feature: Torch dependency in sagameker-core to be made optional (5457) (aws#5713)

* feature: Torch dependency in sagameker-core to be made optional (5457)

* fix: address review comments (iteration #1)

* fix: address review comments (iteration #1)

* fix: use subprocess instead of importlib.reload to avoid breaking six.with_metaclass super()

Author: aviruthen

sagemaker-core/pyproject.toml

@@ -32,7 +32,6 @@ dependencies = [
     "smdebug_rulesconfig>=1.0.1",
     "schema>=0.7.5",
     "omegaconf>=2.1.0",
-    "torch>=1.9.0",
     "scipy>=1.5.0",
     # Remote function dependencies
     "cloudpickle>=2.0.0",
@@ -52,6 +51,12 @@ classifiers = [
 ]
 
 [project.optional-dependencies]
+torch = [
+    "torch>=1.9.0",
+]
+all = [
+    "sagemaker-core[torch]",
+]
 codegen = [
     "black>=24.3.0, <25.0.0",
     "pandas>=2.0.0, <3.0.0",
@@ -62,6 +67,7 @@ test = [
     "pytest>=8.0.0, <9.0.0",
     "pytest-cov>=4.0.0",
     "pytest-xdist>=3.0.0",
+    "sagemaker-core[torch]",
 ]
 
 [project.urls]

sagemaker-core/src/sagemaker/core/deserializers/base.py

@@ -365,8 +365,11 @@ def __init__(self, accept="tensor/pt"):
             from torch import from_numpy
 
             self.convert_npy_to_tensor = from_numpy
-        except ImportError:
-            raise Exception("Unable to import pytorch.")
+        except ImportError as e:
+            raise ImportError(
+                "Unable to import torch. Please install torch to use TorchTensorDeserializer: "
+                "pip install 'sagemaker-core[torch]'"
+            ) from e
 
     def deserialize(self, stream, content_type="tensor/pt"):
         """Deserialize streamed data to TorchTensor

sagemaker-core/src/sagemaker/core/serializers/base.py

@@ -443,9 +443,16 @@ class TorchTensorSerializer(SimpleBaseSerializer):
 
     def __init__(self, content_type="tensor/pt"):
         super(TorchTensorSerializer, self).__init__(content_type=content_type)
-        from torch import Tensor
+        try:
+            from torch import Tensor
+
+            self.torch_tensor = Tensor
+        except ImportError as e:
+            raise ImportError(
+                "Unable to import torch. Please install torch to use TorchTensorSerializer: "
+                "pip install 'sagemaker-core[torch]'"
+            ) from e
 
-        self.torch_tensor = Tensor
         self.numpy_serializer = NumpySerializer()
 
     def serialize(self, data):

sagemaker-core/tests/unit/test_optional_torch_dependency.py

@@ -0,0 +1,188 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+"""Tests to verify torch dependency is optional in sagemaker-core.
+
+The "module imports without torch" tests use subprocess instead of
+importlib.reload to avoid poisoning the class hierarchy in the current
+process.  six.with_metaclass + old-style super() breaks when a module
+is reloaded because the class identity changes, causing
+``TypeError: super(type, obj): obj must be an instance or subtype of type``
+in subsequent tests that instantiate serializers/deserializers.
+"""
+from __future__ import absolute_import
+
+import io
+import subprocess
+import sys
+import textwrap
+
+import numpy as np
+import pytest
+
+
+def _block_torch():
+    """Block torch imports by setting sys.modules['torch'] to None.
+
+    Returns a dict of saved torch submodule entries so they can be restored.
+    """
+    torch_keys = [key for key in sys.modules if key.startswith("torch.")]
+    saved = {key: sys.modules.pop(key) for key in torch_keys}
+    saved["torch"] = sys.modules.get("torch")
+    sys.modules["torch"] = None
+    return saved
+
+
+def _restore_torch(saved):
+    """Restore torch modules from saved dict."""
+    original_torch = saved.pop("torch", None)
+    if original_torch is not None:
+        sys.modules["torch"] = original_torch
+    elif "torch" in sys.modules:
+        del sys.modules["torch"]
+    for key, val in saved.items():
+        sys.modules[key] = val
+
+
+def test_serializer_module_imports_without_torch():
+    """Verify that non-torch serializers can be imported and instantiated without torch.
+
+    Runs in a subprocess to avoid polluting the current process's class
+    hierarchy via importlib.reload (which breaks six.with_metaclass).
+    """
+    code = textwrap.dedent("""\
+        import sys
+        # Block torch before any sagemaker imports
+        sys.modules["torch"] = None
+
+        from sagemaker.core.serializers.base import (
+            CSVSerializer,
+            NumpySerializer,
+            JSONSerializer,
+            IdentitySerializer,
+        )
+
+        assert CSVSerializer() is not None
+        assert NumpySerializer() is not None
+        assert JSONSerializer() is not None
+        assert IdentitySerializer() is not None
+        print("OK")
+    """)
+    result = subprocess.run(
+        [sys.executable, "-c", code],
+        capture_output=True,
+        text=True,
+    )
+    assert result.returncode == 0, (
+        f"Subprocess failed:\nstdout: {result.stdout}\nstderr: {result.stderr}"
+    )
+
+
+def test_deserializer_module_imports_without_torch():
+    """Verify that non-torch deserializers can be imported and instantiated without torch.
+
+    Runs in a subprocess for the same reason as the serializer test above.
+    """
+    code = textwrap.dedent("""\
+        import sys
+        sys.modules["torch"] = None
+
+        from sagemaker.core.deserializers.base import (
+            StringDeserializer,
+            BytesDeserializer,
+            CSVDeserializer,
+            NumpyDeserializer,
+            JSONDeserializer,
+        )
+
+        assert StringDeserializer() is not None
+        assert BytesDeserializer() is not None
+        assert CSVDeserializer() is not None
+        assert NumpyDeserializer() is not None
+        assert JSONDeserializer() is not None
+        print("OK")
+    """)
+    result = subprocess.run(
+        [sys.executable, "-c", code],
+        capture_output=True,
+        text=True,
+    )
+    assert result.returncode == 0, (
+        f"Subprocess failed:\nstdout: {result.stdout}\nstderr: {result.stderr}"
+    )
+
+
+def test_torch_tensor_serializer_raises_import_error_without_torch():
+    """Verify TorchTensorSerializer raises ImportError when torch is not installed."""
+    import sagemaker.core.serializers.base as ser_module
+
+    saved = {}
+    try:
+        saved = _block_torch()
+
+        with pytest.raises(ImportError, match="Unable to import torch"):
+            ser_module.TorchTensorSerializer()
+    finally:
+        _restore_torch(saved)
+
+
+def test_torch_tensor_deserializer_raises_import_error_without_torch():
+    """Verify TorchTensorDeserializer raises ImportError when torch is not installed."""
+    import sagemaker.core.deserializers.base as deser_module
+
+    saved = {}
+    try:
+        saved = _block_torch()
+
+        with pytest.raises(ImportError, match="Unable to import torch"):
+            deser_module.TorchTensorDeserializer()
+    finally:
+        _restore_torch(saved)
+
+
+def test_torch_tensor_serializer_works_with_torch():
+    """Verify TorchTensorSerializer works when torch is available."""
+    try:
+        import torch
+    except ImportError:
+        pytest.skip("torch is not installed")
+
+    from sagemaker.core.serializers.base import TorchTensorSerializer
+
+    serializer = TorchTensorSerializer()
+    tensor = torch.tensor([1.0, 2.0, 3.0])
+    result = serializer.serialize(tensor)
+    assert result is not None
+    # Verify the result can be loaded back as numpy
+    array = np.load(io.BytesIO(result))
+    assert np.array_equal(array, np.array([1.0, 2.0, 3.0]))
+
+
+def test_torch_tensor_deserializer_works_with_torch():
+    """Verify TorchTensorDeserializer works when torch is available."""
+    try:
+        import torch
+    except ImportError:
+        pytest.skip("torch is not installed")
+
+    from sagemaker.core.deserializers.base import TorchTensorDeserializer
+
+    deserializer = TorchTensorDeserializer()
+    # Create a numpy array, save it, and deserialize to tensor
+    array = np.array([1.0, 2.0, 3.0])
+    buffer = io.BytesIO()
+    np.save(buffer, array)
+    buffer.seek(0)
+
+    result = deserializer.deserialize(buffer, "tensor/pt")
+    assert isinstance(result, torch.Tensor)
+    assert torch.equal(result, torch.tensor([1.0, 2.0, 3.0]))

sagemaker-core/tests/unit/test_serializer_implementations.py

@@ -162,3 +162,25 @@ def test_numpy_serializer_import(self):
     def test_record_serializer_deprecated(self):
         """Test that numpy_to_record_serializer is available as deprecated."""
         assert hasattr(implementations, "numpy_to_record_serializer")
+
+
+class TestTorchSerializerWithOptionalDependency:
+    """Test torch serializer/deserializer with optional torch dependency."""
+
+    def test_torch_tensor_serializer_instantiation(self):
+        """Test that TorchTensorSerializer can be instantiated when torch is available."""
+        torch = pytest.importorskip("torch")
+        from sagemaker.core.serializers.base import TorchTensorSerializer
+
+        serializer = TorchTensorSerializer()
+        assert serializer is not None
+        assert serializer.content_type == "tensor/pt"
+
+    def test_torch_tensor_deserializer_instantiation(self):
+        """Test that TorchTensorDeserializer can be instantiated when torch is available."""
+        torch = pytest.importorskip("torch")
+        from sagemaker.core.deserializers.base import TorchTensorDeserializer
+
+        deserializer = TorchTensorDeserializer()
+        assert deserializer is not None
+        assert deserializer.accept == "tensor/pt"