Update app.py

app.py

@@ -1,6 +1,6 @@
 import streamlit as st
 import torch
-from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer
+from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
 from PIL import Image
 from byaldi import RAGMultiModalModel
 from qwen_vl_utils import process_vision_info
@@ -8,24 +8,24 @@ from qwen_vl_utils import process_vision_info
 # Model and processor names
 RAG_MODEL = "vidore/colpali"
 QWN_MODEL = "Qwen/Qwen2-VL-7B-Instruct"
-QWN_PROCESSOR = "Qwen/Qwen2-VL-2B-Instruct"
 
 @st.cache_resource
 def load_models():
     RAG = RAGMultiModalModel.from_pretrained(RAG_MODEL)
     
-    model = AutoModelForCausalLM.from_pretrained(
+    model = Qwen2VLForConditionalGeneration.from_pretrained(
         QWN_MODEL,
         torch_dtype=torch.bfloat16,
+        attn_implementation="flash_attention_2",
+        device_map="auto",
         trust_remote_code=True
-    ).cuda().eval()
+    ).eval()
     
-    processor = AutoProcessor.from_pretrained(QWN_PROCESSOR, trust_remote_code=True)
-    tokenizer = AutoTokenizer.from_pretrained(QWN_PROCESSOR, trust_remote_code=True)
+    processor = AutoProcessor.from_pretrained(QWN_MODEL, trust_remote_code=True)
     
-    return RAG, model, processor, tokenizer
+    return RAG, model, processor
 
-RAG, model, processor, tokenizer = load_models()
+RAG, model, processor = load_models()
 
 def document_rag(text_query, image):
     messages = [
@@ -40,7 +40,7 @@ def document_rag(text_query, image):
             ],
         }
     ]
-    text = tokenizer.apply_chat_template(
+    text = processor.apply_chat_template(
         messages, tokenize=False, add_generation_prompt=True
     )
     image_inputs, video_inputs = process_vision_info(messages)
@@ -51,12 +51,12 @@ def document_rag(text_query, image):
         padding=True,
         return_tensors="pt",
     )
-    inputs = inputs.to("cuda")
+    inputs = inputs.to(model.device)
     generated_ids = model.generate(**inputs, max_new_tokens=50)
     generated_ids_trimmed = [
         out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
     ]
-    output_text = tokenizer.batch_decode(
+    output_text = processor.batch_decode(
         generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
     )
     return output_text[0]