removing unnecessary cases from slice_scatter impl and adding test case

Author: apbose

py/torch_tensorrt/dynamo/lowering/_decompositions.py

@@ -183,17 +183,11 @@ def slice_scatter_decomposition(
         step = 1
 
     src_dim = src_tensor.shape
-    step_dim = (end - start) // step
-    end_dim = end
-    if step_dim > src_dim[dim]:
-        end_dim = src_dim[dim]
-    else:
-        # In this case src first step_dim need to be selected
-        indices = torch.arange(0, step_dim)
-        src = torch.index_select(src_tensor, dim, indices)
+    # step == 0 is not a valid torch case
+    # also src_dim should be equal to slice dimension
 
-    if start == 0 and end == dim_size and step == 0:
-        return input_tensor
+    if start == 0 and end == dim_size and step == 1:
+        return src_tensor
 
     cat_tensors = []
     index_tensor_shape = []

tests/py/dynamo/lowering/test_decompositions.py

@@ -547,6 +547,70 @@ def forward(self, x, src, dim, start=None, end=None, step=1):
             f"Slice_scatter TRT outputs don't match with the original model.",
         )
 
+    def test_lowering_slice_scatter_dimZero_StepTwo_module(self):
+        class sliceScatter(torch.nn.Module):
+            def __init__(self, *args, **kwargs) -> None:
+                super().__init__(*args, **kwargs)
+
+            def forward(self, x, src, dim, start, end, step):
+                y = torch.ops.aten.slice_scatter.default(x, src, dim, start, end, step)
+                return y
+
+        # Operations expected to be removed in the traced graph after decompositions
+        expected_ops = {
+            torch.ops.aten.index.Tensor,
+            torch.ops.aten.scatter.src,
+        }
+        unexpected_ops = {torch.ops.aten.slice_scatter}
+
+        inputs = [torch.zeros(8, 8).cuda(), torch.ones(2, 8).cuda(), 0, 2, 6, 2]
+
+        fx_graph = torch.fx.symbolic_trace(sliceScatter())
+
+        unexpected_ops_seen, expected_ops_unseen = lower_graph_testing(
+            fx_graph,
+            inputs,
+            expected_ops=expected_ops,
+            unexpected_ops=unexpected_ops,
+            min_block_size=1,
+        )
+
+        self.assertEquals(
+            len(unexpected_ops_seen),
+            0,
+            f"The following unexpected ops were encountered: {unexpected_ops_seen}",
+        )
+
+        self.assertEquals(
+            len(expected_ops_unseen),
+            0,
+            f"The following expected ops were not encountered: {expected_ops_unseen}",
+        )
+
+        torch._dynamo.reset()
+
+        # Validate that the results between Torch and Torch-TRT are similar
+        optimized_model = torch_tensorrt.compile(
+            fx_graph,
+            "torch_compile",
+            inputs,
+            min_block_size=1,
+            truncate_long_and_double=True,
+            pass_through_build_failures=True,
+        )
+        optimized_model_results = optimized_model(*inputs).detach().cpu()
+        torch_model_results = fx_graph(*inputs).detach().cpu()
+
+        max_diff = float(
+            torch.max(torch.abs(optimized_model_results - torch_model_results))
+        )
+        self.assertAlmostEqual(
+            max_diff,
+            0,
+            DECIMALS_OF_AGREEMENT,
+            f"Slice_scatter TRT outputs don't match with the original model.",
+        )
+
 
 if __name__ == "__main__":
     run_tests()