Revert "[inductor] add lowering for repeat_interleave.Tensor with output size specified (#147160) (#158462)"

test/inductor/test_torchinductor.py

@@ -13654,31 +13654,6 @@ def forward(self, x):
         inputs = (torch.randn(4, device=self.device),)
         self.common(Model(), inputs)
 
-    @parametrize("dtype", [torch.int32, torch.int64])
-    def test_repeat_interleave_Tensor_decomp(self, dtype):
-        # https://github.com/pytorch/pytorch/issues/147160
-        def f(input, repeats):
-            return torch.repeat_interleave(input, repeats, dim=0, output_size=3) + 1
-
-        input = torch.tensor([[1, 2], [3, 4]], dtype=dtype, device=self.device)
-        repeat = torch.tensor([1, 2], device=self.device)
-
-        if input.device.type == "mps" and dtype == torch.int64:
-            raise unittest.SkipTest(
-                "torch.compile fails this test with mps & int64, "
-                "see https://github.com/pytorch/pytorch/issues/159408"
-            )
-
-        f_compiled = torch.compile(f)
-        output, (code,) = run_and_get_code(f_compiled, input, repeat)
-        reference = f(input, repeat)
-        self.assertEqual(output, reference)
-        # we don't lower when the cpp_wrapper is used because it cannot generate
-        # proper examples during autotune
-        can_lower = (not config.cpp_wrapper) and (input.device.type != "mps")
-        has_lowered = not re.search(r"repeat_interleave.Tensor", code)
-        self.assertEqual(has_lowered, can_lower)
-
 
 @dataclasses.dataclass
 class TestFailure:

test/inductor/test_torchinductor_codegen_dynamic_shapes.py

@@ -347,7 +347,7 @@ def run(*ex, **kwargs):
     "test_rand_like_deterministic_dynamic_shapes": TestFailure(
         ("cpu", "cuda", "xpu"), is_skip=True
     ),
-    "test_repeat_interleave_2_dynamic_shapes": TestFailure(("cpu", "xpu")),
+    "test_repeat_interleave_2_dynamic_shapes": TestFailure(("cpu", "cuda", "xpu")),
     "test_slice_mutation2_dynamic_shapes": TestFailure(
         ("cpu", "cuda", "xpu"), is_skip=True
     ),

torch/_inductor/decomposition.py

@@ -1154,25 +1154,3 @@ def rrelu_with_noise_functional(
     else:
         negative_slope = (lower + upper) / 2
         return aten.leaky_relu(self, negative_slope), torch.Tensor()
-
-
-@register_decomposition(aten.repeat_interleave.Tensor)
-def repeat_interleave_Tensor(
-    repeat: torch.Tensor,
-    output_size: Optional[int] = None,
-) -> torch.Tensor:
-    if config.triton.autotune_at_compile_time:
-        # We can't compile-time auto-tune this because
-        # it expects specific data in `repeat`
-        return NotImplemented
-    if output_size is None or type(output_size) is not int:
-        return NotImplemented
-    if repeat.device.type == "mps":
-        return NotImplemented
-    assert repeat.dtype in [torch.int32, torch.int64]
-    assert repeat.ndim == 1
-    cumsum = repeat.cumsum(0)
-    pos = torch.arange(output_size, device=repeat.device)
-    return torch.searchsorted(
-        cumsum, pos, out_int32=(repeat.dtype == torch.int32), right=True
-    )

torch/_inductor/lowering.py

@@ -2879,7 +2879,6 @@ def is_aligned(x):
 
 # index_reduce requires fallback when use_scatter_fallback(...) returns True
 make_fallback(aten.index_reduce)
-make_fallback(aten.repeat_interleave.Tensor, override_decomp=True)
 
 
 # Register with type_promotion_kind None.