bfsdbv

app.py

@@ -33,6 +33,18 @@ id_model = Idefics3ForConditionalGeneration.from_pretrained("HuggingFaceM4/Idefi
 BAD_WORDS_IDS = id_processor.tokenizer(["<image>", "<fake_token_around_image>"], add_special_tokens=False).input_ids
 EOS_WORDS_IDS = [id_processor.tokenizer.eos_token_id]
 
+# Load colpali model
+model_name = "vidore/colpali-v1.2"
+token = os.environ.get("HF_TOKEN")
+model = ColPali.from_pretrained(
+    "vidore/colpaligemma-3b-pt-448-base", torch_dtype=torch.bfloat16, device_map="cuda", token = token).eval()
+
+model.load_adapter(model_name)
+model = model.eval()
+processor = AutoProcessor.from_pretrained(model_name, token = token)
+
+mock_image = Image.new("RGB", (448, 448), (255, 255, 255))
+
 @spaces.GPU
 def model_inference(
     images, text, assistant_prefix= None, decoding_strategy = "Greedy", temperature= 0.4, max_new_tokens=512,
@@ -91,17 +103,6 @@ def model_inference(
     generated_texts = id_processor.batch_decode(generated_ids[:, generation_args["input_ids"].size(1):], skip_special_tokens=True)
     return generated_texts[0]
 
-# Load model
-model_name = "vidore/colpali-v1.2"
-token = os.environ.get("HF_TOKEN")
-model = ColPali.from_pretrained(
-    "vidore/colpaligemma-3b-pt-448-base", torch_dtype=torch.bfloat16, device_map="cuda", token = token).eval()
-
-model.load_adapter(model_name)
-model = model.eval()
-processor = AutoProcessor.from_pretrained(model_name, token = token)
-
-mock_image = Image.new("RGB", (448, 448), (255, 255, 255))
 
 
 @spaces.GPU