feat: support aten._local_scalar_dense converter

[chohk88]

Description

A converter for the torch.ops.aten._local_scalar_dense operation, a low-level operation in PyTorch used internally to extract a scalar value from a tensor containing a single element.

To reviewer, I did not incorporate type-specific considerations in the implementation of the converter and test code. However, all test cases have successfully passed. I would greatly appreciate any feedback or suggestions regarding this aspect—especially if there are potential implications or improvements I might have overlooked.

Fixes # (issue)

Type of change

Please delete options that are not relevant and/or add your own.

• New feature (non-breaking change which adds functionality)

Checklist:

• My code follows the style guidelines of this project (You can use the linters)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas and hacks
• I have made corresponding changes to the documentation
• I have added tests to verify my fix or my feature
• New and existing unit tests pass locally with my changes
• I have added the relevant labels to my PR in so that relevant reviewers are notified

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 08:56:26.557741+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 08:58:19.570898+00:00
@@ -6,28 +6,16 @@


class TestLocalScalarDenseConverter(DispatchTestCase):
    @parameterized.expand(
        [
-            (
-                torch.randn((5, 10, 5), dtype=torch.float32),
-            ),
-            (
-                torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),
-            ),
-            (
-                torch.randn((1), dtype=torch.float32),
-            ),
-            (
-                (torch.tensor([-2.4])),
-            ),
-            (
-                (torch.tensor([5.5, 3.5, 3.6])),
-            ),
-            (
-                (torch.tensor([True])),
-            ),
+            (torch.randn((5, 10, 5), dtype=torch.float32),),
+            (torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),),
+            (torch.randn((1), dtype=torch.float32),),
+            ((torch.tensor([-2.4])),),
+            ((torch.tensor([5.5, 3.5, 3.6])),),
+            ((torch.tensor([True])),),
            (
                torch.tensor(
                    [
                        float("nan"),
                        1.23,
@@ -42,13 +30,11 @@
                        1.23,
                        float("nan"),
                    ]
                ),
            ),
-            (
-                (torch.tensor([float("inf")])),
-            ),
+            ((torch.tensor([float("inf")])),),
        ]
    )
    def test_local_scalar_dense(self, data):
        class local_scalar_dense(nn.Module):
            def forward(self, input):

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 08:56:26.539822+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 08:58:21.048206+00:00
@@ -6,28 +6,16 @@


class TestLocalScalarDenseConverter(DispatchTestCase):
    @parameterized.expand(
        [
-            (
-                torch.randn((5, 10, 5), dtype=torch.float32),
-            ),
-            (
-                torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),
-            ),
-            (
-                torch.randn((1), dtype=torch.float32),
-            ),
-            (
-                (torch.tensor([-2.4])),
-            ),
-            (
-                (torch.tensor([5.5, 3.5, 3.6])),
-            ),
-            (
-                (torch.tensor([True])),
-            ),
+            (torch.randn((5, 10, 5), dtype=torch.float32),),
+            (torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),),
+            (torch.randn((1), dtype=torch.float32),),
+            ((torch.tensor([-2.4])),),
+            ((torch.tensor([5.5, 3.5, 3.6])),),
+            ((torch.tensor([True])),),
            (
                torch.tensor(
                    [
                        float("nan"),
                        1.23,
@@ -42,13 +30,11 @@
                        1.23,
                        float("nan"),
                    ]
                ),
            ),
-            (
-                (torch.tensor([float("inf")])),
-            ),
+            ((torch.tensor([float("inf")])),),
        ]
    )
    def test_local_scalar_dense(self, data):
        class local_scalar_dense(nn.Module):
            def forward(self, input):

[chohk88]

I would appreciate your feedback on the type-related issues mentioned in the description.

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 08:56:28.395579+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 08:59:57.272490+00:00
@@ -6,28 +6,16 @@


class TestLocalScalarDenseConverter(DispatchTestCase):
    @parameterized.expand(
        [
-            (
-                torch.randn((5, 10, 5), dtype=torch.float32),
-            ),
-            (
-                torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),
-            ),
-            (
-                torch.randn((1), dtype=torch.float32),
-            ),
-            (
-                (torch.tensor([-2.4])),
-            ),
-            (
-                (torch.tensor([5.5, 3.5, 3.6])),
-            ),
-            (
-                (torch.tensor([True])),
-            ),
+            (torch.randn((5, 10, 5), dtype=torch.float32),),
+            (torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),),
+            (torch.randn((1), dtype=torch.float32),),
+            ((torch.tensor([-2.4])),),
+            ((torch.tensor([5.5, 3.5, 3.6])),),
+            ((torch.tensor([True])),),
            (
                torch.tensor(
                    [
                        float("nan"),
                        1.23,
@@ -42,13 +30,11 @@
                        1.23,
                        float("nan"),
                    ]
                ),
            ),
-            (
-                (torch.tensor([float("inf")])),
-            ),
+            ((torch.tensor([float("inf")])),),
        ]
    )
    def test_local_scalar_dense(self, data):
        class local_scalar_dense(nn.Module):
            def forward(self, input):

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 08:56:26.670833+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_local_scalar_dense_aten.py	2024-04-16 09:00:15.174874+00:00
@@ -6,28 +6,16 @@


class TestLocalScalarDenseConverter(DispatchTestCase):
    @parameterized.expand(
        [
-            (
-                torch.randn((5, 10, 5), dtype=torch.float32),
-            ),
-            (
-                torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),
-            ),
-            (
-                torch.randn((1), dtype=torch.float32),
-            ),
-            (
-                (torch.tensor([-2.4])),
-            ),
-            (
-                (torch.tensor([5.5, 3.5, 3.6])),
-            ),
-            (
-                (torch.tensor([True])),
-            ),
+            (torch.randn((5, 10, 5), dtype=torch.float32),),
+            (torch.randint(-10, 10, (5, 1, 15), dtype=torch.int32),),
+            (torch.randn((1), dtype=torch.float32),),
+            ((torch.tensor([-2.4])),),
+            ((torch.tensor([5.5, 3.5, 3.6])),),
+            ((torch.tensor([True])),),
            (
                torch.tensor(
                    [
                        float("nan"),
                        1.23,
@@ -42,13 +30,11 @@
                        1.23,
                        float("nan"),
                    ]
                ),
            ),
-            (
-                (torch.tensor([float("inf")])),
-            ),
+            ((torch.tensor([float("inf")])),),
        ]
    )
    def test_local_scalar_dense(self, data):
        class local_scalar_dense(nn.Module):
            def forward(self, input):

[zewenli98]

@chohk88 @narendasan I checked the aten schema where there are only two funcs return Scalar type, i.e.,

- func: item(Tensor self) -> Scalar
- func: _local_scalar_dense(Tensor self) -> Scalar

I found some people have similar issues like: pytorch/xla#909
Their cases are moving data from XLA to CPU, so I'm thinking if this op is only called when moving data to CPUs.

For the functionality, my understanding to this function is to extract the first element in the input, whatever it is a single tensor or list of tensors, and then move it to CPUs. If true, I don't think it's meaningful to implement this converter. Instead, we can have a lowering pass to remove this op to avoid potential graph breaks.

Do you know why this converter is needed? or is this op used in any models?

[narendasan]

Needs #2810 to be implemented to fully work