ENH: cuDF 25.2 compatibility (#853)

Author: hainaweiben

.github/workflows/python.yaml

@@ -290,7 +290,7 @@ jobs:
       if: ${{ matrix.module == 'gpu' }}
       run: |
         source activate ${{ env.CONDA_ENV }}
-        pip install --extra-index-url=https://pypi.nvidia.com cudf-cu12==24.10.*
+        pip install --extra-index-url=https://pypi.nvidia.com cudf-cu12==25.2.*
         pip install ucxx-cu12 cython numpy cloudpickle scikit-learn \
           pyyaml psutil tornado sqlalchemy defusedxml tqdm uvloop coverage \
           pytest pytest-cov pytest-timeout pytest-forked pytest-asyncio pytest-mock

python/xorbits/_mars/deploy/oscar/tests/test_local.py

@@ -1144,7 +1144,7 @@ def test_naive_code_file():
             raise
 
 
-ucp = lazy_import("ucp")
+ucp = lazy_import("ucxx")
 _OSCAR_CONF_TEMPLATE = """
 "@inherits": '@default'
 oscar:

python/xorbits/_mars/lib/nvutils.py

@@ -685,7 +685,41 @@ def _running_process_matches(handle: _nvmlDevice_t) -> bool:
     out : bool
         Whether the device handle has a CUDA context on the running process.
     """
-    return any(os.getpid() == o.pid for o in get_compute_running_processes(handle))
+    current_pid = os.getpid()
+    try:
+        _nvml_check_error(_nvml_lib.nvmlInit_v2())
+        procs = get_compute_running_processes(handle)
+        if any(p.pid == current_pid for p in procs):
+            return True
+    except NVMLAPIError:
+        pass
+    finally:
+        try:
+            _nvml_lib.nvmlShutdown()
+        except Exception:
+            pass
+
+    try:
+        import cupy
+
+        cupy.cuda.runtime.deviceSynchronize()
+        return True
+    except ImportError:
+        pass
+    except Exception:
+        pass
+
+    try:
+        out = subprocess.check_output(
+            ["nvidia-smi", "--query-compute-apps=pid", "--format=csv,noheader"],
+            text=True,
+            stderr=subprocess.DEVNULL,
+        )
+        return str(current_pid) in out.splitlines()
+    except Exception:
+        pass
+
+    return False
 
 
 def get_cuda_context() -> CudaContext:

python/xorbits/_mars/tensor/array_utils.py

@@ -86,8 +86,12 @@ def _get(x):
 
 
 def move_to_device(x, device_id):
-    if hasattr(x, "device") and x.device.id == device_id:
-        return x
+    if hasattr(x, "device"):
+        if isinstance(x.device, str):
+            if x.device == str(device_id):
+                return x
+        elif x.device.id == device_id:
+            return x
 
     assert device_id >= 0
 
@@ -97,7 +101,12 @@ def move_to_device(x, device_id):
     # for dense array, we currently copy from gpu to memory and then copy back to destination device
     # to avoid kernel panic
     with cp.cuda.Device(device_id):
-        return cp.asarray(cp.asnumpy(x))  # remove `cp.asnumpy` call to do directly copy
+        if isinstance(x, np.ndarray):
+            return cp.asarray(x)
+        else:
+            return cp.asarray(
+                cp.asnumpy(x)
+            )  # remove `cp.asnumpy` call to do directly copy
 
 
 def convert_order(x, order):

python/xorbits/_mars/utils.py

@@ -504,7 +504,11 @@ def calc_data_size(dt: Any, shape: Tuple[int] = None) -> int:
     if hasattr(dt, "shape") and len(dt.shape) == 0:
         return 0
     if hasattr(dt, "dtypes") and shape is not None:
-        size = shape[0] * sum(_get_dtype_itemsize(dtype) for dtype in dt.dtypes)
+        dtypes = dt.dtypes
+        try:
+            size = shape[0] * sum(_get_dtype_itemsize(dtype) for dtype in dtypes)
+        except TypeError:
+            size = shape[0] * _get_dtype_itemsize(dtypes)
         try:
             index_value_value = dt.index_value.value
             if hasattr(index_value_value, "dtype") and not isinstance(
@@ -552,7 +556,10 @@ def _is_fast_dtype(dtype):
 
     # handling possible MultiIndex
     if hasattr(index_obj, "dtypes"):
-        dtypes.extend(index_obj.dtypes)
+        try:
+            dtypes.extend(index_obj.dtypes)
+        except TypeError:
+            dtypes.append(index_obj.dtypes)
     else:
         dtypes.append(index_obj.dtype)