Enable max.unary_out (pytorch#85926)

Pull Request resolved: pytorch#85926
Approved by: https://github.com/bdhirsh

aten/src/ATen/native/ReduceAllOps.cpp

@@ -34,9 +34,16 @@ Tensor max(const Tensor &self) {
 }
 
 Tensor& max_unary_out(const Tensor &self, Tensor& out) {
-  Tensor tmp_output = at::max(self);
-  at::native::resize_output(out, tmp_output.sizes());
-  out.copy_(tmp_output);
+  // First check if the devices match (CPU vs GPU)
+  TORCH_CHECK(self.device() == out.device());
+
+  TORCH_CHECK(canCast(
+      typeMetaToScalarType(self.dtype()),
+      typeMetaToScalarType(out.dtype())));
+
+  at::native::resize_output(out, {});
+
+  max_all_stub(self.device().type(), out, self.contiguous());
   return out;
 }
 

aten/src/ATen/native/native_functions.yaml

@@ -8711,13 +8711,6 @@
     MPS: max_mps
     QuantizedCPU: max_quantized_cpu
 
-# Not to be confused with binary op `max.out`. Commented because of failed CI
-# FIXME: enable this
-#- func: max.unary_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
-#  device_check: NoCheck   # TensorIterator
-#  dispatch:
-#    CompositeExplicitAutograd: max_unary_out
-
 - func: fmax(Tensor self, Tensor other) -> Tensor
   structured_delegate: fmax.out
   device_check: NoCheck   # TensorIterator
@@ -8752,6 +8745,12 @@
 - func: max.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
   device_check: NoCheck   # TensorIterator
 
+- func: max.unary_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
+  device_check: NoCheck   # TensorIterator
+  dispatch:
+    CPU, CUDA: max_unary_out
+    QuantizedCPU: max_quantized_unary_out
+
 - func: minimum(Tensor self, Tensor other) -> Tensor
   structured_delegate: minimum.out
   device_check: NoCheck   # TensorIterator

aten/src/ATen/native/quantized/TensorCompare.cpp

@@ -14,6 +14,19 @@ Tensor max_quantized_cpu(const Tensor& self) {
   return std::get<0>(self.reshape({-1}).max(/*dim=*/0));
 }
 
+Tensor& max_quantized_unary_out(const Tensor& self, Tensor& out) {
+  // TODO this implementation is inefficient for now.
+  TORCH_CHECK(self.device() == out.device());
+
+  TORCH_CHECK(canCast(
+      typeMetaToScalarType(self.dtype()),
+      typeMetaToScalarType(out.dtype())));
+  Tensor temp = max_quantized_cpu(self);
+  at::native::resize_output(out, temp.sizes());
+  out.copy_(temp);
+  return out;
+}
+
 Tensor min_quantized_cpu(const Tensor& self) {
   return std::get<0>(self.reshape({-1}).min(/*dim=*/0));
 }

torch/_meta_registrations.py

@@ -110,11 +110,17 @@ def meta_index_select_out(self, dim, index, out):
     return out.copy_(torch.index_select(self, dim, index))
 
 
-@register_meta([aten.max.default, aten.min.default])
+@register_meta([aten.max.default, aten.max.unary_out])
+@out_wrapper()
 def meta_max(self):
     return self.new_empty(())
 
 
+@register_meta([aten.min.default])
+def meta_min(self):
+    return self.new_empty(())
+
+
 @register_meta(aten.angle.default)
 def meta_angle(self):
     if self.is_complex():

torch/csrc/jit/frontend/schema_matching.cpp

@@ -324,16 +324,23 @@ static bool varargsCanBeUsedAsList(
       !typevar_list;
 }
 
-// Note (@zasdfgbnm):
-// This is a workaround for https://github.com/pytorch/pytorch/issues/47964
-// Currently JIT does not distinguish ScalarType vs int, so there is really
-// no way to distinguish x.view(1) vs x.view(torch.int8). So we have to hardcode
-// the aten::view.dtype here to block this overload. This blocklist should be
-// removed when JIT fully suports ScalarType as its own type.
 bool isBlockListedSchema(const FunctionSchema& schema) {
+  // Note (@zasdfgbnm):
+  // This is a workaround for https://github.com/pytorch/pytorch/issues/47964
+  // Currently JIT does not distinguish ScalarType vs int, so there is really
+  // no way to distinguish x.view(1) vs x.view(torch.int8). So we have to
+  // hardcode the aten::view.dtype here to block this overload. This blocklist
+  // should be removed when JIT fully suports ScalarType as its own type.
   if (schema.name() == "aten::view" && schema.overload_name() == "dtype") {
     return true;
   }
+  // Note (@tugsbayasgalan)
+  // TorchScript doesn't suport kwargs so this op collides with aten.max.others
+  // since both of them have 2 Tensor inputs. Since we don't expect users to
+  // use this op in TS, we just skip it
+  if (schema.name() == "aten::max" && schema.overload_name() == "unary_out") {
+    return true;
+  }
   return false;
 }
 

torch/testing/_internal/common_methods_invocations.py

@@ -10238,7 +10238,7 @@ def reference_flatten(input, start_dim=0, end_dim=-1):
     OpInfo('max',
            variant_test_name='reduction_no_dim',
            dtypes=all_types_and(torch.float16, torch.bfloat16, torch.bool),
-           supports_out=False,
+           supports_out=True,
            supports_forward_ad=True,
            supports_fwgrad_bwgrad=True,
            sample_inputs_func=sample_inputs_max_min_reduction_no_dim,