Fix #2541: Add support for multimodal content format in qwen2.5-vl model

Co-Authored-By: Joe Moura <joao@crewai.com>

src/crewai/llm.py

@@ -839,9 +839,13 @@ class LLM(BaseLLM):
 
         # Validate message format first
         for msg in messages:
-            if not isinstance(msg, dict) or "role" not in msg or "content" not in msg:
+            if not isinstance(msg, dict) or "role" not in msg:
                 raise TypeError(
-                    "Invalid message format. Each message must be a dict with 'role' and 'content' keys"
+                    "Invalid message format. Each message must be a dict with 'role' key"
+                )
+            if "content" not in msg and msg["role"] != "system":
+                raise TypeError(
+                    "Invalid message format. Each non-system message must have a 'content' key"
                 )
 
         # Handle O1 models specially
@@ -868,6 +872,19 @@ class LLM(BaseLLM):
                 messages.append({"role": "user", "content": "Please continue."})
             return messages
 
+        if "qwen" in self.model.lower():
+            formatted_messages = []
+            for msg in messages:
+                if not isinstance(msg.get("content"), str):
+                    formatted_messages.append(msg)
+                    continue
+                
+                formatted_messages.append({
+                    "role": msg["role"],
+                    "content": [{"type": "text", "text": msg["content"]}]
+                })
+            return formatted_messages
+
         # Handle Anthropic models
         if not self.is_anthropic:
             return messages

tests/test_qwen_multimodal.py

@@ -0,0 +1,31 @@
+import pytest
+from crewai.llm import LLM
+
+
+@pytest.mark.vcr(filter_headers=["authorization"])
+def test_qwen_multimodal_content_formatting():
+    """Test that multimodal content is properly formatted for Qwen models."""
+    
+    llm = LLM(model="sambanova/Qwen2.5-72B-Instruct", temperature=0.7)
+    
+    message = {"role": "user", "content": "Describe this image"}
+    formatted = llm._format_messages_for_provider([message])
+    assert isinstance(formatted[0]["content"], list)
+    assert formatted[0]["content"][0]["type"] == "text"
+    assert formatted[0]["content"][0]["text"] == "Describe this image"
+    
+    multimodal_content = [
+        {"type": "text", "text": "What's in this image?"}, 
+        {"type": "image_url", "image_url": "https://example.com/image.jpg"}
+    ]
+    message = {"role": "user", "content": multimodal_content}
+    formatted = llm._format_messages_for_provider([message])
+    assert formatted[0]["content"] == multimodal_content
+    
+    messages = [
+        {"role": "system", "content": "You are a visual analysis assistant."},
+        {"role": "user", "content": multimodal_content}
+    ]
+    formatted = llm._format_messages_for_provider(messages)
+    assert isinstance(formatted[0]["content"], list)
+    assert formatted[1]["content"] == multimodal_content