Fix torch's convert_to_tensor not respecting dtype when input is a Variable (#21452)

* Fix torch's `convert_to_tensor` not respecting `dtype` when input is a `Variable`.

* Fix openvino backend.

* Revert the casting and use variables for the tests.

Author: james77777778

keras/src/backend/openvino/core.py

@@ -600,18 +600,17 @@ def convert_to_tensor(x, dtype=None, sparse=None, ragged=None):
         raise ValueError("`sparse=True` is not supported with openvino backend")
     if ragged:
         raise ValueError("`ragged=True` is not supported with openvino backend")
+    if dtype is not None:
+        dtype = standardize_dtype(dtype)
     if isinstance(x, OpenVINOKerasTensor):
         return x
     elif isinstance(x, np.ndarray):
         if dtype is not None:
-            dtype = standardize_dtype(dtype)
             ov_type = OPENVINO_DTYPES[dtype]
             return OpenVINOKerasTensor(ov_opset.constant(x, ov_type).output(0))
         return OpenVINOKerasTensor(ov_opset.constant(x).output(0))
     elif isinstance(x, (list, tuple)):
-        if dtype is not None:
-            dtype = standardize_dtype(dtype)
-        else:
+        if dtype is None:
             # try to properly deduce element type
             elem = _get_first_element(x)
             if isinstance(elem, float):
@@ -624,12 +623,11 @@ def convert_to_tensor(x, dtype=None, sparse=None, ragged=None):
         dtype = standardize_dtype(dtype)
         ov_type = OPENVINO_DTYPES[dtype]
         return OpenVINOKerasTensor(ov_opset.constant(x, ov_type).output(0), x)
-    if dtype is not None:
-        dtype = standardize_dtype(dtype)
     if isinstance(x, Variable):
+        x = x.value
         if dtype and dtype != x.dtype:
-            return x.value.astype(dtype)
-        return x.value
+            x = cast(x, dtype)
+        return x
     if not is_tensor(x) and standardize_dtype(dtype) == "bfloat16":
         return ov.Tensor(np.asarray(x).astype(dtype))
     if dtype is None:

keras/src/backend/torch/core.py

@@ -192,10 +192,10 @@ def convert_to_tensor(x, dtype=None, sparse=None, ragged=None):
     if ragged:
         raise ValueError("`ragged=True` is not supported with torch backend")
     if isinstance(x, Variable):
-        # TorchDynamo has bugs supporting nn.Parameter type check.
-        # Return it directly instead of pass it to the rest of the logic in the
-        # function.
-        return x.value
+        if dtype is None:
+            return x.value
+        x = x.value
+        return x.to(to_torch_dtype(dtype))
     if is_tensor(x):
         device = get_device()
         if x.device != device:

keras/src/ops/core_test.py

@@ -1150,6 +1150,11 @@ def test_convert_to_tensor(self, x, dtype, expected_dtype):
                 ops.convert_to_tensor(x, dtype=dtype), expected_dtype
             )
 
+    @parameterized.named_parameters(named_product(dtype=ALL_DTYPES))
+    def test_convert_to_tensor_with_variable(self, dtype):
+        x = backend.Variable(np.array([1.0, 0.0, 1.0], dtype=np.float32))
+        self.assertDType(ops.convert_to_tensor(x, dtype=dtype), dtype)
+
     @parameterized.named_parameters(named_product(dtype=ALL_DTYPES))
     def test_saturate_cast(self, dtype):
         x = np.ones((1,))

keras/src/ops/nn_test.py

@@ -3137,7 +3137,7 @@ def test_dot_product_attention(self, dtype):
     def test_rms_normalization(self, dtypes):
         input_dtype, weight_dtype = dtypes
         inputs = knp.ones((2, 8), dtype=input_dtype)
-        scale = knp.ones((8,), dtype=weight_dtype)
+        scale = backend.Variable(knp.ones((8,), dtype=weight_dtype))
         expected_dtype = input_dtype
 
         self.assertDType(knn.rms_normalization(inputs, scale), expected_dtype)
@@ -3151,8 +3151,8 @@ def test_rms_normalization(self, dtypes):
     def test_layer_normalization(self, dtypes):
         input_dtype, weight_dtype = dtypes
         inputs = knp.ones((2, 8), dtype=input_dtype)
-        gamma = knp.ones((8,), dtype=weight_dtype)
-        beta = knp.ones((8,), dtype=weight_dtype)
+        gamma = backend.Variable(knp.ones((8,), dtype=weight_dtype))
+        beta = backend.Variable(knp.ones((8,), dtype=weight_dtype))
         expected_dtype = input_dtype
 
         self.assertDType(