app.py

import gradio as gr
import torch
import re
import random
import requests
from transformers import AutoModelForCausalLM, AutoTokenizer
from torchvision import models, transforms
from PIL import Image

# ==========================================
# 1. ЗАГРУЗКА ЗРЕНИЯ (~20MB)
# ==========================================
vision_model = models.squeezenet1_1(weights=models.SqueezeNet1_1_Weights.IMAGENET1K_V1).eval()

LABELS_URL = "https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt"
labels = requests.get(LABELS_URL).text.splitlines()

preprocess = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

# ==========================================
# 2. ТВОИ МОЗГИ (Лимит 340MB)
# ==========================================
MODEL_PATH = "./" 
TOKENIZER_NAME = "sberbank-ai/rugpt3small_based_on_gpt2"

tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_NAME)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    dtype=torch.float32, 
    tie_word_embeddings=False
).to("cpu")

model.config.max_position_embeddings = 128

# ==========================================
# 3. ЛОГИКА
# ==========================================
def predict(image, message, history):
    vision_info = "ничего не вижу"
    
    if image is not None:
        try:
            pil_img = Image.fromarray(image.astype('uint8'), 'RGB')
            input_tensor = preprocess(pil_img).unsqueeze(0)
            with torch.no_grad():
                output = vision_model(input_tensor)
            _, index = torch.max(output, 1)
            detected = labels[index[0]].replace("_", " ")
            vision_info = f"вижу {detected}"
        except Exception:
            vision_info = "туман"

    # Собираем промпт
    prompt = f"User: ({vision_info}) {message}\nBot:"
    
    inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
    curr_len = inputs.input_ids.shape[1]
    
    # Лимит до 128 токенов
    max_to_gen = 128 - curr_len - 1
    if max_to_gen <= 5:
        history.append({"role": "user", "content": message})
        history.append({"role": "assistant", "content": "Слишком длинно, не влезаю в 128!"})
        return history

    with torch.no_grad():
        output_tokens = model.generate(
            **inputs,
            max_new_tokens=max_to_gen,
            do_sample=True,
            temperature=0.35,
            repetition_penalty=1.8,
            pad_token_id=tokenizer.pad_token_id
        )

    answer = tokenizer.decode(output_tokens[0][curr_len:], skip_special_tokens=True).strip()
    answer = re.split(r'User:|Bot:|\n', answer)[0].strip()

    if not answer: answer = "..."

    # В Gradio 6.0 возвращаем обновленный список сообщений
    history.append({"role": "user", "content": message})
    history.append({"role": "assistant", "content": answer})
    return history

# ==========================================
# 4. ИНТЕРФЕЙС (GRADIO 6.0)
# ==========================================
with gr.Blocks() as demo:
    gr.Markdown("# 🍌 BananaVision Lite")
    with gr.Row():
        img_input = gr.Image(label="Глаза")
        chatbot = gr.Chatbot(label="Чат") # БЕЗ type="messages"
    
    msg = gr.Textbox(placeholder="Чё там на картинке?")
    btn = gr.Button("Спросить")

    btn.click(predict, [img_input, msg, chatbot], [chatbot])
    msg.submit(predict, [img_input, msg, chatbot], [chatbot])

# Тема передается здесь
demo.launch(theme=gr.themes.Default(primary_hue="yellow"))