[HiCache] Check in-flight async ops in is_fully_idle() before attach/detach (#20746)

python/sglang/test/test_utils.py

@@ -167,6 +167,25 @@ def download_image_with_retry(image_url: str, max_retries: int = 3) -> Image.Ima
             time.sleep(2**i)
 
 
+def flush_cache_with_retry(
+    base_url: str, retries: int = 5, interval: float = 2.0
+) -> bool:
+    """Flush device cache with retry.
+
+    The scheduler may still have in-flight HiCache async ops (GPU↔Host↔L3)
+    that prevent is_fully_idle() from returning True, so we retry.
+    """
+    for _ in range(retries):
+        try:
+            response = requests.post(f"{base_url}/flush_cache", timeout=10)
+            if response.status_code == 200:
+                return True
+        except requests.RequestException:
+            pass
+        time.sleep(interval)
+    return False
+
+
 def is_in_ci():
     """Return whether it is in CI runner."""
     return get_bool_env_var("SGLANG_IS_IN_CI")