sglang 0.5.2rc2__py3-none-any.whl → 0.5.3rc0__py3-none-any.whl

sglang/srt/eplb/expert_distribution.py

@@ -11,6 +11,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+
+from __future__ import annotations
+
 import logging
 import math
 import os
@@ -19,16 +22,20 @@ from abc import ABC
 from collections import deque
 from contextlib import contextmanager
 from pathlib import Path
-from typing import Any, Dict, List, Literal, Optional, Tuple, Type
+from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional, Tuple, Type
 
 import einops
 import torch
 import torch.distributed
 
-from sglang.srt.eplb.expert_location import ExpertLocationMetadata
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import Withable, get_bool_env_var
+from sglang.srt.utils import Withable, get_bool_env_var, is_npu
+
+_is_npu = is_npu()
+
+if TYPE_CHECKING:
+    from sglang.srt.eplb.expert_location import ExpertLocationMetadata
 
 logger = logging.getLogger(__name__)
 
@@ -43,7 +50,7 @@ class ExpertDistributionRecorder(ABC):
     @staticmethod
     def init_new(
         server_args: ServerArgs,
-        expert_location_metadata: "ExpertLocationMetadata",
+        expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
         if server_args.expert_distribution_recorder_mode is not None:
@@ -118,7 +125,7 @@ class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
     def __init__(
         self,
         server_args: ServerArgs,
-        expert_location_metadata: "ExpertLocationMetadata",
+        expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
         self._server_args = server_args
@@ -211,7 +218,9 @@ class _ExpertDistributionRecorderReal(ExpertDistributionRecorder):
     def _on_hook(self, hook_name: str, **kwargs):
         if self._disable_all:
             return
-        if not (self._recording or torch.cuda.is_current_stream_capturing()):
+        if not (
+            self._recording or torch.get_device_module().is_current_stream_capturing()
+        ):
             return
         gatherer = self._single_pass_gatherers[
             self._accumulator.get_single_pass_gatherer_key(
@@ -279,7 +288,7 @@ class _SinglePassGatherer(ABC):
     @staticmethod
     def init_new(
         server_args: ServerArgs,
-        expert_location_metadata: "ExpertLocationMetadata",
+        expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ) -> "_SinglePassGatherer":
         if server_args.expert_distribution_recorder_mode == "per_token":
@@ -307,7 +316,7 @@ class _SinglePassGatherer(ABC):
 
         return _SelectExpertsSinglePassGatherer(expert_location_metadata, rank)
 
-    def __init__(self, expert_location_metadata: "ExpertLocationMetadata", rank: int):
+    def __init__(self, expert_location_metadata: ExpertLocationMetadata, rank: int):
         self._expert_location_metadata = expert_location_metadata
         self._rank = rank
 
@@ -346,7 +355,7 @@ class _DetailSinglePassGatherer(_SinglePassGatherer):
     def __init__(
         self,
         server_args: ServerArgs,
-        expert_location_metadata: "ExpertLocationMetadata",
+        expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
         super().__init__(expert_location_metadata, rank)
@@ -446,6 +455,10 @@ def _list_sum(a: List, b: List) -> List:
 class _LayerBasedGpuSinglePassGatherer(_SinglePassGatherer):
     def __init__(self, *args, enable_global_physical_experts: bool, **kwargs):
         super().__init__(*args, **kwargs)
+        if not _is_npu:
+            device = "cuda"
+        else:
+            device = "npu"
         self._enable_global_physical_experts = enable_global_physical_experts
         self._data = torch.zeros(
             (
@@ -457,7 +470,7 @@ class _LayerBasedGpuSinglePassGatherer(_SinglePassGatherer):
                 ),
             ),
             dtype=torch.int,
-            device="cuda",
+            device=device,
         )
 
     def reset(self):
@@ -561,7 +574,7 @@ class _Accumulator(ABC):
     @staticmethod
     def init_new(
         server_args: ServerArgs,
-        expert_location_metadata: "ExpertLocationMetadata",
+        expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ) -> "_Accumulator":
         return _Accumulator.get_class(server_args)(
@@ -580,7 +593,7 @@ class _Accumulator(ABC):
     def __init__(
         self,
         server_args: ServerArgs,
-        expert_location_metadata: "ExpertLocationMetadata",
+        expert_location_metadata: ExpertLocationMetadata,
         rank: int,
     ):
         self._server_args = server_args
@@ -779,7 +792,7 @@ class _StatAccumulator(_UtilizationRateAccumulatorMixin):
 
         if self._first_dump:
             self._first_dump = False
-            torch.cuda.empty_cache()
+            torch.get_device_module().empty_cache()
 
         torch.distributed.all_reduce(
             logical_count_of_buffered_step, op=torch.distributed.ReduceOp.SUM

sglang/srt/eplb/expert_location.py

@@ -11,21 +11,26 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+
+from __future__ import annotations
+
 import json
 import logging
 import random
 from dataclasses import dataclass
 from pathlib import Path
-from typing import List, Optional
+from typing import TYPE_CHECKING, List, Optional
 
 import torch
 import torch.distributed
 import torch.nn.functional as F
 
-from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.eplb import eplb_algorithms
 from sglang.srt.model_loader import get_model_architecture
-from sglang.srt.server_args import ServerArgs
+
+if TYPE_CHECKING:
+    from sglang.srt.configs.model_config import ModelConfig
+    from sglang.srt.server_args import ServerArgs
 
 logger = logging.getLogger(__name__)
 

sglang/srt/eplb/expert_location_updater.py

@@ -47,7 +47,7 @@ class ExpertLocationUpdater:
     ):
         if self._first_execution:
             self._first_execution = False
-            torch.cuda.empty_cache()
+            torch.get_device_module().empty_cache()
 
         old_expert_location_metadata = get_global_expert_location_metadata()
         assert old_expert_location_metadata is not None

sglang/srt/function_call/base_format_detector.py

@@ -162,12 +162,9 @@ class BaseFormatDetector(ABC):
 
         try:
             try:
-                if current_text.startswith(self.bot_token):
-                    start_idx = len(self.bot_token)
-                elif self.current_tool_id > 0 and current_text.startswith(
-                    self.tool_call_separator + self.bot_token
-                ):
-                    start_idx = len(self.tool_call_separator + self.bot_token)
+                tool_call_pos = current_text.find(self.bot_token)
+                if tool_call_pos != -1:
+                    start_idx = tool_call_pos + len(self.bot_token)
                 elif self.current_tool_id > 0 and current_text.startswith(
                     self.tool_call_separator
                 ):

sglang/srt/function_call/ebnf_composer.py

@@ -50,19 +50,19 @@ class EBNFComposer:
 
     CALL_RULE_MAP = {
         "pythonic": 'call_{name} ::= "{name}" "(" {arguments_rule} ")"',
-        "json": 'call_{name} ::= "{{" "\\"name\\"" ":" "\\"{name}\\"" ", " "\\"arguments\\"" ":" {arguments_rule} "}}"',
+        "json": 'call_{name} ::= "{{" ws "\\"name\\"" ws ":" ws "\\"{name}\\"" ws "," ws "\\"arguments\\"" ws ":" ws {arguments_rule} ws "}}"',
         "xml": 'call_{name} ::= "<function={name}>\\n" {arguments_rule} "\\n</function>"',
     }
 
     ARGUMENTS_RULE_MAP = {
         "pythonic": "{arg_rules}",
-        "json": '"{{" {arg_rules} "}}"',
+        "json": '"{{" ws {arg_rules} ws "}}"',
         "xml": "{arg_rules}",
     }
 
     KEY_VALUE_RULE_MAP = {
         "pythonic": '"{key}" "=" {valrule}',
-        "json": '"\\"{key}\\"" ":" {valrule}',
+        "json": '"\\"{key}\\"" ws ":" ws {valrule}',
         "xml": '"<parameter={key}>\\n" {valrule} "\\n</parameter>"',
     }
 
@@ -165,7 +165,7 @@ class EBNFComposer:
         tool_call_separator: Optional[str] = None,
         call_rule_fmt: Optional[str] = None,
         key_value_rule_fmt: Optional[str] = None,
-        key_value_separator: str = ",",
+        key_value_separator: str = 'ws "," ws',
     ):
         """
         Generalized EBNF builder for all detectors.
@@ -183,6 +183,10 @@ class EBNFComposer:
             key_value_rule_fmt: Optional custom format string for key-value pairs. It should define how each parameter is formatted,
                 with placeholders {key} for the parameter name and {valrule} for the value rule. If None, a default format
                 based on function_format will be used.
+            key_value_separator: Raw EBNF fragment inserted between key-value pairs.
+                This string is used verbatim (not auto-quoted). Pass:
+                - Quoted terminals when you need a literal token (e.g. '","' or '"\\n"').
+                - Raw/non-terminals when you need grammar tokens (e.g. 'ws "," ws').
         """
         # =================================================================
         # Step 1: Determine the root tool calls rule
@@ -281,9 +285,7 @@ class EBNFComposer:
             # Add required properties joined by commas
             if required:
                 rule_parts.append(
-                    f' "{key_value_separator}" '.join(
-                        prop_kv_pairs[k] for k in required
-                    )
+                    f" {key_value_separator} ".join(prop_kv_pairs[k] for k in required)
                 )
 
             # Add optional properties with flexible ordering
@@ -298,14 +300,14 @@ class EBNFComposer:
                             opt_parts.append(prop_kv_pairs[optional[j]])
                         else:
                             opt_parts.append(
-                                f' ( "{key_value_separator}" {prop_kv_pairs[optional[j]]} )?'
+                                f" ( {key_value_separator} {prop_kv_pairs[optional[j]]} )?"
                             )
                     opt_alternatives.append("".join(opt_parts))
 
                 # Wrap with appropriate comma handling based on whether we have required properties
                 if required:
                     # Required properties exist, so optional group needs outer comma
-                    rule_parts.append(f' ( "{key_value_separator}" ( ')
+                    rule_parts.append(f" ( {key_value_separator} ( ")
                     rule_parts.append(" | ".join(opt_alternatives))
                     rule_parts.append(" ) )?")
                 else:

sglang/srt/function_call/function_call_parser.py

@@ -69,6 +69,8 @@ class FunctionCallParser:
         Returns:
             True if the text contains a tool call, False otherwise
         """
+        if not self.tools:
+            return False
         return self.detector.has_tool_call(text)
 
     def parse_non_stream(self, full_text: str) -> Tuple[str, list[ToolCallItem]]:
@@ -83,6 +85,8 @@ class FunctionCallParser:
             - The remaining text after parsing that was not consumed by the detector (can be treated as normal text)
             - A list of tool calls parsed from the text
         """
+        if not self.tools:
+            return full_text, []
         parsed_result = self.detector.detect_and_parse(full_text, self.tools)
         tool_call_list = parsed_result.calls
         if tool_call_list:
@@ -102,6 +106,8 @@ class FunctionCallParser:
             - The normal text that should be displayed to the user
             - A list of tool calls parsed from the chunk
         """
+        if not self.tools:
+            return chunk_text, []
         final_normal_text = ""
         final_calls = []
 

sglang/srt/function_call/glm4_moe_detector.py

@@ -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/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/grpc/__init__.py

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

sglang/srt/grpc/sglang_scheduler_pb2.py

@@ -0,0 +1,106 @@
+# -*- 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\"\xc7\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\x16\n\x0emax_new_tokens\x18\x08 \x01(\x05\x12\x0c\n\x04stop\x18\t \x03(\t\x12\x16\n\x0estop_token_ids\x18\n \x03(\x05\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\x11\n\tlora_path\x18\x10 \x01(\t\x12\t\n\x01n\x18\x11 \x01(\x05\x12\x15\n\rtoken_healing\x18\x12 \x01(\x08\x12\x16\n\x0emin_new_tokens\x18\x13 \x01(\x05\x12\x12\n\nignore_eos\x18\x14 \x01(\x08\x12\x14\n\x0cno_stop_trim\x18\x15 \x01(\x08\x12\x17\n\x0fstream_interval\x18\x16 \x01(\x05\x12H\n\nlogit_bias\x18\x17 \x03(\x0b\x32\x34.sglang.grpc.scheduler.SamplingParams.LogitBiasEntry\x12\x16\n\x0estructural_tag\x18\x18 \x01(\t\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\nconstraint\"]\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\"\xe9\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(\x05\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\x15\n\rdp_balance_id\x18\x11 \x01(\x05\":\n\x0eTokenizedInput\x12\x15\n\roriginal_text\x18\x01 \x01(\t\x12\x11\n\tinput_ids\x18\x02 \x03(\x05\"\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\"\xf5\x01\n\x13GenerateStreamChunk\x12\x10\n\x08token_id\x18\x01 \x01(\x05\x12\x0c\n\x04text\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\x31\n\x08logprobs\x18\x06 \x01(\x0b\x32\x1f.sglang.grpc.scheduler.LogProbs\x12\x15\n\rhidden_states\x18\x07 \x03(\x02\x12\x17\n\x0fgeneration_time\x18\x08 \x01(\x02\x12\x12\n\nqueue_time\x18\t \x01(\x05\"\xcd\x02\n\x10GenerateComplete\x12\x12\n\noutput_ids\x18\x01 \x03(\x05\x12\x13\n\x0boutput_text\x18\x02 \x01(\t\x12K\n\rfinish_reason\x18\x03 \x01(\x0e\x32\x34.sglang.grpc.scheduler.GenerateComplete.FinishReason\x12\x35\n\x0c\x61ll_logprobs\x18\x0b \x03(\x0b\x32\x1f.sglang.grpc.scheduler.LogProbs\x12>\n\x11\x61ll_hidden_states\x18\x0c \x03(\x0b\x32#.sglang.grpc.scheduler.HiddenStates\"L\n\x0c\x46inishReason\x12\x08\n\x04STOP\x10\x00\x12\n\n\x06LENGTH\x10\x01\x12\r\n\tEOS_TOKEN\x10\x02\x12\x0c\n\x08STOP_STR\x10\x03\x12\t\n\x05\x41\x42ORT\x10\x04\"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\"\x84\x01\n\x08LogProbs\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\x12\x13\n\x0btoken_texts\x18\x04 \x03(\t\"E\n\x0bTopLogProbs\x12\x0e\n\x06values\x18\x01 \x03(\x02\x12\x11\n\ttoken_ids\x18\x02 \x03(\x05\x12\x13\n\x0btoken_texts\x18\x03 \x03(\t\"?\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\"\xbc\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\x17\n\x0fgeneration_time\x18\x05 \x01(\x02\x12:\n\x10\x62\x61tch_embeddings\x18\x06 \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=824
+  _globals['_SAMPLINGPARAMS_LOGITBIASENTRY']._serialized_start=762
+  _globals['_SAMPLINGPARAMS_LOGITBIASENTRY']._serialized_end=810
+  _globals['_DISAGGREGATEDPARAMS']._serialized_start=826
+  _globals['_DISAGGREGATEDPARAMS']._serialized_end=919
+  _globals['_GENERATEREQUEST']._serialized_start=922
+  _globals['_GENERATEREQUEST']._serialized_end=1539
+  _globals['_TOKENIZEDINPUT']._serialized_start=1541
+  _globals['_TOKENIZEDINPUT']._serialized_end=1599
+  _globals['_MULTIMODALINPUTS']._serialized_start=1602
+  _globals['_MULTIMODALINPUTS']._serialized_end=1813
+  _globals['_GENERATERESPONSE']._serialized_start=1816
+  _globals['_GENERATERESPONSE']._serialized_end=2043
+  _globals['_GENERATESTREAMCHUNK']._serialized_start=2046
+  _globals['_GENERATESTREAMCHUNK']._serialized_end=2291
+  _globals['_GENERATECOMPLETE']._serialized_start=2294
+  _globals['_GENERATECOMPLETE']._serialized_end=2627
+  _globals['_GENERATECOMPLETE_FINISHREASON']._serialized_start=2551
+  _globals['_GENERATECOMPLETE_FINISHREASON']._serialized_end=2627
+  _globals['_GENERATEERROR']._serialized_start=2629
+  _globals['_GENERATEERROR']._serialized_end=2704
+  _globals['_LOGPROBS']._serialized_start=2707
+  _globals['_LOGPROBS']._serialized_end=2839
+  _globals['_TOPLOGPROBS']._serialized_start=2841
+  _globals['_TOPLOGPROBS']._serialized_end=2910
+  _globals['_HIDDENSTATES']._serialized_start=2912
+  _globals['_HIDDENSTATES']._serialized_end=2975
+  _globals['_EMBEDREQUEST']._serialized_start=2978
+  _globals['_EMBEDREQUEST']._serialized_end=3308
+  _globals['_EMBEDRESPONSE']._serialized_start=3311
+  _globals['_EMBEDRESPONSE']._serialized_end=3468
+  _globals['_EMBEDCOMPLETE']._serialized_start=3471
+  _globals['_EMBEDCOMPLETE']._serialized_end=3659
+  _globals['_EMBEDDING']._serialized_start=3661
+  _globals['_EMBEDDING']._serialized_end=3703
+  _globals['_EMBEDERROR']._serialized_start=3705
+  _globals['_EMBEDERROR']._serialized_end=3765
+  _globals['_HEALTHCHECKREQUEST']._serialized_start=3767
+  _globals['_HEALTHCHECKREQUEST']._serialized_end=3845
+  _globals['_HEALTHCHECKRESPONSE']._serialized_start=3847
+  _globals['_HEALTHCHECKRESPONSE']._serialized_end=3902
+  _globals['_ABORTREQUEST']._serialized_start=3904
+  _globals['_ABORTREQUEST']._serialized_end=3954
+  _globals['_ABORTRESPONSE']._serialized_start=3956
+  _globals['_ABORTRESPONSE']._serialized_end=4005
+  _globals['_LOADLORAREQUEST']._serialized_start=4007
+  _globals['_LOADLORAREQUEST']._serialized_end=4080
+  _globals['_LOADLORARESPONSE']._serialized_start=4082
+  _globals['_LOADLORARESPONSE']._serialized_end=4154
+  _globals['_UNLOADLORAREQUEST']._serialized_start=4156
+  _globals['_UNLOADLORAREQUEST']._serialized_end=4195
+  _globals['_UNLOADLORARESPONSE']._serialized_start=4197
+  _globals['_UNLOADLORARESPONSE']._serialized_end=4251
+  _globals['_UPDATEWEIGHTSREQUEST']._serialized_start=4253
+  _globals['_UPDATEWEIGHTSREQUEST']._serialized_end=4372
+  _globals['_UPDATEWEIGHTSRESPONSE']._serialized_start=4374
+  _globals['_UPDATEWEIGHTSRESPONSE']._serialized_end=4431
+  _globals['_GETINTERNALSTATEREQUEST']._serialized_start=4433
+  _globals['_GETINTERNALSTATEREQUEST']._serialized_end=4478
+  _globals['_GETINTERNALSTATERESPONSE']._serialized_start=4480
+  _globals['_GETINTERNALSTATERESPONSE']._serialized_end=4546
+  _globals['_SETINTERNALSTATEREQUEST']._serialized_start=4548
+  _globals['_SETINTERNALSTATEREQUEST']._serialized_end=4613
+  _globals['_SETINTERNALSTATERESPONSE']._serialized_start=4615
+  _globals['_SETINTERNALSTATERESPONSE']._serialized_end=4675
+  _globals['_SGLANGSCHEDULER']._serialized_start=4678
+  _globals['_SGLANGSCHEDULER']._serialized_end=5060
+# @@protoc_insertion_point(module_scope)