Convert aten.alias to no-op

tests/lowering/misc/test_alias.py

@@ -0,0 +1,34 @@
+import torch
+import torch_ttnn
+import pytest
+
+
+class AliasModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return torch.ops.aten.alias.default(x)
+
+
+@pytest.mark.parametrize(
+    "input_shape",
+    [(4, 4)],
+)
+def test_alias(device, input_shape):
+    m = AliasModule()
+    input = torch.rand(input_shape, dtype=torch.bfloat16)
+    result_before = m.forward(input)
+
+    option = torch_ttnn.TorchTtnnOption(device=device, gen_graphviz=True)
+    # The compilation is lazy, so we need to run forward once to trigger the compilation
+    m = torch.compile(m, backend=torch_ttnn.backend, options=option)
+    result_after = m.forward(input)
+    option._out_fx_graphs[0].print_tabular()
+
+    # Check the graph has be rewritten
+    nodes = list(option._out_fx_graphs[0].nodes)
+    # There should be no op
+    assert [node.op for node in nodes].count("call_function") == 0
+    # Check inference result
+    assert torch.allclose(result_before, result_after)

torch_ttnn/passes/lowering/to_tt_pass.py

@@ -312,6 +312,10 @@ def call_function(self, target, args, kwargs):
                 return self.call_function_prop_meta(ttnn.squeeze, args, kwargs)
             return self.call_function_prop_meta(target, args, kwargs)
 
+        if target == torch.ops.aten.alias.default:
+            # alias is no-op
+            return args[0]
+
         return self.call_function_prop_meta(target, args, kwargs)