TransformersModel auto-detects VLMs (#378)

* TransformersModel auto-detects VLMs

examples/agent_from_any_llm.py

@@ -9,7 +9,7 @@ from smolagents.agents import CodeAgent, ToolCallingAgent
 available_inferences = ["hf_api", "transformers", "ollama", "litellm"]
 chosen_inference = "transformers"
 
-print(f"Chose model {chosen_inference}")
+print(f"Chose model: '{chosen_inference}'")
 
 if chosen_inference == "hf_api":
     model = HfApiModel(model_id="meta-llama/Llama-3.3-70B-Instruct")

src/smolagents/agents.py

@@ -798,7 +798,7 @@ class ToolCallingAgent(MultiStepAgent):
             tool_arguments = tool_call.function.arguments
 
         except Exception as e:
-            raise AgentGenerationError(f"Error in generating tool call with model:\n{e}", self.logger)
+            raise AgentGenerationError(f"Error in generating tool call with model:\n{e}", self.logger) from e
 
         log_entry.tool_calls = [ToolCall(name=tool_name, arguments=tool_arguments, id=tool_call_id)]
 

src/smolagents/models.py

@@ -418,9 +418,6 @@ class TransformersModel(Model):
             The torch_dtype to initialize your model with.
         trust_remote_code (bool, default `False`):
             Some models on the Hub require running remote code: for this model, you would have to set this flag to True.
-        flatten_messages_as_text (`bool`, default `True`):
-            Whether to flatten messages as text: this must be sent to False to use VLMs (as opposed to LLMs for which this flag can be ignored).
-            Caution: this parameter is experimental and will be removed in an upcoming PR as we auto-detect VLMs.
         kwargs (dict, *optional*):
             Any additional keyword arguments that you want to use in model.generate(), for instance `max_new_tokens` or `device`.
         **kwargs:
@@ -449,7 +446,6 @@ class TransformersModel(Model):
         device_map: Optional[str] = None,
         torch_dtype: Optional[str] = None,
         trust_remote_code: bool = False,
-        flatten_messages_as_text: bool = True,
         **kwargs,
     ):
         super().__init__(**kwargs)
@@ -469,6 +465,7 @@ class TransformersModel(Model):
         if device_map is None:
             device_map = "cuda" if torch.cuda.is_available() else "cpu"
         logger.info(f"Using device: {device_map}")
+        self._is_vlm = False
         try:
             self.model = AutoModelForCausalLM.from_pretrained(
                 model_id,
@@ -481,6 +478,7 @@ class TransformersModel(Model):
             if "Unrecognized configuration class" in str(e):
                 self.model = AutoModelForImageTextToText.from_pretrained(model_id, device_map=device_map)
                 self.processor = AutoProcessor.from_pretrained(model_id)
+                self._is_vlm = True
             else:
                 raise e
         except Exception as e:
@@ -490,7 +488,6 @@ class TransformersModel(Model):
             self.model_id = default_model_id
             self.tokenizer = AutoTokenizer.from_pretrained(default_model_id)
             self.model = AutoModelForCausalLM.from_pretrained(model_id, device_map=device_map, torch_dtype=torch_dtype)
-        self.flatten_messages_as_text = flatten_messages_as_text
 
     def make_stopping_criteria(self, stop_sequences: List[str], tokenizer) -> "StoppingCriteriaList":
         from transformers import StoppingCriteria, StoppingCriteriaList
@@ -526,8 +523,7 @@ class TransformersModel(Model):
             messages=messages,
             stop_sequences=stop_sequences,
             grammar=grammar,
-            tools_to_call_from=tools_to_call_from,
+            flatten_messages_as_text=(not self._is_vlm),
-            flatten_messages_as_text=self.flatten_messages_as_text,
             **kwargs,
         )
 
@@ -595,9 +591,19 @@ class TransformersModel(Model):
         else:
             if "Action:" in output:
                 output = output.split("Action:", 1)[1].strip()
-            parsed_output = json.loads(output)
+            try:
-            tool_name = parsed_output.get("tool_name")
+                start_index = output.index("{")
-            tool_arguments = parsed_output.get("tool_arguments")
+                end_index = output.rindex("}")
+                output = output[start_index : end_index + 1]
+            except Exception as e:
+                raise Exception("No json blob found in output!") from e
+
+            try:
+                parsed_output = json.loads(output)
+            except json.JSONDecodeError as e:
+                raise ValueError(f"Tool call '{output}' has an invalid JSON structure: {e}")
+            tool_name = parsed_output.get("name")
+            tool_arguments = parsed_output.get("arguments")
             return ChatMessage(
                 role="assistant",
                 content="",

tests/test_agents.py

@@ -28,11 +28,7 @@ from smolagents.agents import (
     ToolCallingAgent,
 )
 from smolagents.default_tools import PythonInterpreterTool
-from smolagents.models import (
+from smolagents.models import ChatMessage, ChatMessageToolCall, ChatMessageToolCallDefinition, TransformersModel
-    ChatMessage,
-    ChatMessageToolCall,
-    ChatMessageToolCallDefinition,
-)
 from smolagents.tools import tool
 from smolagents.types import AgentImage, AgentText
 from smolagents.utils import BASE_BUILTIN_MODULES
@@ -620,3 +616,26 @@ nested_answer()
 
         output = agent.run("Count to 3")
         assert output == "Correct!"
+
+    def test_transformers_toolcalling_agent(self):
+        @tool
+        def get_weather(location: str, celsius: bool = False) -> str:
+            """
+            Get weather in the next days at given location.
+            Secretly this tool does not care about the location, it hates the weather everywhere.
+
+            Args:
+                location: the location
+                celsius: the temperature type
+            """
+            return "The weather is UNGODLY with torrential rains and temperatures below -10°C"
+
+        model = TransformersModel(
+            model_id="HuggingFaceTB/SmolLM2-360M-Instruct",
+            max_new_tokens=100,
+            device_map="auto",
+            do_sample=False,
+        )
+        agent = ToolCallingAgent(model=model, tools=[get_weather], max_steps=1)
+        agent.run("What's the weather in Paris?")
+        assert agent.logs[2].tool_calls[0].name == "get_weather"

tests/test_models.py

@@ -57,7 +57,6 @@ class ModelTests(unittest.TestCase):
             max_new_tokens=5,
             device_map="auto",
             do_sample=False,
-            flatten_messages_as_text=True,
         )
         messages = [{"role": "user", "content": [{"type": "text", "text": "Hello!"}]}]
         output = model(messages, stop_sequences=["great"]).content
@@ -72,7 +71,6 @@ class ModelTests(unittest.TestCase):
             max_new_tokens=5,
             device_map="auto",
             do_sample=False,
-            flatten_messages_as_text=False,
         )
         messages = [{"role": "user", "content": [{"type": "text", "text": "Hello!"}, {"type": "image", "image": img}]}]
         output = model(messages, stop_sequences=["great"]).content