fix missing revision arg when loading tokenizer (#2982)

python/sglang/srt/managers/detokenizer_manager.py

@@ -71,6 +71,7 @@ class DetokenizerManager:
                 server_args.tokenizer_path,
                 tokenizer_mode=server_args.tokenizer_mode,
                 trust_remote_code=server_args.trust_remote_code,
+                revision=server_args.revision,
             )
 
         self.decode_status = LimitedCapacityDict()

python/sglang/srt/managers/scheduler.py

@@ -206,6 +206,7 @@ class Scheduler:
                     server_args.tokenizer_path,
                     tokenizer_mode=server_args.tokenizer_mode,
                     trust_remote_code=server_args.trust_remote_code,
+                    revision=server_args.revision,
                 )
                 self.tokenizer = self.processor.tokenizer
             else:
@@ -213,6 +214,7 @@ class Scheduler:
                     server_args.tokenizer_path,
                     tokenizer_mode=server_args.tokenizer_mode,
                     trust_remote_code=server_args.trust_remote_code,
+                    revision=server_args.revision,
                 )
 
         # Check whether overlap can be enabled

python/sglang/srt/managers/tokenizer_manager.py

@@ -158,6 +158,7 @@ class TokenizerManager:
                     server_args.tokenizer_path,
                     tokenizer_mode=server_args.tokenizer_mode,
                     trust_remote_code=server_args.trust_remote_code,
+                    revision=server_args.revision,
                 )
                 self.tokenizer = self.processor.tokenizer
                 os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -171,6 +172,7 @@ class TokenizerManager:
                     server_args.tokenizer_path,
                     tokenizer_mode=server_args.tokenizer_mode,
                     trust_remote_code=server_args.trust_remote_code,
+                    revision=server_args.revision,
                 )
 
         # Store states

python/sglang/srt/managers/tp_worker.py

@@ -83,6 +83,7 @@ class TpModelWorker:
                     server_args.tokenizer_path,
                     tokenizer_mode=server_args.tokenizer_mode,
                     trust_remote_code=server_args.trust_remote_code,
+                    revision=server_args.revision,
                 )
                 self.tokenizer = self.processor.tokenizer
             else:
@@ -90,6 +91,7 @@ class TpModelWorker:
                     server_args.tokenizer_path,
                     tokenizer_mode=server_args.tokenizer_mode,
                     trust_remote_code=server_args.trust_remote_code,
+                    revision=server_args.revision,
                 )
         self.device = self.model_runner.device
 

python/sglang/srt/server.py

@@ -1027,6 +1027,7 @@ class Runtime:
             self.server_args.tokenizer_path,
             tokenizer_mode=self.server_args.tokenizer_mode,
             trust_remote_code=self.server_args.trust_remote_code,
+            revision=self.server_args.revision,
         )
 
     async def async_generate(