[Core] Support load and unload LoRA in api server (#6566)

Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>

vllm/utils.py

@@ -1224,3 +1224,28 @@ async def _run_task_with_lock(task: Callable, lock: asyncio.Lock, *args,
 def supports_dynamo() -> bool:
     base_torch_version = Version(Version(torch.__version__).base_version)
     return base_torch_version >= Version("2.4.0")
+
+
+class AtomicCounter:
+    """An atomic, thread-safe counter"""
+
+    def __init__(self, initial=0):
+        """Initialize a new atomic counter to given initial value"""
+        self._value = initial
+        self._lock = threading.Lock()
+
+    def inc(self, num=1):
+        """Atomically increment the counter by num and return the new value"""
+        with self._lock:
+            self._value += num
+            return self._value
+
+    def dec(self, num=1):
+        """Atomically decrement the counter by num and return the new value"""
+        with self._lock:
+            self._value -= num
+            return self._value
+
+    @property
+    def value(self):
+        return self._value