Transformers

You are viewing v4.44.2 version. A newer version v4.46.3 is available.

Join the Hugging Face community

and get access to the augmented documentation experience

Collaborate on models, datasets and Spaces

Faster examples with accelerated inference

Switch between documentation themes

to get started

Image-text-to-text

Image-text-to-text models, also known as vision language models (VLMs), are language models that take an image input. These models can tackle various tasks, from visual question answering to image segmentation. This task shares many similarities with image-to-text, but with some overlapping use cases like image captioning. Image-to-text models only take image inputs and often accomplish a specific task, whereas VLMs take open-ended text and image inputs and are more generalist models.

In this guide, we provide a brief overview of VLMs and show how to use them with Transformers for inference.

To begin with, there are multiple types of VLMs:

base models used for fine-tuning
chat fine-tuned models for conversation
instruction fine-tuned models

This guide focuses on inference with an instruction-tuned model.

Let’s begin installing the dependencies.

pip install -q transformers accelerate flash_attn

Let’s initialize the model and the processor.

from transformers import AutoProcessor, Idefics2ForConditionalGeneration
import torch

device = torch.device("cuda")
model = Idefics2ForConditionalGeneration.from_pretrained(
    "HuggingFaceM4/idefics2-8b",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
).to(device)

processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b")

This model has a chat template that helps user parse chat outputs. Moreover, the model can also accept multiple images as input in a single conversation or message. We will now prepare the inputs.

The image inputs look like the following.

from PIL import Image
import requests

img_urls =["https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png",
           "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"]
images = [Image.open(requests.get(img_urls[0], stream=True).raw),
          Image.open(requests.get(img_urls[1], stream=True).raw)]

Below is an example of the chat template. We can feed conversation turns and the last message as an input by appending it at the end of the template.

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What do we see in this image?"},
        ]
    },
    {
        "role": "assistant",
        "content": [
            {"type": "text", "text": "In this image we can see two cats on the nets."},
        ]
    },
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "And how about this image?"},
        ]
    },       
]

We will now call the processors’ apply_chat_template() method to preprocess its output along with the image inputs.

prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=prompt, images=[images[0], images[1]], return_tensors="pt").to(device)

We can now pass the preprocessed inputs to the model.

with torch.no_grad():
    generated_ids = model.generate(**inputs, max_new_tokens=500)
generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)

print(generated_texts)
## ['User: What do we see in this image? \nAssistant: In this image we can see two cats on the nets. \nUser: And how about this image? \nAssistant: In this image we can see flowers, plants and insect.']

Streaming

We can use text streaming for a better generation experience. Transformers supports streaming with the TextStreamer or TextIteratorStreamer classes. We will use the TextIteratorStreamer with IDEFICS-8B.

Assume we have an application that keeps chat history and takes in the new user input. We will preprocess the inputs as usual and initialize TextIteratorStreamer to handle the generation in a separate thread. This allows you to stream the generated text tokens in real-time. Any generation arguments can be passed to TextIteratorStreamer.

import time
from transformers import TextIteratorStreamer
from threading import Thread

def model_inference(
    user_prompt,
    chat_history,
    max_new_tokens,
    images
):
    user_prompt = {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": user_prompt},
        ]
    }
    chat_history.append(user_prompt)
    streamer = TextIteratorStreamer(
        processor.tokenizer,
        skip_prompt=True,
        timeout=5.0,
    )

    generation_args = {
        "max_new_tokens": max_new_tokens,
        "streamer": streamer,
        "do_sample": False
    }

    # add_generation_prompt=True makes model generate bot response
    prompt = processor.apply_chat_template(chat_history, add_generation_prompt=True)
    inputs = processor(
        text=prompt,
        images=images,
        return_tensors="pt",
    ).to(device)
    generation_args.update(inputs)

    thread = Thread(
        target=model.generate,
        kwargs=generation_args,
    )
    thread.start()

    acc_text = ""
    for text_token in streamer:
        time.sleep(0.04)
        acc_text += text_token
        if acc_text.endswith("<end_of_utterance>"):
            acc_text = acc_text[:-18]
        yield acc_text
    
    thread.join()

Now let’s call the model_inference function we created and stream the values.

generator = model_inference(
    user_prompt="And what is in this image?",
    chat_history=messages,
    max_new_tokens=100,
    images=images
)

for value in generator:
  print(value)

# In
# In this
# In this image ...

Fit models in smaller hardware

VLMs are often large and need to be optimized to fit in smaller hardware. Transformers supports many model quantization libraries, and here we will only show int8 quantization with Quanto. int8 quantization offers memory improvements up to 75 percent (if all weights are quantized). However it is no free lunch, since 8-bit is not a CUDA-native precision, the weights are quantized back and forth on the fly, which adds up to latency.

First, install dependencies.

pip install -U quanto bitsandbytes

To quantize a model during loading, we need to first create QuantoConfig. Then load the model as usual, but pass quantization_config during model initialization.

from transformers import Idefics2ForConditionalGeneration, AutoTokenizer, QuantoConfig

model_id = "HuggingFaceM4/idefics2-8b"
quantization_config = QuantoConfig(weights="int8")
quantized_model = Idefics2ForConditionalGeneration.from_pretrained(model_id, device_map="cuda", quantization_config=quantization_config)

And that’s it, we can use the model the same way with no changes.

Transformers

Image-text-to-text

Streaming

Fit models in smaller hardware

Further Reading