[BugFix] Preserve tokenizer worker fanout when skip_tokenizer_init is enabled (#30682)

Co-authored-by: hjzhang <zhanghjzzz@qq.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>

python/sglang/srt/managers/scheduler_components/ipc_channels.py

@@ -53,7 +53,8 @@ class SchedulerIpcChannels:
                 context, zmq.PUSH, port_args.tokenizer_ipc_name, False
             )
             if skip_tokenizer_init:
-                # Directly send to the TokenizerManager
+                # No decode work: send outputs straight to the tokenizer side
+                # (MultiTokenizerRouter fans out when tokenizer_worker_num > 1).
                 send_to_detokenizer_raw = get_zmq_socket(
                     context, zmq.PUSH, port_args.tokenizer_ipc_name, False
                 )

python/sglang/srt/managers/tokenizer_manager.py

@@ -601,6 +601,8 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 (FreezeGCReq, lambda x: None),
                 # For handling case when scheduler skips detokenizer and forwards back to the tokenizer manager, we ignore it.
                 (HealthCheckOutput, lambda x: None),
+                # Same skip-detokenizer forwarding case as above.
+                (ConfigureLoggingReq, lambda x: None),
                 (ActiveRanksOutput, self.update_active_ranks),
             ]
         )

python/sglang/srt/server_args.py

@@ -6089,16 +6089,12 @@ class ServerArgs:
             )
 
         if self.skip_tokenizer_init:
-            if self.tokenizer_worker_num != 1:
+            # Tokenizer workers still serve HTTP / state / output work, so
-                logger.warning(
+            # their fanout is preserved; detokenizer workers only decode.
-                    "skip_tokenizer_init=True disables tokenizer workers; forcing tokenizer_worker_num=1 "
-                    f"(requested {self.tokenizer_worker_num})."
-                )
-                self.tokenizer_worker_num = 1
             if self.detokenizer_worker_num != 1:
                 logger.warning(
-                    "skip_tokenizer_init=True disables detokenizer workers; forcing detokenizer_worker_num=1 "
+                    "skip_tokenizer_init=True leaves no decode work for detokenizer workers; "
-                    f"(requested {self.detokenizer_worker_num})."
+                    f"forcing detokenizer_worker_num=1 (requested {self.detokenizer_worker_num})."
                 )
                 self.detokenizer_worker_num = 1
 

test/registered/tokenizer/test_skip_tokenizer_init.py

@@ -36,7 +36,12 @@ class TestSkipTokenizerInit(CustomTestCase):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--skip-tokenizer-init", "--incremental-streaming-output"],
+            other_args=[
+                "--skip-tokenizer-init",
+                "--incremental-streaming-output",
+                "--tokenizer-worker-num",
+                "4",
+            ],
         )
         cls.eos_token_id = [119690]
         cls.tokenizer = AutoTokenizer.from_pretrained(
@@ -216,7 +221,11 @@ class TestSkipTokenizerInitVLM(TestSkipTokenizerInit):
             cls.model,
             cls.base_url,
             timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
-            other_args=["--skip-tokenizer-init"],
+            other_args=[
+                "--skip-tokenizer-init",
+                "--tokenizer-worker-num",
+                "4",
+            ],
         )
         cls.eos_token_id = [cls.tokenizer.eos_token_id]
 

test/registered/unit/server_args/test_server_args.py

@@ -232,6 +232,22 @@ class TestLoadBalanceMethod(unittest.TestCase):
         self.assertEqual(server_args.disaggregation_transfer_backend, "mooncake")
 
 
+class TestSkipTokenizerInit(unittest.TestCase):
+    def test_skip_tokenizer_worker_counts(self):
+        server_args = ServerArgs(
+            model_path="dummy",
+            skip_tokenizer_init=True,
+            tokenizer_worker_num=4,
+            detokenizer_worker_num=3,
+        )
+
+        server_args._handle_tokenizer_batching()
+
+        # Tokenizer fanout preserved; detokenizer coerced to 1 (no decode work).
+        self.assertEqual(server_args.tokenizer_worker_num, 4)
+        self.assertEqual(server_args.detokenizer_worker_num, 1)
+
+
 class TestHiSparseDsaBackendPolicy(unittest.TestCase):
     # The backend selection moved to the resolution pipeline; these policy
     # tests drive the pass through its read-only view.