Update inference.py (#6)

- Update inference.py (6b651309a79be38f32bf79c9135c5bcba8ab2ff4)


Co-authored-by: GRINDA AI <FINGU-AI@users.noreply.huggingface.co>

inference.py

@@ -1,9 +1,19 @@
 import os
 import json
+import subprocess
+import sys
 import torch
+from typing import List, Dict
+
+# Ensure vllm is installed
+try:
+    import vllm
+except ImportError:
+    subprocess.check_call([sys.executable, "-m", "pip", "install", "vllm"])
+
+# Import the necessary modules after installation
 from vllm import LLM, SamplingParams
 from vllm.utils import random_uuid
-from typing import List, Dict
 
 # Function to format chat messages using Qwen's chat template
 def format_chat(messages: List[Dict[str, str]]) -> str: