config: publishing is the process entry's job (#36251)

test/manual/test_forward_split_prefill.py

@@ -19,6 +19,7 @@ from sglang.srt.distributed.parallel_state_wrapper import ParallelState
 from sglang.srt.managers.schedule_batch import Req, ScheduleBatch
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.model_executor.model_runner import ModelRunner
+from sglang.srt.runtime_context import publish
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
@@ -52,6 +53,8 @@ class TestForwardSplitPrefill(CustomTestCase):
 
         cls.port_args = PortArgs.init_new(cls.server_args)
 
+        publish(cls.server_args, role="scheduler")
+
         # Load model and tokenizer
         cls.model_config = ModelConfig.from_server_args(cls.server_args)
         cls.model_runner = ModelRunner(

test/manual/test_tokenizer_batch_encode.py

@@ -15,6 +15,7 @@ from unittest.mock import Mock, patch
 
 from sglang.srt.managers.io_struct import GenerateReqInput
 from sglang.srt.managers.tokenizer_manager import TokenizerManager
+from sglang.srt.runtime_context import publish
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST
 
@@ -39,6 +40,7 @@ class TestTokenizerBatchEncode(unittest.TestCase):
         ):
 
             mock_tokenizer.return_value = Mock(vocab_size=32000)
+            publish(self.server_args, role="tokenizer")
             self.tokenizer_manager = TokenizerManager(self.server_args, self.port_args)
 
     def test_batch_encode_enabled(self):

test/manual/test_tokenizer_manager.py

@@ -24,6 +24,7 @@ from sglang.srt.managers.tokenizer_manager import (
     TokenizerManager,
 )
 from sglang.srt.observability.req_time_stats import APIServerReqTimeStats
+from sglang.srt.runtime_context import publish
 from sglang.srt.server_args import PortArgs, ServerArgs
 from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST
 
@@ -45,6 +46,7 @@ class TestInputFormatDetection(unittest.TestCase):
             ) as mock_tokenizer,
         ):
             mock_tokenizer.return_value = Mock(vocab_size=32000)
+            publish(self.server_args, role="tokenizer")
             self.tokenizer_manager = TokenizerManager(self.server_args, self.port_args)
 
     def test_detect_single_string(self):
@@ -143,6 +145,7 @@ class TestTokenizerInputPreparation(unittest.TestCase):
             ) as mock_tokenizer,
         ):
             mock_tokenizer.return_value = Mock(vocab_size=32000)
+            publish(self.server_args, role="tokenizer")
             self.tokenizer_manager = TokenizerManager(self.server_args, self.port_args)
 
     def test_prepare_single_string_input(self):
@@ -203,6 +206,7 @@ class TestTokenizerResultExtraction(unittest.TestCase):
             ) as mock_tokenizer,
         ):
             mock_tokenizer.return_value = Mock(vocab_size=32000)
+            publish(self.server_args, role="tokenizer")
             self.tokenizer_manager = TokenizerManager(self.server_args, self.port_args)
 
     def test_extract_single_string_results(self):
@@ -327,6 +331,7 @@ class TestTokenizerManagerIntegration(unittest.TestCase):
             ) as mock_tokenizer,
         ):
             mock_tokenizer.return_value = Mock(vocab_size=32000)
+            publish(self.server_args, role="tokenizer")
             self.tokenizer_manager = TokenizerManager(self.server_args, self.port_args)
 
     def test_full_workflow_single_string(self):

test/manual/test_vlm_accuracy.py

@@ -20,6 +20,7 @@ from sglang.srt.managers.schedule_batch import (
 from sglang.srt.model_executor.model_runner import ModelRunner
 from sglang.srt.multimodal.processors.base_processor import BaseMultimodalProcessor
 from sglang.srt.parser.conversation import generate_chat_conv
+from sglang.srt.runtime_context import publish
 from sglang.srt.server_args import ServerArgs
 from sglang.test.test_utils import download_image_with_retry
 
@@ -141,16 +142,18 @@ class VisionLLMLogitsBase(unittest.IsolatedAsyncioTestCase):
         return inputs
 
     def get_sglang_model(self):
+        server_args = ServerArgs(
+            model_path=self.model_path,
+            disable_cuda_graph=True,
+        )
+        publish(server_args, role="scheduler")
         self.model_runner = ModelRunner(
             model_config=ModelConfig(self.model_path, model_override_args="{}"),
             mem_fraction_static=0.8,
             gpu_id=0,
             ps=ParallelState.trivial(),
             nccl_port=12435,
-            server_args=ServerArgs(
-                model_path=self.model_path,
-                disable_cuda_graph=True,
-            ),
+            server_args=server_args,
         )
         return self.model_runner.model