fix: InternS1 don't recognize image, updates image token for InternVL processor (#9381)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

python/sglang/srt/multimodal/processors/internvl.py

@@ -44,7 +44,7 @@ class InternVLImageProcessor(BaseMultimodalProcessor):
         self.img_start_token_id = tokenizer.convert_tokens_to_ids(self.IMG_START_TOKEN)
         self.img_end_token_id = tokenizer.convert_tokens_to_ids(self.IMG_END_TOKEN)
         self.mm_tokens = MultimodalSpecialTokens(
-            image_token="<image>",
+            image_token="<IMG_CONTEXT>",
             image_token_id=tokenizer.convert_tokens_to_ids(self.IMG_CONTEXT_TOKEN),
         ).build(_image_processor)
 
@@ -218,13 +218,18 @@ class InternVLImageProcessor(BaseMultimodalProcessor):
 
         pixel_values = torch.cat(pixel_values, dim=0)
 
+        original_placeholder = "<<<__IMG_CONTEXT_PLACEHOLDER__>>>"
+        input_text = input_text.replace(self.IMG_CONTEXT_TOKEN, original_placeholder)
+
         for idx, num_patches in enumerate(num_patches_list):
             image_tokens = (
                 self.IMG_START_TOKEN
                 + self.IMG_CONTEXT_TOKEN * self.num_image_token * num_patches
                 + self.IMG_END_TOKEN
             )
-            input_text = input_text.replace("<image>", image_tokens, 1)
+            input_text = input_text.replace(original_placeholder, image_tokens, 1)
+
+        input_text = input_text.replace(original_placeholder, self.IMG_CONTEXT_TOKEN)
 
         input_ids = self.tokenizer(input_text, return_tensors="pt")[
             "input_ids"