Load LLM before gradio

app.py

@@ -18,6 +18,12 @@ sampling_params = SamplingParams(max_tokens=8192, temperature=0.7)
 max_tokens_per_img = 4096
 max_img_per_msg = 5
 
+llm = LLM(model="mistralai/Pixtral-12B-2409",
+          tokenizer_mode="mistral",
+          max_model_len=65536,
+          max_num_batched_tokens=max_img_per_msg * max_tokens_per_img,
+          limit_mm_per_prompt={"image": max_img_per_msg})  # Name or path of your model
+
 def encode_image(image: Image.Image, image_format="PNG") -> str:
     im_file = BytesIO()
     image.save(im_file, format=image_format)
@@ -28,13 +34,6 @@ def encode_image(image: Image.Image, image_format="PNG") -> str:
 
 # @spaces.GPU #[uncomment to use ZeroGPU]
 def infer(image_url, prompt, progress=gr.Progress(track_tqdm=True)):
-    # tokenize image urls and text
-    llm = LLM(model="mistralai/Pixtral-12B-2409",
-              tokenizer_mode="mistral",
-              max_model_len=65536,
-              max_num_batched_tokens=max_img_per_msg * max_tokens_per_img,
-              limit_mm_per_prompt={"image": max_img_per_msg})  # Name or path of your model
-
     image = Image.open(BytesIO(requests.get(image_url).content))
     image = image.resize((3844, 2408))
     new_image_url = f"data:image/png;base64,{encode_image(image, image_format='PNG')}"