[air] Suppress "NumPy array is not writable" error in torch conversion (

#29808)

When we convert NumPy arrays to torch tesnors in Ray Data, we run into a verbose Numpy error:

    The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior.

Since we don't write to numpy tensors we suppress this warning in the conversion. The alternative of copying the array would duplicate our memory usage.

Signed-off-by: Kai Fricke <[email protected]>

python/ray/air/_internal/torch_utils.py

@@ -1,3 +1,4 @@
+import warnings
 from typing import Dict, List, Optional, Union
 
 import numpy as np
@@ -132,7 +133,15 @@ def convert_ndarray_to_torch_tensor(
     Returns: A Torch Tensor.
     """
     ndarray = _unwrap_ndarray_object_type_if_needed(ndarray)
-    return torch.as_tensor(ndarray, dtype=dtype, device=device)
+
+    # The numpy array is not always writeable as it can come from the Ray object store.
+    # Numpy will throw a verbose warning here, which we suppress, as we don't write
+    # to the tensors. We also don't want to copy the array to avoid memory overhead.
+    # Original warning: https://github.com/pytorch/pytorch/blob/v1.13.0/
+    # torch/csrc/utils/tensor_numpy.cpp#L198-L206
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        return torch.as_tensor(ndarray, dtype=dtype, device=device)
 
 
 def convert_ndarray_batch_to_torch_tensor_batch(

python/ray/air/tests/test_data_batch_conversion.py

@@ -4,6 +4,7 @@
 import numpy as np
 import pyarrow as pa
 
+from ray.air._internal.torch_utils import convert_ndarray_to_torch_tensor
 from ray.air.constants import TENSOR_COLUMN_NAME
 from ray.air.util.data_batch_conversion import (
     convert_batch_type_to_pandas,
@@ -185,6 +186,17 @@ def test_numpy_object_pandas():
     )
 
 
+@pytest.mark.parametrize("writable", [False, True])
+def test_numpy_to_tensor_warning(writable):
+    input_data = np.array([[1, 2, 3]], dtype=int)
+    input_data.setflags(write=writable)
+
+    with pytest.warns(None) as record:
+        tensor = convert_ndarray_to_torch_tensor(input_data)
+    assert not record.list, [w.message for w in record.list]
+    assert tensor is not None
+
+
 def test_dict_fail():
     input_data = {"x": "y"}
     with pytest.raises(ValueError):