From 2e4a1e2d0570f66662cfb98fc2006b0aba2717fa Mon Sep 17 00:00:00 2001
From: Yinghai Lu <yinghai@thinkingmachines.ai>
Date: Sun, 16 Mar 2025 18:10:09 -0700
Subject: [PATCH] Initialize image processor for skip-tokenizer-init codepath
 (#4479)

Co-authored-by: Alex Kirillov <alex@iterationlab.org>
---
 .../sglang/srt/managers/tokenizer_manager.py  | 39 +++++++++++--------
 1 file changed, 22 insertions(+), 17 deletions(-)

diff --git a/python/sglang/srt/managers/tokenizer_manager.py b/python/sglang/srt/managers/tokenizer_manager.py
index 3132060ed..c211d76ff 100644
--- a/python/sglang/srt/managers/tokenizer_manager.py
+++ b/python/sglang/srt/managers/tokenizer_manager.py
@@ -168,27 +168,32 @@ class TokenizerManager:
         self.context_len = self.model_config.context_len
         self.image_token_id = self.model_config.image_token_id
 
-        # Create image processor placeholder
-        self.image_processor = get_dummy_image_processor()
+        if self.model_config.is_multimodal:
+            _processor = get_processor(
+                server_args.tokenizer_path,
+                tokenizer_mode=server_args.tokenizer_mode,
+                trust_remote_code=server_args.trust_remote_code,
+                revision=server_args.revision,
+            )
 
-        # Create tokenizer
-        if server_args.skip_tokenizer_init:
-            self.tokenizer = self.processor = None
-        else:
-            if self.model_config.is_multimodal:
-                self.processor = get_processor(
-                    server_args.tokenizer_path,
-                    tokenizer_mode=server_args.tokenizer_mode,
-                    trust_remote_code=server_args.trust_remote_code,
-                    revision=server_args.revision,
-                )
+            # We want to parallelize the image pre-processing so we create an executor for it
+            # We creat image_processor for any skip_tokenizer_init to make sure we still encode
+            # images even with skip_tokenizer_init=False.
+            self.image_processor = get_image_processor(
+                self.model_config.hf_config, server_args, _processor
+            )
+
+            if server_args.skip_tokenizer_init:
+                self.tokenizer = self.processor = None
+            else:
+                self.processor = _processor
                 self.tokenizer = self.processor.tokenizer
                 os.environ["TOKENIZERS_PARALLELISM"] = "false"
+        else:
+            self.image_processor = get_dummy_image_processor()
 
-                # We want to parallelize the image pre-processing so we create an executor for it
-                self.image_processor = get_image_processor(
-                    self.model_config.hf_config, server_args, self.processor
-                )
+            if server_args.skip_tokenizer_init:
+                self.tokenizer = self.processor = None
             else:
                 self.tokenizer = get_tokenizer(
                     server_args.tokenizer_path,