Fix deprecation warning of sre_parse for python 3.13 (#16247)

python/sglang/srt/managers/scheduler.py

@@ -733,8 +733,6 @@ class Scheduler(
         self.running_batch: ScheduleBatch = ScheduleBatch(reqs=[], batch_is_full=False)
         # The current forward batch
         self.cur_batch: Optional[ScheduleBatch] = None
-        # The current split prefill batch
-        self.split_prefill_batch: Optional[ScheduleBatch] = None
         # The last forward batch
         self.last_batch: Optional[ScheduleBatch] = None
         self.forward_ct = 0

python/sglang/srt/managers/tokenizer_manager.py

@@ -219,11 +219,12 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
 
     def init_model_config(self):
         server_args = self.server_args
+        model_config_class = getattr(self, "model_config_class", ModelConfig)
 
         # Read model args
         self.model_path = server_args.model_path
         self.served_model_name = server_args.served_model_name
-        self.model_config = ModelConfig.from_server_args(server_args)
+        self.model_config = model_config_class.from_server_args(server_args)
         self.is_generation = self.model_config.is_generation
         self.is_image_gen = self.model_config.is_image_gen
         self.context_len = self.model_config.context_len
@@ -232,11 +233,15 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
         speculative_algorithm = SpeculativeAlgorithm.from_string(
             server_args.speculative_algorithm
         )
-        self.reserve_input_token_num = (
+        if speculative_algorithm.is_eagle():
-            0
+            # In the current eagle implementation, we store the draft tokens in the output token slots,
-            if speculative_algorithm.is_none()
+            # so we need to reserve the space for the draft tokens.
-            else server_args.speculative_num_draft_tokens
+            self.num_reserved_tokens = max(
+                server_args.speculative_eagle_topk * server_args.speculative_num_steps,
+                server_args.speculative_num_draft_tokens,
             )
+        else:
+            self.num_reserved_tokens = 0
         self.validate_total_tokens = True
 
     def init_tokenizer_and_processor(self):
@@ -457,6 +462,9 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
         )
         self.init_communicators(self.server_args)
 
+        self.sampling_params_class = SamplingParams
+        self.signal_handler_class = SignalHandler
+
     async def generate_request(
         self,
         obj: Union[GenerateReqInput, EmbeddingReqInput],
@@ -724,7 +732,7 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
         # FIXME: unify the length validation logic with the one in the scheduler.
         _max_req_len = self.context_len
         input_token_num = len(input_ids) if input_ids is not None else 0
-        input_token_num += self.reserve_input_token_num
+        input_token_num += self.num_reserved_tokens
 
         # Validate input length
         if input_token_num >= self.context_len:
@@ -860,9 +868,6 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
                     f"The input_ids {input_ids} contains values greater than the vocab size ({vocab_size})."
                 )
 
-    def _get_sampling_params(self, sampling_kwargs: Dict) -> SamplingParams:
-        return SamplingParams(**sampling_kwargs)
-
     def _create_tokenized_object(
         self,
         obj: Union[GenerateReqInput, EmbeddingReqInput],
@@ -880,7 +885,7 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
             sampling_kwargs = {**self.preferred_sampling_params, **obj.sampling_params}
         else:
             sampling_kwargs = obj.sampling_params
-        sampling_params = self._get_sampling_params(sampling_kwargs)
+        sampling_params = self.sampling_params_class(**sampling_kwargs)
         sampling_params.normalize(self.tokenizer)
         sampling_params.verify(self.model_config.vocab_size)
 
@@ -1412,21 +1417,22 @@ class TokenizerManager(TokenizerCommunicatorMixin, TokenizerManagerMultiItemMixi
         )
         self.event_loop = loop
 
-        # We cannot add signal handler when the tokenizer manager is not in
-        # the main thread due to the CPython limitation.
         if threading.current_thread() is threading.main_thread():
-            signal_handler = SignalHandler(self)
+            signal_handler = self.signal_handler_class(self)
             loop.add_signal_handler(signal.SIGTERM, signal_handler.sigterm_handler)
             # Update the signal handler for the process. It overrides the sigquit handler in the launch phase.
             loop.add_signal_handler(
                 signal.SIGQUIT, signal_handler.running_phase_sigquit_handler
             )
         else:
+            # We cannot add signal handler when the tokenizer manager is not in
+            # the main thread due to the CPython limitation.
             logger.warning(
                 "Signal handler is not added because the tokenizer manager is "
                 "not in the main thread. This disables graceful shutdown of the "
                 "tokenizer manager when SIGTERM is received."
             )
+
         self.asyncio_tasks.add(
             loop.create_task(print_exception_wrapper(self.sigterm_watchdog))
         )

python/sglang/srt/multiplex/multiplexing_mixin.py

@@ -5,7 +5,7 @@ Mixin class providing multiplexing scheduling logic
 from __future__ import annotations
 
 import logging
-from typing import TYPE_CHECKING
+from typing import TYPE_CHECKING, Optional
 
 import torch
 import torch.distributed as dist
@@ -23,6 +23,7 @@ from sglang.srt.multiplex.pdmux_context import (
 )
 
 if TYPE_CHECKING:
+    from sglang.srt.managers.schedule_batch import ScheduleBatch
     from sglang.srt.managers.scheduler import Scheduler
 
 logger = logging.getLogger(__name__)
@@ -31,6 +32,9 @@ logger = logging.getLogger(__name__)
 class SchedulerMultiplexMixin:
 
     def init_pdmux(self: Scheduler):
+        # The current split prefill batch
+        self.split_prefill_batch: Optional[ScheduleBatch] = None
+
         # for pd_multiplexing, Init stream_groups, exclude normal stream for prefill only and decode only
         self.pdmux_config = load_pdmux_config(self.server_args.pdmux_config_path)
         initialize_stream_groups(self.gpu_id, self.pdmux_config)

python/sglang/srt/sampling/sampling_params.py

@@ -14,9 +14,14 @@
 """Sampling parameters for text generation."""
 
 import logging
-import sre_parse
 from typing import Any, Dict, List, Optional, Union
 
+# sre_parse is deprecated in Python 3.11+, use re._parser instead
+try:
+    import re._parser as sre_parse
+except ImportError:
+    import sre_parse  # Python < 3.11
+
 _SAMPLING_EPS = 1e-6
 TOP_K_ALL = 1 << 30
 

python/sglang/srt/server_args.py

@@ -5198,4 +5198,4 @@ def auto_choose_speculative_params(self: ServerArgs):
         return (5, 4, 8)
     else:
         # The default value for all other models
-        return (5, 4, 8)
+        return (3, 1, 4)