app.py

import torch
import gradio as gr
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from deepseek_vl.models import VLChatProcessor, MultiModalityCausalLM

# 模型路徑
model_path = "deepseek-ai/deepseek-vl-7b-chat"

# ==== BitsAndBytes 4-bit 量化設定 ====
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,  # 強制 float16
    bnb_4bit_use_double_quant=True
)

# 載入 processor 和 tokenizer
vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer

# 載入模型 (4-bit 量化 + float16)
vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(
    model_path,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
).eval()

# ==== 單張圖片推理函式 ====
def chat_with_image(image, user_message):
    try:
        # 建立對話
        conversation = [
            {"role": "User", "content": "<image_placeholder>" + user_message, "images": [image]},
            {"role": "Assistant", "content": ""}
        ]

        # 直接傳入 PIL.Image，不使用 load_pil_images
        prepare_inputs = vl_chat_processor(
            conversations=conversation,
            images=[image],
            force_batchify=True
        ).to(vl_gpt.device)

        # 🚨 將 BatchedVLChatProcessorOutput 轉 dict
        prepare_inputs = {k: getattr(prepare_inputs, k) for k in prepare_inputs.__dataclass_fields__.keys()}

        # 正確 dtype：input_ids/labels 保持 long，其他 tensor 轉 float16
        new_inputs = {}
        for k, v in prepare_inputs.items():
            if torch.is_tensor(v):
                if k in ["input_ids", "labels"]:
                    new_inputs[k] = v.to(torch.long)
                else:
                    new_inputs[k] = v.to(torch.float16)
            else:
                new_inputs[k] = v
        prepare_inputs = new_inputs

        # 取得 embeddings
        inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)

        # 生成回答
        outputs = vl_gpt.language_model.generate(
            inputs_embeds=inputs_embeds,
            attention_mask=prepare_inputs["attention_mask"],
            pad_token_id=tokenizer.eos_token_id,
            bos_token_id=tokenizer.bos_token_id,
            eos_token_id=tokenizer.eos_token_id,
            max_new_tokens=128,  # 減少記憶體
            do_sample=False,
            use_cache=True
        )

        # 解碼
        answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=True)
        return answer

    except Exception as e:
        return f"Error: {str(e)}"

# ==== Gradio Web UI ====
demo = gr.Interface(
    fn=chat_with_image,
    inputs=[gr.Image(type="pil", label="Upload Image"),
            gr.Textbox(lines=2, placeholder="Ask about the image...")],
    outputs="text",
    title="DeepSeek-VL-7B-Chat Demo (4-bit, float16)",
    description="上傳圖片並輸入問題，模型會生成與圖片相關的回答"
)

if __name__ == "__main__":
    demo.launch()