#8364: Disable fallback for ttnn.repeat_interleave

tenstorrent · May 22, 2024 · 6a1209e · 6a1209e
1 parent 9c0ecf3
commit 6a1209e
Show file tree

Hide file tree

Showing 4 changed files with 14 additions and 7 deletions.
diff --git a/models/experimental/functional_mistral/tt/ttnn_functional_attention.py b/models/experimental/functional_mistral/tt/ttnn_functional_attention.py
@@ -64,8 +64,8 @@ def repeat_kv(key, values, repeats, device):
     values = ttnn.to_layout(
         ttnn.to_device(ttnn.from_torch(values, dtype=ttnn.bfloat16), device), layout=ttnn.TILE_LAYOUT
     )
-    keys = ttnn.repeat_interleave(keys, repeats, dim)
-    values = ttnn.repeat_interleave(values, repeats, dim)
+    keys = ttnn.get_fallback_function(ttnn.repeat_interleave)(keys, repeats, dim)
+    values = ttnn.get_fallback_function(ttnn.repeat_interleave)(values, repeats, dim)
     return keys, values
 
 

diff --git a/tests/ttnn/unit_tests/operations/test_repeat_interleave.py b/tests/ttnn/unit_tests/operations/test_repeat_interleave.py
@@ -11,6 +11,7 @@
 from tests.ttnn.utils_for_testing import assert_with_pcc
 
 
+@pytest.mark.skip(reason="ttnn.repeat_interleave only supports repeat over dim 0 or 1")
 def test_repeat_interleave(device):
     torch_input_tensor = torch.tensor([[1, 2], [3, 4]])
     torch_result = torch.repeat_interleave(torch_input_tensor, 2, dim=0)
@@ -23,6 +24,7 @@ def test_repeat_interleave(device):
     assert_with_pcc(torch_result, output, 0.9999)
 
 
+@pytest.mark.skip(reason="ttnn.repeat_interleave only supports repeat over dim 0 or 1")
 def test_repeat_interleave_with_repeat_tensor(device):
     torch_input_tensor = torch.tensor([[1, 2], [3, 4]], dtype=torch.bfloat16)
     torch_repeats = torch.tensor([1, 2])

diff --git a/ttnn/ttnn/decorators.py b/ttnn/ttnn/decorators.py
@@ -757,11 +757,13 @@ def operation_decorator(function: callable):
         global OPERATION_TO_FALLBACK_FUNCTION
 
         def fallback_function(*function_args, **function_kwargs):
-            updated_function_args, updated_function_kwargs = preprocess_golden_function_inputs(
-                function_args, function_kwargs
-            )
+            preprocess_inputs = preprocess_golden_function_inputs or default_preprocess_golden_function_inputs
+            postprocess_outputs = postprocess_golden_function_outputs or default_postprocess_golden_function_outputs
+
+            updated_function_args, updated_function_kwargs = preprocess_inputs(function_args, function_kwargs)
             output = golden_function(*updated_function_args, **updated_function_kwargs)
-            output = postprocess_golden_function_outputs(output, function_args, function_kwargs)
+            output = postprocess_outputs(output, function_args, function_kwargs)
+
             return output
 
         if ttnn.CONFIG.enable_fast_runtime_mode:

diff --git a/ttnn/ttnn/operations/data_movement.py b/ttnn/ttnn/operations/data_movement.py
@@ -299,11 +299,14 @@ def _repeat_interleave_validate_input_tensors(operation_name, input_tensor, *arg
     )
 
 
+# This operation does not support the following cases:
+#   - Shape([2[32], 2[32]]) -> repeats = 2, dim = 0
+#   - Shape([2[32], 2[32]]) -> repeats = Tensor[1,2], dim = 1
 @ttnn.register_operation(
     name="ttnn.repeat_interleave",
     validate_input_tensors=_repeat_interleave_validate_input_tensors,
     golden_function=_golden_function,
-    allow_to_fallback_to_golden_function_on_failure=True,
+    allow_to_fallback_to_golden_function_on_failure=False,
 )
 def repeat_interleave(input_tensor: ttnn.Tensor, repeats: Union[ttnn.Tensor, int], dim: int = 0) -> ttnn.Tensor:
     r"""