/chat/completions requests"""
 
     def __init__(
-        self, tokenizer_manager: TokenizerManager, template_manager: TemplateManager
+        self,
+        tokenizer_manager: TokenizerManager,
+        template_manager: TemplateManager,
     ):
         super().__init__(tokenizer_manager)
         self.template_manager = template_manager
@@ -67,6 +69,18 @@ class OpenAIServingChat(OpenAIServingBase):
         ):
             return "Tools cannot be empty if tool choice is set to required."
 
+        max_output_tokens = request.max_completion_tokens or request.max_tokens
+        server_context_length = self.tokenizer_manager.server_args.context_length
+        if (
+            max_output_tokens
+            and server_context_length
+            and max_output_tokens > server_context_length
+        ):
+            return (
+                f"max_completion_tokens is too large: {max_output_tokens}."
+                f"This model supports at most {server_context_length} completion tokens."
+            )
+
         return None
 
     def _convert_to_internal_request(
@@ -81,7 +95,9 @@ class OpenAIServingChat(OpenAIServingBase):
 
         # Build sampling parameters
         sampling_params = self._build_sampling_params(
-            request, processed_messages.stop, processed_messages.tool_call_constraint
+            request,
+            processed_messages.stop,
+            processed_messages.tool_call_constraint,
         )
 
         # Handle single vs multiple requests
@@ -196,14 +212,15 @@ class OpenAIServingChat(OpenAIServingBase):
                 tokenize=True,
                 add_generation_prompt=True,
                 tools=tools,
+                reasoning_effort=request.reasoning_effort,
                 **(
                     request.chat_template_kwargs if request.chat_template_kwargs else {}
                 ),
             )
         except Exception:
-            #  This except branch will be triggered when the chosen model
-            #  has a different tools input format that is not compatible
-            #  with openAI's apply_chat_template tool_call format, like Mistral.
+            # This except branch will be triggered when the chosen model
+            # has a different tools input format that is not compatible
+            # with openAI's apply_chat_template tool_call format, like Mistral.
             tools = (
                 [t if "function" in t else {"function": t} for t in tools]
                 if tools
@@ -214,6 +231,7 @@ class OpenAIServingChat(OpenAIServingBase):
                 tokenize=True,
                 add_generation_prompt=True,
                 tools=tools,
+                reasoning_effort=request.reasoning_effort,
                 **(
                     request.chat_template_kwargs if request.chat_template_kwargs else {}
                 ),
@@ -277,6 +295,8 @@ class OpenAIServingChat(OpenAIServingBase):
                 prompt = prompt[: -len(conv.sep2)]
         else:
             prompt = conv.get_prompt()
+            if self._get_enable_thinking_from_request(request):
+                prompt += "<think>"  # Note(Xinyuan): hard code thinking token
 
         image_data = conv.image_data if conv.image_data else None
         video_data = conv.video_data if conv.video_data else None
@@ -448,7 +468,6 @@ class OpenAIServingChat(OpenAIServingBase):
                     )
                     yield f"data: {chunk.model_dump_json()}\n\n"
 
-                # Process content delta
                 stream_buffer = stream_buffers.get(index, "")
                 delta = content["text"][len(stream_buffer) :]
                 stream_buffers[index] = stream_buffer + delta
@@ -502,7 +521,7 @@ class OpenAIServingChat(OpenAIServingBase):
                     if delta:
                         choice_data = ChatCompletionResponseStreamChoice(
                             index=index,
-                            delta=DeltaMessage(content=delta if delta else None),
+                            delta=DeltaMessage(content=delta),
                             finish_reason=None,
                             matched_stop=None,
                             logprobs=choice_logprobs,
@@ -645,9 +664,15 @@ class OpenAIServingChat(OpenAIServingBase):
             reasoning_text = None
             reasoning_parser = self.tokenizer_manager.server_args.reasoning_parser
             if reasoning_parser and request.separate_reasoning:
+                is_force_reasoning = (
+                    self.template_manager.force_reasoning
+                    or self._get_enable_thinking_from_request(request)
+                )
                 try:
                     parser = ReasoningParser(
-                        model_type=reasoning_parser, stream_reasoning=False
+                        model_type=reasoning_parser,
+                        stream_reasoning=False,
+                        force_reasoning=is_force_reasoning,
                     )
                     reasoning_text, text = parser.parse_non_stream(text)
                 except Exception as e:
@@ -810,14 +835,19 @@ class OpenAIServingChat(OpenAIServingBase):
     ) -> tuple[Optional[str], str]:
         """Process reasoning content in streaming response"""
         if index not in reasoning_parser_dict:
+            is_force_reasoning = (
+                self.template_manager.force_reasoning
+                or self._get_enable_thinking_from_request(request)
+            )
             reasoning_parser_dict[index] = ReasoningParser(
                 self.tokenizer_manager.server_args.reasoning_parser,
                 request.stream_reasoning,
+                is_force_reasoning,
             )
         reasoning_parser = reasoning_parser_dict[index]
         return reasoning_parser.parse_stream_chunk(delta)
 
-    def _get_enable_thinking_from_request(request: ChatCompletionRequest) -> bool:
+    def _get_enable_thinking_from_request(self, request: ChatCompletionRequest) -> bool:
         """Extracts the 'enable_thinking' flag from request chat_template_kwargs.
 
         NOTE: This parameter is only useful for models that support enable_thinking
@@ -826,7 +856,7 @@ class OpenAIServingChat(OpenAIServingBase):
         Args:
             request_obj: The request object (or an item from a list of requests).
         Returns:
-            The boolean value of 'enable_thinking' if found and not True, otherwise True.
+            The boolean value of 'enable_thinking' if found, otherwise False.
         """
         if (
             hasattr(request, "chat_template_kwargs")
@@ -834,7 +864,7 @@ class OpenAIServingChat(OpenAIServingBase):
             and request.chat_template_kwargs.get("enable_thinking") is not None
         ):
             return request.chat_template_kwargs.get("enable_thinking")
-        return True
+        return False
 
     async def _process_tool_call_stream(
         self,