fix: modify the gemma3 chat template to be compatible with openai api

Author: kossum

llama_cpp/llama_chat_format.py

@@ -3564,7 +3564,7 @@ class Gemma3ChatHandler(Llava15ChatHandler):
         "{{ message['content'] | trim }}"
         "{%- elif message['content'] is iterable -%}"
         "{%- for item in message['content'] -%}"
-        "{%- if item['type'] == 'image' -%}"
+        "{%- if item['type'] == 'image_url' -%}"
         "{{ '<start_of_image>' }}"
         "{%- elif item['type'] == 'text' -%}"
         "{{ item['text'] | trim }}"
@@ -3604,21 +3604,6 @@ def split_text_on_image_urls(text: str, image_urls: List[str]):
                 remaining = ""
         return split_text
 
-    @staticmethod
-    def get_image_urls(messages: List[llama_types.ChatCompletionRequestMessage]):
-        image_urls: List[str] = []
-        for message in messages:
-            if message["role"] == "user":
-                if message.get("content") is None:
-                    continue
-                for content in message["content"]:
-                    if isinstance(content, dict) and content.get("type") == "image":
-                        if isinstance(content.get("image"), dict) and isinstance(content["image"].get("url"), str):
-                            image_urls.append(content["image"]["url"])
-                        elif isinstance(content.get("url"), str):
-                            image_urls.append(content["url"])
-        return image_urls
-
     def eval_image(self, llama: llama.Llama, image_url: str):
         import llama_cpp
 

llama_cpp/llava_cpp.py

@@ -245,7 +245,8 @@ def clip_image_batch_encode(
     ctx: clip_ctx_p,
     n_threads: c_int,
     imgs: "_Pointer[clip_image_f32_batch]",
-    vec: c_void_p
+    vec: c_void_p,
+    /,
 ) -> bool:
     ...