Ensure managed memory is supported in cudf.pandas.

python/cudf/cudf/pandas/__init__.py

@@ -5,6 +5,8 @@
 import os
 import warnings
 
+from cuda import cudart
+
 import rmm.mr
 
 from cudf._lib import pylibcudf
@@ -26,6 +28,27 @@
 }
 
 
+def _ensure_managed_memory_is_supported(rmm_mode):
+    # WSL does not support managed memory. We raise if the user requests
+    # managed memory on a system that does not support it.
+    if "managed" in rmm_mode:
+        # Ensure CUDA is initialized before checking cudaDevAttrConcurrentManagedAccess
+        cudart.cudaFree(0)
+
+        device_id = 0
+        err, supports_managed_access = cudart.cudaDeviceGetAttribute(
+            cudart.cudaDeviceAttr.cudaDevAttrConcurrentManagedAccess, device_id
+        )
+        if err != cudart.cudaError_t.cudaSuccess:
+            raise RuntimeError(
+                f"Failed to check cudaDevAttrConcurrentManagedAccess with error {err}"
+            )
+        if supports_managed_access == 0:
+            raise ValueError(
+                f"Managed memory is unsupported on this system, so {rmm_mode=} is not allowed."
+            )
+
+
 def _enable_managed_prefetching(rmm_mode):
     if "managed" in rmm_mode:
         for key in _SUPPORTED_PREFETCHES:
@@ -50,6 +73,8 @@ def install():
         )
         return
 
+    _ensure_managed_memory_is_supported(rmm_mode)
+
     free_memory, _ = rmm.mr.available_device_memory()
     free_memory = int(round(float(free_memory) * 0.80 / 256) * 256)
     new_mr = current_mr