import gradio as gr from transformers import AutoProcessor, Idefics3ForConditionalGeneration import re import time from PIL import Image import torch import spaces import subprocess subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True) processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM_converted_4") model = Idefics3ForConditionalGeneration.from_pretrained("HuggingFaceTB/SmolVLM_converted_4", torch_dtype=torch.bfloat16, _attn_implementation="flash_attention_2").to("cuda") @spaces.GPU def model_inference( images, text, assistant_prefix, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p ): if text == "" and not images: gr.Error("Please input a query and optionally image(s).") if text == "" and images: gr.Error("Please input a text query along the image(s).") if isinstance(images, Image.Image): images = [images] resulting_messages = [ { "role": "user", "content": [{"type": "image"}] + [ {"type": "text", "text": text} ] } ] if assistant_prefix: text = f"{assistant_prefix} {text}" prompt = processor.apply_chat_template(resulting_messages, add_generation_prompt=True) inputs = processor(text=prompt, images=[images], return_tensors="pt") inputs = {k: v.to("cuda") for k, v in inputs.items()} generation_args = { "max_new_tokens": max_new_tokens, "repetition_penalty": repetition_penalty, } assert decoding_strategy in [ "Greedy", "Top P Sampling", ] if decoding_strategy == "Greedy": generation_args["do_sample"] = False elif decoding_strategy == "Top P Sampling": generation_args["temperature"] = temperature generation_args["do_sample"] = True generation_args["top_p"] = top_p generation_args.update(inputs) # Generate generated_ids = model.generate(**generation_args) generated_texts = processor.batch_decode(generated_ids[:, generation_args["input_ids"].size(1):], skip_special_tokens=True) return generated_texts[0] with gr.Blocks(fill_height=True) as demo: gr.Markdown("## SmolVLM 🐶") gr.Markdown("Play with [HuggingFaceTB/SmolVLM](https://huggingface.co/HuggingFaceTB/SmolVLM) in this demo. To get started, upload an image and text or try one of the examples.") gr.Markdown("**Disclaimer:** SmolVLM does not include an RLHF alignment stage, so it may not consistently follow prompts or handle complex tasks. However, this doesn't mean it is incapable of doing so. Adding a prefix to the assistant's response, such as Let's think step for a reasoning question or `` for HTML code generation, can significantly improve the output in practice. You could also play with the parameters such as the temperature in non-greedy mode.") with gr.Column(): image_input = gr.Image(label="Upload your Image", type="pil", scale=1) query_input = gr.Textbox(label="Prompt") assistant_prefix = gr.Textbox(label="Assistant Prefix", placeholder="Let's think step by step.") submit_btn = gr.Button("Submit") output = gr.Textbox(label="Output") with gr.Accordion(label="Example Inputs and Advanced Generation Parameters"): examples=[ ["example_images/mmmu_example.jpeg", "Chase wants to buy 4 kilograms of oval beads and 5 kilograms of star-shaped beads. How much will he spend?", "Let's think step by step.", "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/rococo_1.jpg", "What art era is this?", None, "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/paper_with_text.png", "Read what's written on the paper", None, "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/dragons_playing.png","What's unusual about this image?",None, "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/example_images_ai2d_example_2.jpeg", "What happens to fish if pelicans increase?", None, "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/travel_tips.jpg", "I want to go somewhere similar to the one in the photo. Give me destinations and travel tips.", None, "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/dummy_pdf.png", "How much percent is the order status?", None, "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/art_critic.png", "As an art critic AI assistant, could you describe this painting in details and make a thorough critic?.",None, "Greedy", 0.4, 512, 1.2, 0.8], ["example_images/s2w_example.png", "What is this UI about?", None,"Greedy", 0.4, 512, 1.2, 0.8]] # Hyper-parameters for generation max_new_tokens = gr.Slider( minimum=8, maximum=1024, value=512, step=1, interactive=True, label="Maximum number of new tokens to generate", ) repetition_penalty = gr.Slider( minimum=0.01, maximum=5.0, value=1.2, step=0.01, interactive=True, label="Repetition penalty", info="1.0 is equivalent to no penalty", ) temperature = gr.Slider( minimum=0.0, maximum=5.0, value=0.4, step=0.1, interactive=True, label="Sampling temperature", info="Higher values will produce more diverse outputs.", ) top_p = gr.Slider( minimum=0.01, maximum=0.99, value=0.8, step=0.01, interactive=True, label="Top P", info="Higher values is equivalent to sampling more low-probability tokens.", ) decoding_strategy = gr.Radio( [ "Greedy", "Top P Sampling", ], value="Greedy", label="Decoding strategy", interactive=True, info="Higher values is equivalent to sampling more low-probability tokens.", ) decoding_strategy.change( fn=lambda selection: gr.Slider( visible=( selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"] ) ), inputs=decoding_strategy, outputs=temperature, ) decoding_strategy.change( fn=lambda selection: gr.Slider( visible=( selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"] ) ), inputs=decoding_strategy, outputs=repetition_penalty, ) decoding_strategy.change( fn=lambda selection: gr.Slider(visible=(selection in ["Top P Sampling"])), inputs=decoding_strategy, outputs=top_p, ) gr.Examples( examples = examples, inputs=[image_input, query_input, assistant_prefix, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p], outputs=output, fn=model_inference ) submit_btn.click(model_inference, inputs = [image_input, query_input, assistant_prefix, decoding_strategy, temperature, max_new_tokens, repetition_penalty, top_p], outputs=output) demo.launch(debug=True)