add vllm

app.py

@@ -38,11 +38,22 @@ with open(f'{model_path}/params.json', 'r') as f:
 with open(f'{model_path}/tekken.json', 'r') as f:
     tokenizer_config = json.load(f)
 
-llm = LLM(model=repo_id,
-          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})
+@spaces.GPU()
+def initialize_llm():
+    try:
+        llm = LLM(
+            model=repo_id,
+            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}
+        )
+        return llm
+    except Exception as e:
+        print("LLM initialization failed:", e)
+        return None
+
+llm = initialize_llm()
 
 def encode_image(image: Image.Image, image_format="PNG") -> str:
     im_file = BytesIO()