app.py

import gradio as gr
import spaces
from mistral_inference.transformer import Transformer
from mistral_inference.generate import generate
from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
from mistral_common.protocol.instruct.messages import UserMessage, TextChunk, ImageURLChunk
from mistral_common.protocol.instruct.request import ChatCompletionRequest
from huggingface_hub import snapshot_download
from pathlib import Path

# モデルのダウンロードと準備
mistral_models_path = Path.home().joinpath('mistral_models', 'Pixtral')
mistral_models_path.mkdir(parents=True, exist_ok=True)

snapshot_download(repo_id="mistral-community/pixtral-12b-240910", 
                  allow_patterns=["params.json", "consolidated.safetensors", "tekken.json"], 
                  local_dir=mistral_models_path)

# トークナイザーとモデルのロード
tokenizer = MistralTokenizer.from_file(f"{mistral_models_path}/tekken.json")
model = Transformer.from_folder(mistral_models_path)

# 推論処理
@spaces.GPU
def mistral_inference(prompt, image_url):
    completion_request = ChatCompletionRequest(
        messages=[UserMessage(content=[ImageURLChunk(image_url=image_url), TextChunk(text=prompt)])]
    )
    
    encoded = tokenizer.encode_chat_completion(completion_request)
    images = encoded.images
    tokens = encoded.tokens

    out_tokens, _ = generate([tokens], model, images=[images], max_tokens=1024, temperature=0.35, eos_id=tokenizer.instruct_tokenizer.tokenizer.eos_id)
    result = tokenizer.decode(out_tokens[0])
    
    return result

# Gradio インターフェース
def process_input(text, image_url):
    result = mistral_inference(text, image_url)
    return result

with gr.Blocks() as demo:
    gr.Markdown("## Pixtralモデルによる画像説明生成")
    
    with gr.Row():
        text_input = gr.Textbox(label="テキストプロンプト", placeholder="例: Describe the image.")
        image_input = gr.Textbox(label="画像URL", placeholder="例: https://example.com/image.png")
    
    result_output = gr.Textbox(label="モデルの出力結果")
    
    submit_button = gr.Button("推論を実行")
    
    submit_button.click(process_input, inputs=[text_input, image_input], outputs=result_output)

demo.launch()