slice scatter support for dynamic cases

Author: apbose

py/torch_tensorrt/dynamo/lowering/_decompositions.py

@@ -213,22 +213,26 @@ def slice_scatter_decomposition(
         return src_tensor
 
     # Ensure start, end, and step are all integers
-    assert isinstance(start, int), "start must be an integer"
-    assert isinstance(end, int), "end must be an integer"
-    assert isinstance(step, int), "step must be an integer"
-
-    cat_tensors = []
-    index_tensor_shape = []
-    for i, src_each_dim in enumerate(list(src_dim)):
-        if i != dim:
-            index_tensor_shape.append(src_each_dim)
-    for index in range(start, end, step):
-        cat_tensors.append(index * torch.ones(index_tensor_shape, dtype=torch.int64))
-    index_tensor = torch.stack(cat_tensors, dim)
-    index_tensor = index_tensor.to(device_input_tensor)
-    index_tensor_64 = index_tensor.to(torch.int64)
-    output_tensor = torch.scatter(input_tensor, dim, index_tensor_64, src_tensor)
-    return output_tensor
+    # Ensure start, end, and step are all integers
+    assert isinstance(start, (int, torch.SymInt)), "start must be an int or SymInt"
+    assert isinstance(end, (int, torch.SymInt)), "end must be an int or SymInt"
+    assert isinstance(step, (int, torch.SymInt)), "step must be an int or SymInt"
+
+    src_dim = src_tensor.shape
+    # 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 == 1:
+        return src_tensor
+
+    indices = torch.arange(
+        start, end, step, device=device_input_tensor, dtype=torch.int64
+    )
+    index_tensor = indices.view(
+        [-1 if i == dim else 1 for i in range(input_tensor.dim())]
+    )
+    index_tensor = index_tensor.expand_as(src_tensor)
+    return torch.scatter(input_tensor.clone(), dim, index_tensor, src_tensor)
 
 
 @register_torch_trt_decomposition(

tests/py/dynamo/lowering/test_decompositions.py

@@ -812,6 +812,79 @@ def forward(self, x, src, dim, start, end, step):
             f"Slice_scatter TRT outputs don't match with the original model.",
         )
 
+    def test_lowering_slice_scatter_dynamic_module(self):
+        class sliceScatter(torch.nn.Module):
+            def __init__(self, *args, **kwargs) -> None:
+                super().__init__(*args, **kwargs)
+
+            def forward(self, x, src, dim, start=None, end=None, step=1):
+                y = torch.ops.aten.slice_scatter(x, src, dim, start, end, step)
+                return y
+
+        # Operations expected to be removed in the traced graph after decompositions
+        expected_ops = {
+            torch.ops.aten.scatter.src,
+        }
+        unexpected_ops = {torch.ops.aten.select_scatter}
+
+        a = torch.zeros(8, 8).cuda()
+        b = torch.ones(8, 2).cuda()
+
+        # 0-D tensors for dynamic scalar values
+        start = torch.tensor(1, dtype=torch.int64).cuda()
+        end = torch.tensor(6, dtype=torch.int64).cuda()
+        step = torch.tensor(1, dtype=torch.int64).cuda()
+
+        # Mark scalar tensors as dynamic (note: shape = ())
+        torch._dynamo.mark_dynamic(start, (), min=1, max=3)
+        torch._dynamo.mark_dynamic(end, (), min=4, max=6)
+
+        inputs = (a, b, start, end, None, step)
+        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.assertEqual(
+            len(unexpected_ops_seen),
+            0,
+            f"The following unexpected ops were encountered: {unexpected_ops_seen}",
+        )
+
+        self.assertEqual(
+            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_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.",
+        )
+
     def test_lowering_select_scatter_dimZero_module(self):
         class selectScatter(torch.nn.Module):
             def __init__(self, *args, **kwargs) -> None: