How to run agents from `smolagents` locally?

Inference Endpoints on the Hub
1

Is there any successful example of running CodeAgent or ToolCallingAgent entirely locally?
I couldn’t find anything in the course or in the forum. Please help!
If this is impossible, I would just like anybody from the HF’s people to assure that (for example, maybe the smolagents framework works explicitly with HF’s APIs). Thank you!

2

It seems ongoing issue. In short, it’s a bug.

3

I have been running everything locally by using either the TransfomerModel or MLXModel classes instead of the HfApiModel class.

from smolagents import CodeAgent, MLXModel

model = MLXModel(
    "mlx-community/Qwen2.5-Coder-32B-Instruct-4bit",
    {
        "temperature": 0.7,
        "top_k": 20,  
        "top_p": 0.8, 
        "min_p": 0.05,
        "num_ctx": 32768,
    },
)
agent = CodeAgent(tools=[], model=model, add_base_tools=True)
agent.run(
    "Could you give me the 40th number in the Fibonacci sequence?",
)

You can also use Ollama through the LiteLLMModel if you prepend “ollama_chat/” to the model name like such

from smolagents import CodeAgent, LiteLLMModel

model = LiteLLMModel(
    model_id="ollama_chat/qwen2.5-coder:14b-instruct-q4_K_M",
    api_base="http://localhost:11434",
    num_ctx=8192,
)

agent = CodeAgent(tools=[], model=model, add_base_tools=True)

agent.run(
    "Could you give me the 40th number in the Fibonacci sequence?",
)

Works like a charm!

4

Hi there!
I have been using the same setup with ollama and agree that this works like a charm. But i just cannot get this same setup to work with images.
Whenever i try to send images to a model like llava:7b ollama throws a 400 bad request error.
Any workaround for this problem?

5

It’s probably the first phenomenon.