app.py

import gradio as gr
import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
import warnings

# disable some warnings
transformers.logging.set_verbosity_error()
transformers.logging.disable_progress_bar()
warnings.filterwarnings('ignore')

model_name = 'cognitivecomputations/dolphin-vision-72b'

# Set up GPU memory optimization
torch.cuda.empty_cache()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

# Load model with memory optimizations
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
    device_map="auto",
    trust_remote_code=True,
    offload_folder="offload",  # Offload to disk if necessary
    offload_state_dict=True,   # Offload state dict to CPU
    max_memory={0: "40GB"}     # Limit GPU memory usage
)

def inference(prompt, image, temperature, beam_size):
    messages = [
        {"role": "user", "content": f'<image>\n{prompt}'}
    ]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    text_chunks = [tokenizer(chunk).input_ids for chunk in text.split('<image>')]
    input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1], dtype=torch.long).unsqueeze(0).to(device)

    image_tensor = model.process_images([image], model.config).to(device)

    # Clear GPU memory
    torch.cuda.empty_cache()

    # Generate with memory optimization
    with torch.cuda.amp.autocast():
        output_ids = model.generate(
            input_ids,
            images=image_tensor,
            max_new_tokens=1024,
            temperature=temperature,
            num_beams=beam_size,
            use_cache=True,
            do_sample=True,
            repetition_penalty=1.1,
            length_penalty=1.0,
            no_repeat_ngram_size=3
        )[0]

    # Clear GPU memory again
    torch.cuda.empty_cache()

    return tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip()

# Create Gradio interface
with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(label="Prompt", placeholder="Describe this image in detail")
            image_input = gr.Image(label="Image", type="pil")
            temperature_input = gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.1, label="Temperature")
            beam_size_input = gr.Slider(minimum=1, maximum=10, value=4, step=1, label="Beam Size")
            submit_button = gr.Button("Submit")
        with gr.Column():
            output_text = gr.Textbox(label="Output")

    submit_button.click(
        fn=inference, 
        inputs=[prompt_input, image_input, temperature_input, beam_size_input], 
        outputs=output_text
    )

# Launch the app
demo.launch()