sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc2__py3-none-any.whl

sglang/srt/function_call/glm4_moe_detector.py

@@ -39,7 +39,7 @@ def parse_arguments(json_value):
 
 class Glm4MoeDetector(BaseFormatDetector):
     """
-    Detector for GLM-4.5 models.
+    Detector for GLM-4.5 and GLM-4.6 models.
     Assumes function call format:
       <tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>北京</arg_value>\n<arg_key>date</arg_key>\n<arg_value>2024-06-27</arg_value>\n</tool_call>\n<tool_call>get_weather\n<arg_key>city</arg_key>\n<arg_value>上海</arg_value>\n<arg_key>date</arg_key>\n<arg_value>2024-06-27</arg_value>\n</tool_call>
     """
@@ -53,7 +53,7 @@ class Glm4MoeDetector(BaseFormatDetector):
         self.func_arg_regex = r"<arg_key>(.*?)</arg_key>\s*<arg_value>(.*?)</arg_value>"
 
     def has_tool_call(self, text: str) -> bool:
-        """Check if the text contains a glm-4.5 format tool call."""
+        """Check if the text contains a glm-4.5 / glm-4.6 format tool call."""
         return self.bot_token in text
 
     def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
@@ -102,7 +102,7 @@ class Glm4MoeDetector(BaseFormatDetector):
         self, new_text: str, tools: List[Tool]
     ) -> StreamingParseResult:
         """
-        Streaming incremental parsing tool calls for GLM-4.5 format.
+        Streaming incremental parsing tool calls for GLM-4.5 and GLM-4.6 format.
         """
         self._buffer += new_text
         current_text = self._buffer
@@ -160,5 +160,5 @@ class Glm4MoeDetector(BaseFormatDetector):
             function_format="xml",
             call_rule_fmt='"{name}" "\\n" ( {arguments_rule} "\\n" )?',
             key_value_rule_fmt='"<arg_key>{key}</arg_key>" "\\n" "<arg_value>" {valrule} "</arg_value>"',
-            key_value_separator="\\n",
+            key_value_separator='"\\n"',
         )

sglang/srt/function_call/gpt_oss_detector.py

@@ -81,6 +81,29 @@ class GptOssDetector(BaseFormatDetector):
         # Always use HarmonyParser for parsing to ensure proper filtering
         events = self.harmony_parser.parse(new_text)
 
+        # If there are no parsed events and the chunk contains no Harmony structural
+        # markers, treat it as plain text and pass it through. This fixes a bug where
+        # normal content was held in the buffer when tools were provided but not used.
+        if not events:
+            has_harmony_markers = any(
+                marker in self._buffer
+                for marker in (
+                    "<|start|>",
+                    "<|channel|>",
+                    "<|message|>",
+                    "<|constrain|>",
+                    "<|end|>",
+                    "<|call|>",
+                    "<|return|>",
+                    "assistantfinal",
+                )
+            )
+            if not has_harmony_markers:
+                # Plain text with no tool markers — emit as normal content
+                out = self._buffer
+                self._buffer = ""
+                return StreamingParseResult(normal_text=out, calls=[])
+
         # Quick check if we might have tool calls
         if (
             "<|channel|>commentary to=" not in self._buffer

sglang/srt/function_call/json_array_parser.py

@@ -0,0 +1,63 @@
+import json
+import re
+from typing import List
+
+from sglang.srt.entrypoints.openai.protocol import Tool
+from sglang.srt.function_call.base_format_detector import BaseFormatDetector
+from sglang.srt.function_call.core_types import StreamingParseResult
+
+
+class JsonArrayParser(BaseFormatDetector):
+    """
+    Parser for JSON array tool calls when JSON schema constraints are active.
+
+    This parser is used when tool_choice="required" or a specific tool is named,
+    bypassing model-specific parsers in favor of direct JSON array parsing.
+    """
+
+    def __init__(self):
+        super().__init__()
+        # Configure for JSON array parsing
+        self.bot_token = "["
+        self.eot_token = "]"
+        self.tool_call_separator = ","
+
+    def has_tool_call(self, text: str) -> bool:
+        """
+        Check if the given text contains a JSON tool call (array or single object).
+        """
+        return "[" in text or "{" in text
+
+    def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
+        """
+        Parse JSON tool calls using the base class implementation.
+        """
+        raise NotImplementedError(
+            "Detect and parse not supported for JSON schema constraints."
+        )
+
+    def build_ebnf(self, tools: List[Tool]) -> str:
+        """
+        Build an EBNF grammar for constrained generation.
+        This is not used for JSON schema constraints as they are handled
+        by the constraint backends directly.
+        """
+        raise NotImplementedError(
+            "EBNF generation is not supported for JSON schema constraints."
+        )
+
+    def parse_streaming_increment(
+        self, new_text: str, tools: List[Tool]
+    ) -> StreamingParseResult:
+        """
+        Streaming incremental parsing with tool validation.
+        """
+        return super().parse_streaming_increment(new_text, tools)
+
+    def structure_info(self) -> callable:
+        """
+        Return a function that creates StructureInfo for constrained generation.
+        This is not used for JSON schema constraints as they are handled
+        by the constraint backends directly.
+        """
+        raise NotImplementedError("structure_info not used for JSON schema constraints")

sglang/srt/function_call/kimik2_detector.py

@@ -50,6 +50,11 @@ class KimiK2Detector(BaseFormatDetector):
 
         self._last_arguments = ""
 
+        # Robust parser for ids like "functions.search:0" or fallback "search:0"
+        self.tool_call_id_regex = re.compile(
+            r"^(?:functions\.)?(?P<name>[\w\.]+):(?P<index>\d+)$"
+        )
+
     def has_tool_call(self, text: str) -> bool:
         """Check if the text contains a KimiK2 format tool call."""
         return self.bot_token in text
@@ -76,14 +81,18 @@ class KimiK2Detector(BaseFormatDetector):
             tool_calls = []
             for match in function_call_tuples:
                 function_id, function_args = match
-                function_name = function_id.split(".")[1].split(":")[0]
-                function_idx = int(function_id.split(".")[1].split(":")[1])
+                m = self.tool_call_id_regex.match(function_id)
+                if not m:
+                    logger.warning("Unexpected tool_call_id format: %s", function_id)
+                    continue
+                function_name = m.group("name")
+                function_idx = int(m.group("index"))
 
                 logger.info(f"function_name {function_name}")
 
                 tool_calls.append(
                     ToolCallItem(
-                        tool_index=function_idx,  # Use the call index in the response, not tool position
+                        tool_index=function_idx,
                         name=function_name,
                         parameters=function_args,
                     )
@@ -128,7 +137,11 @@ class KimiK2Detector(BaseFormatDetector):
                 function_id = match.group("tool_call_id")
                 function_args = match.group("function_arguments")
 
-                function_name = function_id.split(".")[1].split(":")[0]
+                m = self.tool_call_id_regex.match(function_id)
+                if not m:
+                    logger.warning("Unexpected tool_call_id format: %s", function_id)
+                    return StreamingParseResult(normal_text="", calls=calls)
+                function_name = m.group("name")
 
                 # Initialize state if this is the first tool call
                 if self.current_tool_id == -1:

sglang/srt/function_call/qwen3_coder_detector.py

@@ -358,5 +358,5 @@ class Qwen3CoderDetector(BaseFormatDetector):
             function_format="xml",
             call_rule_fmt='"<function={name}>\\n" {arguments_rule} "\\n</function>"',
             key_value_rule_fmt='"<parameter={key}>\\n" {valrule} "\\n</parameter>"',
-            key_value_separator="\\n",
+            key_value_separator='"\\n"',
         )

sglang/srt/function_call/utils.py

@@ -1,10 +1,13 @@
 import json
 from json import JSONDecodeError, JSONDecoder
-from typing import Any, Tuple
+from json.decoder import WHITESPACE
+from typing import Any, List, Literal, Optional, Tuple, Union
 
 import partial_json_parser
 from partial_json_parser.core.options import Allow
 
+from sglang.srt.entrypoints.openai.protocol import Tool, ToolChoice
+
 
 def _find_common_prefix(s1: str, s2: str) -> str:
     prefix = ""
@@ -37,10 +40,12 @@ def _partial_json_loads(input_str: str, flags: Allow) -> Tuple[Any, int]:
     """
     try:
         return (partial_json_parser.loads(input_str, flags), len(input_str))
-    except JSONDecodeError as e:
-        if "Extra data" in e.msg:
-            dec = JSONDecoder()
-            return dec.raw_decode(input_str)
+    except (JSONDecodeError, IndexError) as e:
+        msg = getattr(e, "msg", str(e))
+        if "Extra data" in msg or "pop from empty list" in msg:
+            start = WHITESPACE.match(input_str, 0).end()
+            obj, end = JSONDecoder().raw_decode(input_str, start)
+            return obj, end
         raise
 
 
@@ -50,3 +55,89 @@ def _is_complete_json(input_str: str) -> bool:
         return True
     except JSONDecodeError:
         return False
+
+
+def _get_tool_schema_defs(tools: List[Tool]) -> dict:
+    """
+    Get consolidated $defs from all tools, validating for conflicts.
+
+    Args:
+        tools: List of tools to process
+
+    Returns:
+        Dictionary of consolidated $defs from all tools
+
+    Raises:
+        ValueError: If conflicting $defs are found
+    """
+    all_defs = {}
+    for tool in tools:
+        if tool.function.parameters is None:
+            continue
+        defs = tool.function.parameters.get("$defs", {})
+        for def_name, def_schema in defs.items():
+            if def_name in all_defs and all_defs[def_name] != def_schema:
+                raise ValueError(
+                    f"Tool definition '{def_name}' has "
+                    "multiple schemas, which is not "
+                    "supported."
+                )
+            else:
+                all_defs[def_name] = def_schema
+    return all_defs
+
+
+def _get_tool_schema(tool: Tool) -> dict:
+    return {
+        "properties": {
+            "name": {"type": "string", "enum": [tool.function.name]},
+            "parameters": (
+                tool.function.parameters
+                if tool.function.parameters
+                else {"type": "object", "properties": {}}
+            ),
+        },
+        "required": ["name", "parameters"],
+    }
+
+
+def get_json_schema_constraint(
+    tools: List[Tool], tool_choice: Union[ToolChoice, Literal["required"]]
+) -> Optional[dict]:
+    """
+    Get the JSON schema constraint for the specified tool choice.
+
+    Args:
+        tool_choice: The tool choice specification
+
+    Returns:
+        JSON schema dict, or None if no valid tools found
+    """
+
+    if isinstance(tool_choice, ToolChoice):
+        # For specific function choice, return the user's parameters schema directly
+        fn_name = tool_choice.function.name
+        for tool in tools:
+            if tool.function.name == fn_name:
+                return {
+                    "type": "array",
+                    "minItems": 1,
+                    "maxItems": 1,
+                    "items": _get_tool_schema(tool),
+                }
+        return None
+    elif tool_choice == "required":
+        json_schema = {
+            "type": "array",
+            "minItems": 1,
+            "items": {
+                "type": "object",
+                "anyOf": [_get_tool_schema(tool) for tool in tools],
+            },
+        }
+        json_schema_defs = _get_tool_schema_defs(tools)
+        if json_schema_defs:
+            json_schema["$defs"] = json_schema_defs
+        return json_schema
+
+    return None

sglang/srt/grpc/__init__.py

@@ -0,0 +1 @@
+# SGLang gRPC module

sglang/srt/grpc/compile_proto.py

@@ -0,0 +1,245 @@
+#!/usr/bin/env python3
+"""
+Compile protobuf files for SGLang gRPC server.
+
+This script compiles .proto files to Python code using grpc_tools.protoc.
+It generates:
+- *_pb2.py (protobuf message classes)
+- *_pb2_grpc.py (gRPC service classes)
+- *_pb2.pyi (type hints for mypy/IDEs)
+
+Usage:
+    python compile_proto.py [--check] [--proto-file PROTO_FILE]
+
+Options:
+    --check         Check if regeneration is needed (exit 1 if needed)
+    --proto-file    Specify proto file (default: sglang_scheduler.proto)
+
+### Install Dependencies
+pip install "grpcio==1.74.0" "grpcio-tools==1.74.0"
+
+### Run Script
+cd python/sglang/srt/grpc
+python compile_proto.py
+"""
+
+
+import argparse
+import subprocess
+import sys
+from importlib.metadata import version
+from pathlib import Path
+
+GRPC_VERSION = "1.74.0"
+
+
+def get_file_mtime(path: Path) -> float:
+    """Get file modification time, return 0 if file doesn't exist."""
+    try:
+        return path.stat().st_mtime
+    except FileNotFoundError:
+        return 0.0
+
+
+def check_regeneration_needed(proto_file: Path, output_dir: Path) -> bool:
+    """Check if proto files are newer than generated files."""
+    proto_mtime = get_file_mtime(proto_file)
+
+    generated_files = [
+        output_dir / f"{proto_file.stem}_pb2.py",
+        output_dir / f"{proto_file.stem}_pb2_grpc.py",
+        output_dir / f"{proto_file.stem}_pb2.pyi",
+    ]
+
+    for gen_file in generated_files:
+        if get_file_mtime(gen_file) < proto_mtime:
+            return True
+
+    return False
+
+
+def compile_proto(proto_file: Path, output_dir: Path, verbose: bool = True) -> bool:
+    """Compile the protobuf file to Python."""
+
+    if not proto_file.exists():
+        print(f"Error: Proto file not found: {proto_file}")
+        return False
+
+    if verbose:
+        print(f"Found proto file: {proto_file}")
+
+    # Check if grpc_tools is available
+    try:
+        import grpc_tools.protoc
+    except ImportError:
+        print("Error: grpcio-tools not installed")
+        print(
+            f'Install with: pip install "grpcio-tools=={GRPC_VERSION}" "grpcio=={GRPC_VERSION}"'
+        )
+        return False
+
+    grpc_tools_version = version("grpcio-tools")
+    grpc_version = version("grpcio")
+    if grpc_tools_version != GRPC_VERSION or grpc_version != GRPC_VERSION:
+        raise RuntimeError(
+            f"Error: grpcio-tools version {grpc_tools_version} and grpcio version {grpc_version} detected, but {GRPC_VERSION} is required."
+        )
+
+    # Compile command
+    cmd = [
+        sys.executable,
+        "-m",
+        "grpc_tools.protoc",
+        f"-I{proto_file.parent}",
+        f"--python_out={output_dir}",
+        f"--grpc_python_out={output_dir}",
+        f"--pyi_out={output_dir}",  # Generate type stubs
+        str(proto_file.name),
+    ]
+
+    if verbose:
+        print(f"Running: {' '.join(cmd)}")
+
+    # Run protoc
+    result = subprocess.run(cmd, capture_output=True, text=True, cwd=proto_file.parent)
+
+    if result.returncode != 0:
+        print(f"Error compiling proto:")
+        print(result.stderr)
+        if result.stdout:
+            print(result.stdout)
+        return False
+
+    # Verify generated files exist
+    generated_files = [
+        f"{proto_file.stem}_pb2.py",
+        f"{proto_file.stem}_pb2_grpc.py",
+        f"{proto_file.stem}_pb2.pyi",
+    ]
+
+    missing_files = []
+    for gen_file in generated_files:
+        if not (output_dir / gen_file).exists():
+            missing_files.append(gen_file)
+
+    if missing_files:
+        print(f"Error: Expected generated files not found: {missing_files}")
+        return False
+
+    if verbose:
+        print("Successfully compiled protobuf files:")
+        for gen_file in generated_files:
+            print(f"  - {output_dir}/{gen_file}")
+
+    # Fix imports in generated files
+    fix_imports(output_dir, proto_file.stem, verbose)
+
+    return True
+
+
+def fix_imports(output_dir: Path, proto_stem: str, verbose: bool = True) -> None:
+    """Fix imports in generated files to use relative imports."""
+    grpc_file = output_dir / f"{proto_stem}_pb2_grpc.py"
+
+    if grpc_file.exists():
+        content = grpc_file.read_text()
+        # Change absolute import to relative import
+        old_import = f"import {proto_stem}_pb2"
+        new_import = f"from . import {proto_stem}_pb2"
+
+        if old_import in content:
+            content = content.replace(old_import, new_import)
+            grpc_file.write_text(content)
+            if verbose:
+                print("Fixed imports in generated files")
+
+
+def add_generation_header(output_dir: Path, proto_stem: str) -> None:
+    """Add header to generated files indicating they are auto-generated."""
+    header = """# This file is auto-generated. Do not edit manually.
+# Regenerate with: python compile_proto.py
+
+"""
+
+    files_to_update = [f"{proto_stem}_pb2.py", f"{proto_stem}_pb2_grpc.py"]
+
+    for filename in files_to_update:
+        file_path = output_dir / filename
+        if file_path.exists():
+            content = file_path.read_text()
+            if not content.startswith("# This file is auto-generated"):
+                file_path.write_text(header + content)
+
+
+def main():
+    """Main entry point."""
+    parser = argparse.ArgumentParser(
+        description="Compile protobuf files for SGLang gRPC server",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=__doc__,
+    )
+
+    parser.add_argument(
+        "--check",
+        action="store_true",
+        help="Check if regeneration is needed (exit 1 if needed)",
+    )
+
+    parser.add_argument(
+        "--proto-file",
+        type=str,
+        default="sglang_scheduler.proto",
+        help="Proto file to compile (default: sglang_scheduler.proto)",
+    )
+
+    parser.add_argument(
+        "-v",
+        "--verbose",
+        action="store_true",
+        default=True,
+        help="Verbose output (default: True)",
+    )
+
+    parser.add_argument(
+        "-q", "--quiet", action="store_true", help="Quiet mode (overrides verbose)"
+    )
+
+    args = parser.parse_args()
+
+    # Handle verbosity
+    verbose = args.verbose and not args.quiet
+
+    # Get paths
+    script_dir = Path(__file__).parent
+    proto_file = script_dir / args.proto_file
+    output_dir = script_dir
+
+    # Check mode
+    if args.check:
+        if check_regeneration_needed(proto_file, output_dir):
+            if verbose:
+                print("Proto files need regeneration")
+            sys.exit(1)
+        else:
+            if verbose:
+                print("Generated files are up to date")
+            sys.exit(0)
+
+    # Compile mode
+    success = compile_proto(proto_file, output_dir, verbose)
+
+    if success:
+        # Add generation headers
+        add_generation_header(output_dir, proto_file.stem)
+
+        if verbose:
+            print("\n✅ Protobuf compilation successful!")
+            print("Generated files are ready for use")
+    else:
+        if verbose:
+            print("\n❌ Protobuf compilation failed!")
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

sglang/srt/grpc/sglang_scheduler_pb2.py

@@ -0,0 +1,111 @@
+# This file is auto-generated. Do not edit manually.
+# Regenerate with: python compile_proto.py
+
+# -*- coding: utf-8 -*-
+# Generated by the protocol buffer compiler.  DO NOT EDIT!
+# NO CHECKED-IN PROTOBUF GENCODE
+# source: sglang_scheduler.proto
+# Protobuf Python Version: 6.31.1
+"""Generated protocol buffer code."""
+from google.protobuf import descriptor as _descriptor
+from google.protobuf import descriptor_pool as _descriptor_pool
+from google.protobuf import runtime_version as _runtime_version
+from google.protobuf import symbol_database as _symbol_database
+from google.protobuf.internal import builder as _builder
+_runtime_version.ValidateProtobufRuntimeVersion(
+    _runtime_version.Domain.PUBLIC,
+    6,
+    31,
+    1,
+    '',
+    'sglang_scheduler.proto'
+)
+# @@protoc_insertion_point(imports)
+
+_sym_db = _symbol_database.Default()
+
+
+from google.protobuf import timestamp_pb2 as google_dot_protobuf_dot_timestamp__pb2
+from google.protobuf import struct_pb2 as google_dot_protobuf_dot_struct__pb2
+
+
+DESCRIPTOR = _descriptor_pool.Default().AddSerializedFile(b'\n\x16sglang_scheduler.proto\x12\x15sglang.grpc.scheduler\x1a\x1fgoogle/protobuf/timestamp.proto\x1a\x1cgoogle/protobuf/struct.proto\"\xe1\x05\n\x0eSamplingParams\x12\x13\n\x0btemperature\x18\x01 \x01(\x02\x12\r\n\x05top_p\x18\x02 \x01(\x02\x12\r\n\x05top_k\x18\x03 \x01(\x05\x12\r\n\x05min_p\x18\x04 \x01(\x02\x12\x19\n\x11\x66requency_penalty\x18\x05 \x01(\x02\x12\x18\n\x10presence_penalty\x18\x06 \x01(\x02\x12\x1a\n\x12repetition_penalty\x18\x07 \x01(\x02\x12\x1b\n\x0emax_new_tokens\x18\x08 \x01(\x05H\x01\x88\x01\x01\x12\x0c\n\x04stop\x18\t \x03(\t\x12\x16\n\x0estop_token_ids\x18\n \x03(\r\x12\x1b\n\x13skip_special_tokens\x18\x0b \x01(\x08\x12%\n\x1dspaces_between_special_tokens\x18\x0c \x01(\x08\x12\x0f\n\x05regex\x18\r \x01(\tH\x00\x12\x15\n\x0bjson_schema\x18\x0e \x01(\tH\x00\x12\x16\n\x0c\x65\x62nf_grammar\x18\x0f \x01(\tH\x00\x12\x18\n\x0estructural_tag\x18\x10 \x01(\tH\x00\x12\x11\n\tlora_path\x18\x11 \x01(\t\x12\t\n\x01n\x18\x12 \x01(\x05\x12\x15\n\rtoken_healing\x18\x13 \x01(\x08\x12\x16\n\x0emin_new_tokens\x18\x14 \x01(\x05\x12\x12\n\nignore_eos\x18\x15 \x01(\x08\x12\x14\n\x0cno_stop_trim\x18\x16 \x01(\x08\x12\x17\n\x0fstream_interval\x18\x17 \x01(\x05\x12H\n\nlogit_bias\x18\x18 \x03(\x0b\x32\x34.sglang.grpc.scheduler.SamplingParams.LogitBiasEntry\x12.\n\rcustom_params\x18\x19 \x01(\x0b\x32\x17.google.protobuf.Struct\x1a\x30\n\x0eLogitBiasEntry\x12\x0b\n\x03key\x18\x01 \x01(\t\x12\r\n\x05value\x18\x02 \x01(\x02:\x02\x38\x01\x42\x0c\n\nconstraintB\x11\n\x0f_max_new_tokens\"]\n\x13\x44isaggregatedParams\x12\x16\n\x0e\x62ootstrap_host\x18\x01 \x01(\t\x12\x16\n\x0e\x62ootstrap_port\x18\x02 \x01(\x05\x12\x16\n\x0e\x62ootstrap_room\x18\x03 \x01(\x05\"\xe2\x04\n\x0fGenerateRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\ttokenized\x18\x02 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\x12:\n\tmm_inputs\x18\x03 \x01(\x0b\x32\'.sglang.grpc.scheduler.MultimodalInputs\x12>\n\x0fsampling_params\x18\x04 \x01(\x0b\x32%.sglang.grpc.scheduler.SamplingParams\x12\x16\n\x0ereturn_logprob\x18\x05 \x01(\x08\x12\x19\n\x11logprob_start_len\x18\x06 \x01(\x05\x12\x18\n\x10top_logprobs_num\x18\x07 \x01(\x05\x12\x19\n\x11token_ids_logprob\x18\x08 \x03(\r\x12\x1c\n\x14return_hidden_states\x18\t \x01(\x08\x12H\n\x14\x64isaggregated_params\x18\n \x01(\x0b\x32*.sglang.grpc.scheduler.DisaggregatedParams\x12\x1e\n\x16\x63ustom_logit_processor\x18\x0b \x01(\t\x12-\n\ttimestamp\x18\x0c \x01(\x0b\x32\x1a.google.protobuf.Timestamp\x12\x13\n\x0blog_metrics\x18\r \x01(\x08\x12\x14\n\x0cinput_embeds\x18\x0e \x03(\x02\x12\x0f\n\x07lora_id\x18\x0f \x01(\t\x12\x1a\n\x12\x64\x61ta_parallel_rank\x18\x10 \x01(\x05\x12\x0e\n\x06stream\x18\x11 \x01(\x08\":\n\x0eTokenizedInput\x12\x15\n\roriginal_text\x18\x01 \x01(\t\x12\x11\n\tinput_ids\x18\x02 \x03(\r\"\xd3\x01\n\x10MultimodalInputs\x12\x12\n\nimage_urls\x18\x01 \x03(\t\x12\x12\n\nvideo_urls\x18\x02 \x03(\t\x12\x12\n\naudio_urls\x18\x03 \x03(\t\x12\x33\n\x12processed_features\x18\x04 \x01(\x0b\x32\x17.google.protobuf.Struct\x12\x12\n\nimage_data\x18\x05 \x03(\x0c\x12\x12\n\nvideo_data\x18\x06 \x03(\x0c\x12\x12\n\naudio_data\x18\x07 \x03(\x0c\x12\x12\n\nmodalities\x18\x08 \x03(\t\"\xe3\x01\n\x10GenerateResponse\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12;\n\x05\x63hunk\x18\x02 \x01(\x0b\x32*.sglang.grpc.scheduler.GenerateStreamChunkH\x00\x12;\n\x08\x63omplete\x18\x03 \x01(\x0b\x32\'.sglang.grpc.scheduler.GenerateCompleteH\x00\x12\x35\n\x05\x65rror\x18\x04 \x01(\x0b\x32$.sglang.grpc.scheduler.GenerateErrorH\x00\x42\n\n\x08response\"\x95\x02\n\x13GenerateStreamChunk\x12\x11\n\ttoken_ids\x18\x01 \x03(\r\x12\x15\n\rprompt_tokens\x18\x02 \x01(\x05\x12\x19\n\x11\x63ompletion_tokens\x18\x03 \x01(\x05\x12\x15\n\rcached_tokens\x18\x04 \x01(\x05\x12>\n\x0foutput_logprobs\x18\x05 \x01(\x0b\x32%.sglang.grpc.scheduler.OutputLogProbs\x12\x15\n\rhidden_states\x18\x06 \x03(\x02\x12<\n\x0einput_logprobs\x18\x07 \x01(\x0b\x32$.sglang.grpc.scheduler.InputLogProbs\x12\r\n\x05index\x18\x08 \x01(\r\"\x9b\x03\n\x10GenerateComplete\x12\x12\n\noutput_ids\x18\x01 \x03(\r\x12\x15\n\rfinish_reason\x18\x02 \x01(\t\x12\x15\n\rprompt_tokens\x18\x03 \x01(\x05\x12\x19\n\x11\x63ompletion_tokens\x18\x04 \x01(\x05\x12\x15\n\rcached_tokens\x18\x05 \x01(\x05\x12>\n\x0foutput_logprobs\x18\x06 \x01(\x0b\x32%.sglang.grpc.scheduler.OutputLogProbs\x12>\n\x11\x61ll_hidden_states\x18\x07 \x03(\x0b\x32#.sglang.grpc.scheduler.HiddenStates\x12\x1a\n\x10matched_token_id\x18\x08 \x01(\rH\x00\x12\x1a\n\x10matched_stop_str\x18\t \x01(\tH\x00\x12<\n\x0einput_logprobs\x18\n \x01(\x0b\x32$.sglang.grpc.scheduler.InputLogProbs\x12\r\n\x05index\x18\x0b \x01(\rB\x0e\n\x0cmatched_stop\"K\n\rGenerateError\x12\x0f\n\x07message\x18\x01 \x01(\t\x12\x18\n\x10http_status_code\x18\x02 \x01(\t\x12\x0f\n\x07\x64\x65tails\x18\x03 \x01(\t\"u\n\x0eOutputLogProbs\x12\x16\n\x0etoken_logprobs\x18\x01 \x03(\x02\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\x12\x38\n\x0ctop_logprobs\x18\x03 \x03(\x0b\x32\".sglang.grpc.scheduler.TopLogProbs\"\x9e\x01\n\rInputLogProbs\x12@\n\x0etoken_logprobs\x18\x01 \x03(\x0b\x32(.sglang.grpc.scheduler.InputTokenLogProb\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\x12\x38\n\x0ctop_logprobs\x18\x03 \x03(\x0b\x32\".sglang.grpc.scheduler.TopLogProbs\"1\n\x11InputTokenLogProb\x12\x12\n\x05value\x18\x01 \x01(\x02H\x00\x88\x01\x01\x42\x08\n\x06_value\"0\n\x0bTopLogProbs\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\"?\n\x0cHiddenStates\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\r\n\x05layer\x18\x02 \x01(\x05\x12\x10\n\x08position\x18\x03 \x01(\x05\"\xca\x02\n\x0c\x45mbedRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\ttokenized\x18\x02 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\x12:\n\tmm_inputs\x18\x04 \x01(\x0b\x32\'.sglang.grpc.scheduler.MultimodalInputs\x12>\n\x0fsampling_params\x18\x05 \x01(\x0b\x32%.sglang.grpc.scheduler.SamplingParams\x12\x13\n\x0blog_metrics\x18\x06 \x01(\x08\x12\x16\n\x0etoken_type_ids\x18\x07 \x03(\x05\x12\x1a\n\x12\x64\x61ta_parallel_rank\x18\x08 \x01(\x05\x12\x18\n\x10is_cross_encoder\x18\t \x01(\x08\x12\r\n\x05texts\x18\n \x03(\t\"\x9d\x01\n\rEmbedResponse\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x38\n\x08\x63omplete\x18\x02 \x01(\x0b\x32$.sglang.grpc.scheduler.EmbedCompleteH\x00\x12\x32\n\x05\x65rror\x18\x03 \x01(\x0b\x32!.sglang.grpc.scheduler.EmbedErrorH\x00\x42\n\n\x08response\"\xa3\x01\n\rEmbedComplete\x12\x11\n\tembedding\x18\x01 \x03(\x02\x12\x15\n\rprompt_tokens\x18\x02 \x01(\x05\x12\x15\n\rcached_tokens\x18\x03 \x01(\x05\x12\x15\n\rembedding_dim\x18\x04 \x01(\x05\x12:\n\x10\x62\x61tch_embeddings\x18\x05 \x03(\x0b\x32 .sglang.grpc.scheduler.Embedding\"*\n\tEmbedding\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\r\n\x05index\x18\x02 \x01(\x05\"<\n\nEmbedError\x12\x0f\n\x07message\x18\x01 \x01(\t\x12\x0c\n\x04\x63ode\x18\x02 \x01(\t\x12\x0f\n\x07\x64\x65tails\x18\x03 \x01(\t\"N\n\x12HealthCheckRequest\x12\x38\n\ttokenized\x18\x01 \x01(\x0b\x32%.sglang.grpc.scheduler.TokenizedInput\"7\n\x13HealthCheckResponse\x12\x0f\n\x07healthy\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"2\n\x0c\x41\x62ortRequest\x12\x12\n\nrequest_id\x18\x01 \x01(\t\x12\x0e\n\x06reason\x18\x02 \x01(\t\"1\n\rAbortResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"I\n\x0fLoadLoRARequest\x12\x12\n\nadapter_id\x18\x01 \x01(\t\x12\x14\n\x0c\x61\x64\x61pter_path\x18\x02 \x01(\t\x12\x0c\n\x04rank\x18\x03 \x01(\x05\"H\n\x10LoadLoRAResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x12\n\nadapter_id\x18\x02 \x01(\t\x12\x0f\n\x07message\x18\x03 \x01(\t\"\'\n\x11UnloadLoRARequest\x12\x12\n\nadapter_id\x18\x01 \x01(\t\"6\n\x12UnloadLoRAResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"w\n\x14UpdateWeightsRequest\x12\x13\n\tdisk_path\x18\x01 \x01(\tH\x00\x12\x15\n\x0btensor_data\x18\x02 \x01(\x0cH\x00\x12\x14\n\nremote_url\x18\x03 \x01(\tH\x00\x12\x13\n\x0bweight_name\x18\x04 \x01(\tB\x08\n\x06source\"9\n\x15UpdateWeightsResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t\"-\n\x17GetInternalStateRequest\x12\x12\n\nstate_keys\x18\x01 \x03(\t\"B\n\x18GetInternalStateResponse\x12&\n\x05state\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\"A\n\x17SetInternalStateRequest\x12&\n\x05state\x18\x01 \x01(\x0b\x32\x17.google.protobuf.Struct\"<\n\x18SetInternalStateResponse\x12\x0f\n\x07success\x18\x01 \x01(\x08\x12\x0f\n\x07message\x18\x02 \x01(\t2\xfe\x02\n\x0fSglangScheduler\x12]\n\x08Generate\x12&.sglang.grpc.scheduler.GenerateRequest\x1a\'.sglang.grpc.scheduler.GenerateResponse0\x01\x12R\n\x05\x45mbed\x12#.sglang.grpc.scheduler.EmbedRequest\x1a$.sglang.grpc.scheduler.EmbedResponse\x12\x64\n\x0bHealthCheck\x12).sglang.grpc.scheduler.HealthCheckRequest\x1a*.sglang.grpc.scheduler.HealthCheckResponse\x12R\n\x05\x41\x62ort\x12#.sglang.grpc.scheduler.AbortRequest\x1a$.sglang.grpc.scheduler.AbortResponseb\x06proto3')
+
+_globals = globals()
+_builder.BuildMessageAndEnumDescriptors(DESCRIPTOR, _globals)
+_builder.BuildTopDescriptorsAndMessages(DESCRIPTOR, 'sglang_scheduler_pb2', _globals)
+if not _descriptor._USE_C_DESCRIPTORS:
+  DESCRIPTOR._loaded_options = None
+  _globals['_SAMPLINGPARAMS_LOGITBIASENTRY']._loaded_options = None
+  _globals['_SAMPLINGPARAMS_LOGITBIASENTRY']._serialized_options = b'8\001'
+  _globals['_SAMPLINGPARAMS']._serialized_start=113
+  _globals['_SAMPLINGPARAMS']._serialized_end=850
+  _globals['_SAMPLINGPARAMS_LOGITBIASENTRY']._serialized_start=769
+  _globals['_SAMPLINGPARAMS_LOGITBIASENTRY']._serialized_end=817
+  _globals['_DISAGGREGATEDPARAMS']._serialized_start=852
+  _globals['_DISAGGREGATEDPARAMS']._serialized_end=945
+  _globals['_GENERATEREQUEST']._serialized_start=948
+  _globals['_GENERATEREQUEST']._serialized_end=1558
+  _globals['_TOKENIZEDINPUT']._serialized_start=1560
+  _globals['_TOKENIZEDINPUT']._serialized_end=1618
+  _globals['_MULTIMODALINPUTS']._serialized_start=1621
+  _globals['_MULTIMODALINPUTS']._serialized_end=1832
+  _globals['_GENERATERESPONSE']._serialized_start=1835
+  _globals['_GENERATERESPONSE']._serialized_end=2062
+  _globals['_GENERATESTREAMCHUNK']._serialized_start=2065
+  _globals['_GENERATESTREAMCHUNK']._serialized_end=2342
+  _globals['_GENERATECOMPLETE']._serialized_start=2345
+  _globals['_GENERATECOMPLETE']._serialized_end=2756
+  _globals['_GENERATEERROR']._serialized_start=2758
+  _globals['_GENERATEERROR']._serialized_end=2833
+  _globals['_OUTPUTLOGPROBS']._serialized_start=2835
+  _globals['_OUTPUTLOGPROBS']._serialized_end=2952
+  _globals['_INPUTLOGPROBS']._serialized_start=2955
+  _globals['_INPUTLOGPROBS']._serialized_end=3113
+  _globals['_INPUTTOKENLOGPROB']._serialized_start=3115
+  _globals['_INPUTTOKENLOGPROB']._serialized_end=3164
+  _globals['_TOPLOGPROBS']._serialized_start=3166
+  _globals['_TOPLOGPROBS']._serialized_end=3214
+  _globals['_HIDDENSTATES']._serialized_start=3216
+  _globals['_HIDDENSTATES']._serialized_end=3279
+  _globals['_EMBEDREQUEST']._serialized_start=3282
+  _globals['_EMBEDREQUEST']._serialized_end=3612
+  _globals['_EMBEDRESPONSE']._serialized_start=3615
+  _globals['_EMBEDRESPONSE']._serialized_end=3772
+  _globals['_EMBEDCOMPLETE']._serialized_start=3775
+  _globals['_EMBEDCOMPLETE']._serialized_end=3938
+  _globals['_EMBEDDING']._serialized_start=3940
+  _globals['_EMBEDDING']._serialized_end=3982
+  _globals['_EMBEDERROR']._serialized_start=3984
+  _globals['_EMBEDERROR']._serialized_end=4044
+  _globals['_HEALTHCHECKREQUEST']._serialized_start=4046
+  _globals['_HEALTHCHECKREQUEST']._serialized_end=4124
+  _globals['_HEALTHCHECKRESPONSE']._serialized_start=4126
+  _globals['_HEALTHCHECKRESPONSE']._serialized_end=4181
+  _globals['_ABORTREQUEST']._serialized_start=4183
+  _globals['_ABORTREQUEST']._serialized_end=4233
+  _globals['_ABORTRESPONSE']._serialized_start=4235
+  _globals['_ABORTRESPONSE']._serialized_end=4284
+  _globals['_LOADLORAREQUEST']._serialized_start=4286
+  _globals['_LOADLORAREQUEST']._serialized_end=4359
+  _globals['_LOADLORARESPONSE']._serialized_start=4361
+  _globals['_LOADLORARESPONSE']._serialized_end=4433
+  _globals['_UNLOADLORAREQUEST']._serialized_start=4435
+  _globals['_UNLOADLORAREQUEST']._serialized_end=4474
+  _globals['_UNLOADLORARESPONSE']._serialized_start=4476
+  _globals['_UNLOADLORARESPONSE']._serialized_end=4530
+  _globals['_UPDATEWEIGHTSREQUEST']._serialized_start=4532
+  _globals['_UPDATEWEIGHTSREQUEST']._serialized_end=4651
+  _globals['_UPDATEWEIGHTSRESPONSE']._serialized_start=4653
+  _globals['_UPDATEWEIGHTSRESPONSE']._serialized_end=4710
+  _globals['_GETINTERNALSTATEREQUEST']._serialized_start=4712
+  _globals['_GETINTERNALSTATEREQUEST']._serialized_end=4757
+  _globals['_GETINTERNALSTATERESPONSE']._serialized_start=4759
+  _globals['_GETINTERNALSTATERESPONSE']._serialized_end=4825
+  _globals['_SETINTERNALSTATEREQUEST']._serialized_start=4827
+  _globals['_SETINTERNALSTATEREQUEST']._serialized_end=4892
+  _globals['_SETINTERNALSTATERESPONSE']._serialized_start=4894
+  _globals['_SETINTERNALSTATERESPONSE']._serialized_end=4954
+  _globals['_SGLANGSCHEDULER']._serialized_start=4957
+  _globals['_SGLANGSCHEDULER']._serialized_end=5339
+# @@protoc_insertion_point(module_scope)