[aotd] capture rrelu_with_noise noise mutation in compile (#141867)

Rebase-copy of long standing already approved PR https://github.com/pytorch/pytorch/pull/138503 that was blocked on landing by xla build issues.

Got a new  PR with the same content (ghstack checkout was failing due to changed submodules)

Corresponding xla PR:
https://github.com/pytorch/xla/pull/8363

Pull Request resolved: https://github.com/pytorch/pytorch/pull/141867
Approved by: https://github.com/bdhirsh

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-2ec22641e390cda25ec7c61fcbce07507727d584
+e0684027996c60dcbb99fddf205385c208fb9ed7

aten/src/ATen/functorch/BatchRulesBinaryOps.cpp

@@ -280,6 +280,44 @@ static void fill__Tensor_batch_rule(
   std::get<0>(self_and_other).copy_(std::get<1>(self_and_other));
 }
 
+static
+std::tuple<Tensor, std::optional<int64_t>,Tensor, std::optional<int64_t>>
+rrelu_with_noise_batch_rule(
+    const Tensor& self,
+    std::optional<int64_t> self_bdim,
+    Tensor& noise,
+    std::optional<int64_t> noise_bdim,
+    const at::Scalar& lower,
+    const at::Scalar& upper,
+    bool training,
+    std::optional<at::Generator> generator) {
+
+  auto self_ = moveBatchDimToFront(self, self_bdim);
+  auto noise_ = moveBatchDimToFront(self, noise_bdim);
+
+  auto ret = at::rrelu_with_noise(self_, noise_, lower, upper, training, std::move(generator));
+
+  return std::make_tuple(ret, 0, noise_, 0);
+}
+
+static Tensor rrelu_with_noise_batch(
+    const Tensor& self,
+    Tensor& noise,
+    const Scalar& lower,
+    const Scalar& upper,
+    bool training,
+    std::optional<Generator> generator) {
+  c10::impl::ExcludeDispatchKeyGuard guard(DispatchKey::FuncTorchBatched);
+  auto maybe_layer = maybeCurrentDynamicLayer();
+  vmap_check_escaped(maybe_layer, "gen_vmap_plumbing");
+  int64_t cur_level = maybe_layer->layerId();
+  auto [self_value, self_bdim] = unwrapTensorAtLevel(self, cur_level);
+  auto [noise_value, noise_bdim] = unwrapTensorAtLevel(noise, cur_level);
+  TORCH_CHECK(!noise_bdim.has_value(), "vmap: Attempted to vmap over 'noise' in torch.rrelu_with_noise. This is not supported.");
+  auto res = rrelu_with_noise_batch_rule(self_value, self_bdim, noise_value, noise_bdim, lower, upper, training, std::move(generator));
+  return makeBatched(std::get<0>(res), std::get<1>(res), cur_level);
+}
+
 static std::tuple<Tensor, std::optional<int64_t>> log_sigmoid_backward_batch_rule(
   Tensor& grad, std::optional<int64_t> grad_bdim,
   Tensor& self, std::optional<int64_t> self_bdim,
@@ -421,7 +459,6 @@ TORCH_LIBRARY_IMPL(aten, FuncTorchBatched, m) {
   POINTWISE_BOXED(polygamma);
   BINARY_SCALAR_2(sub, Tensor, Scalar);
   BINARY_SCALAR_3(remainder, Tensor, Scalar, Scalar_Tensor);
-  BINARY_POINTWISE(rrelu_with_noise);
   BINARY_SCALAR_2(rsub, Tensor, Scalar);
 
   BINARY_SCALAR_3_Tensor(special_xlog1py, other_scalar, self_scalar);
@@ -509,6 +546,7 @@ TORCH_LIBRARY_IMPL(aten, FuncTorchBatched, m) {
   VMAP_SUPPORT(masked_select_backward, masked_select_backward_batch_rule);
 
   VMAP_SUPPORT2(fill_, Tensor, fill__Tensor_batch_rule);
+  m.impl("rrelu_with_noise", rrelu_with_noise_batch);
 }
 
 } // namespace at::functorch

aten/src/ATen/native/Activation.cpp

@@ -576,7 +576,7 @@ template <typename scalar_t>
 inline void _rrelu_with_noise_train(
     Tensor& output,
     const Tensor& input,
-    const Tensor& noise,
+    Tensor& noise,
     const Scalar& lower_,
     const Scalar& upper_,
     std::optional<Generator> generator) {
@@ -606,7 +606,7 @@ inline void _rrelu_with_noise_train(
 }
 
 Tensor& rrelu_with_noise_out_cpu(const Tensor& self,
-    const Tensor& noise,
+    Tensor& noise,
     const Scalar& lower,
     const Scalar& upper,
     bool training,
@@ -629,7 +629,7 @@ Tensor& rrelu_with_noise_out_cpu(const Tensor& self,
 
 Tensor rrelu_with_noise_cpu(
     const Tensor& self,
-    const Tensor& noise,
+    Tensor& noise,
     const Scalar& lower,
     const Scalar& upper,
     bool training,
@@ -641,7 +641,7 @@ Tensor rrelu_with_noise_cpu(
 
 Tensor& rrelu_with_noise_cpu_(
     Tensor& self,
-    const Tensor& noise,
+    Tensor& noise,
     const Scalar& lower,
     const Scalar& upper,
     bool training,
@@ -670,12 +670,14 @@ Tensor rrelu_with_noise_backward(
 
 Tensor rrelu(const Tensor & self, const Scalar& lower, const Scalar& upper, bool training, std::optional<Generator> generator) {
   TORCH_CHECK(lower.to<double>() <= upper.to<double>(), "Lower bound should be less than or equal to the upper bound")
-  return at::rrelu_with_noise(self, at::empty_like(self, LEGACY_CONTIGUOUS_MEMORY_FORMAT), lower, upper, training, std::move(generator));
+  auto noise = at::empty_like(self, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
+  return at::rrelu_with_noise(self, noise, lower, upper, training, std::move(generator));
 }
 
 Tensor & rrelu_(Tensor & self, const Scalar& lower, const Scalar& upper, bool training, std::optional<Generator> generator) {
   TORCH_CHECK(lower.to<double>() <= upper.to<double>(), "Lower bound should be less than or equal to the upper bound")
-  return at::rrelu_with_noise_(self, at::empty_like(self, LEGACY_CONTIGUOUS_MEMORY_FORMAT), lower, upper, training, std::move(generator));
+  auto noise = at::empty_like(self, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
+  return at::rrelu_with_noise_(self, noise, lower, upper, training, std::move(generator));
 }
 
 TORCH_IMPL_FUNC(threshold_out)(const Tensor& self, const Scalar& threshold, const Scalar& value, const Tensor& result) {

aten/src/ATen/native/cuda/RreluWithNoise.cu

@@ -71,7 +71,7 @@ template <typename scalar_t>
 inline void _rrelu_with_noise_cuda_train(
     Tensor& output,
     const Tensor& input_,
-    const Tensor& noise_,
+    Tensor& noise_,
     const Scalar& lower_,
     const Scalar& upper_,
     std::optional<Generator> generator) {
@@ -139,7 +139,7 @@ inline void _rrelu_with_noise_cuda_train(
 }
 
 Tensor& rrelu_with_noise_out_cuda(const Tensor& self,
-    const Tensor& noise,
+    Tensor& noise,
     const Scalar& lower,
     const Scalar& upper,
     bool training,
@@ -173,7 +173,7 @@ Tensor& rrelu_with_noise_out_cuda(const Tensor& self,
 
 Tensor rrelu_with_noise_cuda(
     const Tensor& self,
-    const Tensor& noise,
+    Tensor& noise,
     const Scalar& lower,
     const Scalar& upper,
     bool training,
@@ -184,7 +184,7 @@ Tensor rrelu_with_noise_cuda(
 
 Tensor& rrelu_with_noise_cuda_(
     Tensor& self,
-    const Tensor& noise,
+    Tensor& noise,
     const Scalar& lower,
     const Scalar& upper,
     bool training,

aten/src/ATen/native/native_functions.yaml

@@ -12018,19 +12018,20 @@
     CUDA: log_sigmoid_backward_cuda
     MPS: log_sigmoid_backward_mps
 
-- func: rrelu_with_noise.out(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None, *, Tensor(a!) out) -> Tensor(a!)
+- func: rrelu_with_noise.out(Tensor self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None, *, Tensor(a!) out) -> Tensor(a!)
   python_module: nn
   tags: nondeterministic_seeded
   dispatch:
     CPU: rrelu_with_noise_out_cpu
     CUDA: rrelu_with_noise_out_cuda
 
-- func: rrelu_with_noise(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
+- func: rrelu_with_noise(Tensor self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
   python_module: nn
   dispatch:
     CPU: rrelu_with_noise_cpu
     CUDA: rrelu_with_noise_cuda
   tags: nondeterministic_seeded
+  autogen: rrelu_with_noise_functional
 
 - func: rrelu_with_noise_backward(Tensor grad_output, Tensor self, Tensor noise, Scalar lower, Scalar upper, bool training, bool self_is_result) -> Tensor
   python_module: nn
@@ -12038,7 +12039,7 @@
     CompositeExplicitAutograd: rrelu_with_noise_backward
   autogen: rrelu_with_noise_backward.out
 
-- func: rrelu_with_noise_(Tensor(a!) self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor(a!)
+- func: rrelu_with_noise_(Tensor(a!) self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor(a!)
   python_module: nn
   tags: nondeterministic_seeded
   dispatch:

test/expect/HasDecompTest.test_has_decomposition.expect

@@ -1125,6 +1125,10 @@ aten::resize_as_sparse_
 aten::row_indices
 aten::row_indices_copy
 aten::row_indices_copy.out
+aten::rrelu_with_noise
+aten::rrelu_with_noise.out
+aten::rrelu_with_noise_
+aten::rrelu_with_noise_functional
 aten::scalar_tensor
 aten::scalar_tensor.out
 aten::scatter.reduce

test/forward_backward_compatibility/check_forward_backward_compatibility.py

@@ -137,6 +137,7 @@ ALLOW_LIST = [
     ("_quantized::wrapped_linear_prepack", datetime.date(2024, 12, 31)),
     ("_quantized::wrapped_linear_prepacked", datetime.date(2024, 12, 31)),
     ("_quantized::wrapped_quantized_linear_prepacked", datetime.date(2024, 12, 31)),
+    ("aten::rrelu_with_noise", datetime.date(2024, 12, 31)),
 ]
 
 ALLOW_LIST_COMPILED = [

test/functorch/test_aotdispatch.py

@@ -6327,6 +6327,42 @@ metadata incorrectly.
         out.sum().backward()
         self.assertEqual(ref_x.grad, x.grad)
 
+    def test_rrelu_with_noise_mutation(self):
+        def fn_functional(x):
+            noise = torch.ones_like(x)
+            result, noise_out = torch.ops.aten.rrelu_with_noise_functional(
+                x, noise, 0.2, 0.8, True
+            )
+            return result, noise_out
+
+        def fn_mutation(x):
+            noise = torch.ones_like(x)
+            result = torch.ops.aten.rrelu_with_noise(x, noise, 0.2, 0.8, True)
+            return result, noise
+
+        def fn_inplace(x):
+            noise = torch.ones_like(x, requires_grad=False)
+            torch.ops.aten.rrelu_with_noise_(x, noise, 0.2, 0.8, True)
+            return x, noise
+
+        def _test_fn(fn, check_backward=True):
+            x = -torch.abs(torch.randn(4, 4, dtype=torch.bfloat16, requires_grad=True))
+
+            ref_y, ref_noise = fn(x)
+            self.assertTrue(torch.all(ref_noise < torch.ones_like(ref_noise)).item())
+
+            comp_y, comp_noise = torch.compile(fn, backend="inductor", fullgraph=True)(
+                x
+            )
+
+            if check_backward:
+                comp_y.sum().backward()
+            self.assertTrue(torch.all(comp_noise < torch.ones_like(comp_noise)).item())
+
+        _test_fn(fn_functional)
+        _test_fn(fn_mutation)
+        _test_fn(fn_inplace, check_backward=False)
+
 
 # entries in here don't work and need to be fixed.
 # Each one of these is a bug (or needs to be investigated)

test/test_decomp.py

@@ -656,50 +656,6 @@ class TestDecomp(TestCase):
         for dim in (-1, 0, 1):
             self.assertEqual(torch.cat(inps, dim), cat_inductor(inps, dim))
 
-    def test_rrelu_with_noise(self, device):
-        # rrelu_with_noise behavior depends on a) whether elements in the input
-        # are <= 0, and b) whether we're in training mode. Cover all cases:
-        dtype = torch.float64
-        x = torch.tensor([-3.0, -2.0, -1.0, 0.0, 1.0, 2.0], dtype=dtype, device=device)
-        lower = 1.0
-        upper = 4.0
-        training = False
-
-        torch.manual_seed(123)
-        noise_ref = torch.zeros(x.shape, dtype=dtype, device=device)
-        ref = torch.ops.aten.rrelu_with_noise(x, noise_ref, lower, upper, training)
-
-        torch.manual_seed(123)
-        noise_res = torch.zeros(x.shape, dtype=dtype, device=device)
-        res = torch._decomp.decompositions.rrelu_with_noise(
-            x,
-            noise_res,
-            lower,
-            upper,
-            training,
-        )
-        self.assertEqual(ref, res)
-        self.assertEqual(noise_ref, noise_res)
-
-        # Now with training=True:
-        training = True
-
-        torch.manual_seed(123)
-        noise_ref = torch.zeros(x.shape, dtype=dtype, device=device)
-        ref = torch.ops.aten.rrelu_with_noise(x, noise_ref, lower, upper, training)
-
-        torch.manual_seed(123)
-        noise_res = torch.zeros(x.shape, dtype=dtype, device=device)
-        res = torch._decomp.decompositions.rrelu_with_noise(
-            x,
-            noise_res,
-            lower,
-            upper,
-            training,
-        )
-        self.assertEqual(ref, res)
-        self.assertEqual(noise_ref, noise_res)
-
     @suppress_warnings
     @tf32_off()
     # only tests RNNs since we have py dispsatcher decomps for them

tools/autograd/derivatives.yaml

@@ -2182,13 +2182,17 @@
   result0: at::where(self_p >= 0, grad_output_t, grad_output_t * weight_p + grad_output_p * weight_t)
   result1: at::where(self_p >= 0, at::zeros({}, self_p.options()), grad_output_p * self_t + grad_output_t * self_p)
 
-- name: rrelu_with_noise(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
+- name: rrelu_with_noise(Tensor self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor
   self: rrelu_with_noise_backward(grad, self, noise, lower, upper, training, false)
   result: auto_element_wise
 
-- name: rrelu_with_noise_(Tensor(a!) self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor(a!)
+- name: rrelu_with_noise_(Tensor(a!) self, Tensor(b!) noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> Tensor(a!)
   self: rrelu_with_noise_backward(grad, result, noise, lower, upper, training, true)
 
+- name: rrelu_with_noise_functional(Tensor self, Tensor noise, Scalar lower=0.125, Scalar upper=0.3333333333333333, bool training=False, Generator? generator=None) -> (Tensor, Tensor noise_out)
+  noise: non_differentiable
+  self: rrelu_with_noise_backward(grad, self, noise, lower, upper, training, false)
+
 - name: _softmax(Tensor self, int dim, bool half_to_float) -> Tensor
   self: _softmax_backward_data(grad, result, dim, self.scalar_type())
   result: result * (self_t - logsumexp_jvp(self_p, self_t, {dim}, true))

torch/_decomp/decompositions.py

@@ -305,44 +305,6 @@ def _prelu_kernel_backward(
     return (input_grad, weight_grad)
 
 
-@register_decomposition(aten.rrelu_with_noise)
-@aten.rrelu_with_noise.default.py_impl(DispatchKey.Autograd)
-@out_wrapper()
-@pw_cast_for_opmath
-def rrelu_with_noise(
-    self: Tensor,
-    noise: Tensor,
-    lower: float = 0.125,
-    upper: float = 0.3333333333333333,
-    training: bool = False,
-    generator: Optional[torch.Generator] = None,
-) -> Tensor:
-    assert generator is None
-    if training:
-        not_positive = self <= 0
-        r = aten.uniform(self, lower, upper)
-        output = torch.where(not_positive, self * r, self)
-        noise.copy_(torch.where(not_positive, r, 1))
-        return output
-    else:
-        negative_slope = (lower + upper) / 2
-        return aten.leaky_relu(self, negative_slope)
-
-
-@register_decomposition(aten.rrelu_with_noise_)
-@aten.rrelu_with_noise_.default.py_impl(DispatchKey.Autograd)
-@pw_cast_for_opmath
-def rrelu_with_noise_(
-    self: Tensor,
-    noise: Tensor,
-    lower: float = 0.125,
-    upper: float = 0.3333333333333333,
-    training: bool = False,
-    generator: Optional[torch.Generator] = None,
-) -> Tensor:
-    return self.copy_(rrelu_with_noise(self, noise, lower, upper, training, generator))
-
-
 @register_decomposition(aten.rrelu_with_noise_backward)
 @out_wrapper()
 @pw_cast_for_opmath

torch/_inductor/decomposition.py

@@ -76,6 +76,7 @@ inductor_decompositions = get_decompositions(
         aten.native_layer_norm,
         aten.nll_loss2d_backward,
         aten.permute_copy,
+        aten.rrelu_with_noise_backward,
         aten._softmax,
         aten.sin_,
         aten.sqrt_,
@@ -1022,3 +1023,23 @@ def searchsorted_scalar(
         side=side,
         sorter=sorter,
     )[0]
+
+
+@register_decomposition(aten.rrelu_with_noise_functional)
+def rrelu_with_noise_functional(
+    self: torch.Tensor,
+    noise: torch.Tensor,
+    lower: float = 0.125,
+    upper: float = 0.3333333333333333,
+    training: bool = False,
+    generator: Optional[torch.Generator] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    if training:
+        not_positive = self <= 0
+        r = aten.uniform(self, lower, upper, generator=generator)
+        output = torch.where(not_positive, self * r, self)
+        noise_out = torch.where(not_positive, r, 1)
+        return output, noise_out
+    else:
+        negative_slope = (lower + upper) / 2
+        return aten.leaky_relu(self, negative_slope), torch.Tensor()

torch/_meta_registrations.py

@@ -3734,6 +3734,28 @@ def meta__add_relu(self, other, alpha=1) -> Tensor:
     )
 
 
+@register_meta([aten.rrelu_with_noise])
+@out_wrapper()
+def meta_rrelu_with_noise(
+    self, noise, lower=0.125, upper=0.3333333333333333, training=False, generator=None
+):
+    return torch.empty_like(self)
+
+
+@register_meta([aten.rrelu_with_noise_functional])
+def meta_rrelu_with_noise_functional(
+    self, noise, lower=0.125, upper=0.3333333333333333, training=False, generator=None
+):
+    return torch.empty_like(self), torch.empty_like(noise)
+
+
+@register_meta([aten.rrelu_with_noise_])
+def meta_rrelu_with_noise_(
+    self, lower=0.125, upper=0.3333333333333333, training=False, generator=None
+):
+    return self
+
+
 @register_meta([aten.index_put.default, aten._unsafe_index_put.default])
 def meta_index_put(self, indices, values, accumulate=False):
     return torch.empty_like(self)

torchgen/native_function_generation.py

@@ -448,10 +448,10 @@ def add_generated_native_functions(
                 continue
 
             base_fn = (
-                d[SchemaKind.inplace]
-                if has_inplace
-                else d[SchemaKind.mutable]
+                d[SchemaKind.mutable]
                 if has_mutable
+                else d[SchemaKind.inplace]
+                if has_inplace
                 else d[SchemaKind.out]
                 if has_out
                 else d[SchemaKind.functional]