0.19.1

README.md

@@ -91,6 +91,10 @@ See: [OpenVLM Leaderboard](https://huggingface.co/spaces/opencompass/open_vlm_le
 
 ## Recent updates
 
+Version 0.19.1
+
+- really Fix <|end|> token for Mini-InternVL-Chat-4B-V1-5, thanks again [@Ph0rk0z](https://github.com/Ph0rk0z)
+
 Version 0.19.0
 
 - new model support: tiiuae/falcon-11B-vlm

backend/internvl-chat-v1-5.py

@@ -105,10 +105,10 @@ def __init__(self, model_id: str, device: str, device_map: str = 'auto', extra_p
 
         self.model.img_context_token_id = self.tokenizer.convert_tokens_to_ids('<IMG_CONTEXT>')
 
-        if self.format == 'phintern' and self.tokenizer.convert_tokens_to_ids('<|end|>') != 0:
-            self.eos_token_id = self.tokenizer.convert_tokens_to_ids('<|end|>')
-        elif self.tokenizer.convert_tokens_to_ids('<|im_end|>') != 0:
-            self.eos_token_id = self.tokenizer.convert_tokens_to_ids('<|im_end|>')  # 92542, InternLM2
+        self.eos_token = '<|end|>' if self.format == 'phintern' else '<|im_end|>'
+
+        if self.tokenizer.convert_tokens_to_ids(self.eos_token) != 0:
+            self.eos_token_id = self.tokenizer.convert_tokens_to_ids(self.eos_token)  # 92542, InternLM2
         else:
             self.eos_token_id = self.tokenizer.eos_token_id
 
@@ -151,4 +151,4 @@ async def chat_with_images(self, request: ImageChatRequest) -> str:
         )
         response = self.tokenizer.decode(output[0], skip_special_tokens=True)
 
-        return response.split('<|im_end|>')[0].strip()
+        return response.split(self.eos_token)[0].strip()