Update app.py

app.py

@@ -1,8 +1,7 @@
 import torch
 import gradio as gr
-from transformers import AutoModelForCausalLM
+from transformers import AutoModelForCausalLM, BitsAndBytesConfig
 from deepseek_vl.models import VLChatProcessor
-from deepseek_vl.utils.io import load_pil_images
 
 # 模型路徑
 model_path = "deepseek-ai/deepseek-vl-7b-chat"
@@ -11,45 +10,55 @@ model_path = "deepseek-ai/deepseek-vl-7b-chat"
 vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path)
 tokenizer = vl_chat_processor.tokenizer
 
-# ==== 載入模型 (CPU/GPU 兼容，不使用量化) ====
+# ==== 量化模型設定 (4-bit) ====
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.float16,
+    bnb_4bit_use_double_quant=True
+)
+
 vl_gpt: AutoModelForCausalLM = AutoModelForCausalLM.from_pretrained(
     model_path,
     trust_remote_code=True,
-    device_map="auto",       # 自動分配 GPU / CPU
-    torch_dtype=torch.bfloat16  # 低精度減少 VRAM
+    device_map="auto",
+    quantization_config=bnb_config
 )
 vl_gpt.eval()
 
 # ==== 單張圖片處理 + 減少 max_new_tokens ====
 def generate_answer(image, text):
     try:
+        # 將圖片與文字組合成對話格式
         conversation = [
             {"role": "User", "content": "<image_placeholder>" + text, "images": [image]},
             {"role": "Assistant", "content": ""}
         ]
 
-        pil_images = load_pil_images(conversation)
+        # 🚨 這裡直接傳入 [image]，因為已經是 PIL.Image，不需要 load_pil_images
         prepare_inputs = vl_chat_processor(
             conversations=conversation,
-            images=pil_images,
+            images=[image],
             force_batchify=True
         ).to(vl_gpt.device)
 
+        # 轉換成 embeddings
         inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
 
+        # 產生回答
         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,   # 降低生成長度
+            max_new_tokens=128,   # 降低生成長度以減少記憶體
             do_sample=False,
             use_cache=True
         )
 
         answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=True)
         return f"{prepare_inputs['sft_format'][0]} {answer}"
+
     except Exception as e:
         return f"Error: {str(e)}"
 
@@ -59,7 +68,7 @@ demo = gr.Interface(
     inputs=[gr.Image(type="pil", label="Upload Image"), gr.Textbox(label="Question")],
     outputs="text",
     title="DeepSeek-VL-7B Chat Demo",
-    description="上傳圖片並輸入問題，模型會生成與圖片相關的回答（CPU/GPU 兼容，低記憶體模式）"
+    description="上傳圖片並輸入問題，模型會生成與圖片相關的回答（4-bit 量化，低記憶體模式）"
 )
 
 if __name__ == "__main__":