Passing an `image_url` to a text-only model should fail explicitly

[Wauplin]

(noticed this error while working on https://github.com/huggingface/huggingface_hub/pull/2556)

System Info

Using TGI through Inference API (e.g. mistralai/Mistral-Nemo-Instruct-2407). At the time I open this issue /info returns

{
"model_id": "mistralai/Mistral-Nemo-Instruct-2407",
"model_sha": "e17a136e1dcba9c63ad771f2c85c1c312c563e6b",
"model_pipeline_tag": "text-generation",
"max_concurrent_requests": 128,
"max_best_of": 2,
"max_stop_sequences": 4,
"max_input_tokens": 16000,
"max_total_tokens": 32768,
"validation_workers": 2,
"max_client_batch_size": 4,
"router": "text-generation-router",
"version": "2.2.1-dev0",
"sha": "a0b6a2434503afa5da5f25fa47a3e4589c80941c",
"docker_label": "sha-a0b6a24"
}

Information

• [ ] Docker
• [ ] The CLI directly

Tasks

• [ ] An officially supported command
• [ ] My own modifications

Reproduction

Send a request to a text-only model with a payload containing a text + image content. Here is a curl command to reproduce it. It sends an image as image_url and "Describe this image in one sentence." as text.

curl -X POST \
  -H 'Content-Type: application/json' \
  -H 'authorization: Bearer <HF TOKEN>' \
  -d '{
    "model": "mistralai/Mistral-Nemo-Instruct-2407",
    "messages": [
      {
        "role": "user",
        "content": [
          {"type": "image_url", "image_url": {"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"}},
          {"type": "text", "text": "Describe this image in one sentence."}
        ]
      }
    ]
  }' \
  https://api-inference.huggingface.co/models/mistralai/Mistral-Nemo-Instruct-2407/v1/chat/completions
{"object":"chat.completion","id":"","created":1727279137,"model":"mistralai/Mistral-Nemo-Instruct-2407","system_fingerprint":"2.2.1-dev0-sha-a0b6a24","choices":[{"index":0,"message":{"role":"assistant","content":"The Statue of Liberty stands tall and proud on its pedestal, facing the New York Bay."},"logprobs":null,"finish_reason":"stop"}],"usage":{"prompt_tokens":57,"completion_tokens":19,"total_tokens":76}}

Expected behavior

Currently TGI returns successfully with the sentence "The Statue of Liberty stands tall and proud on its pedestal, facing the New York Bay.". It seems that since the model is not capable of handling the image, the image url is directly passed to the model. Since the url contains Statue-of-Liberty-Island-New-York-Bay.jpg, the answer looks correct but is not generated from the image itself.

{"object":"chat.completion","id":"","created":1727279137,"model":"mistralai/Mistral-Nemo-Instruct-2407","system_fingerprint":"2.2.1-dev0-sha-a0b6a24","choices":[{"index":0,"message":{"role":"assistant","content":"The Statue of Liberty stands tall and proud on its pedestal, facing the New York Bay."},"logprobs":null,"finish_reason":"stop"}],"usage":{"prompt_tokens":57,"completion_tokens":19,"total_tokens":76}}% 

In such a case I would expect either a 400 Bad request or a 422 Unprocessable entity.

I also tried with a base64-encoded URL and the model fails (max tokens exceeded) since the full base64 encoding seems to be tokenized.