Allow int vals to go down the fastpath for _foreach_max (#127303)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/127303
Approved by: https://github.com/albanD
ghstack dependencies: #127187

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

@@ -68,8 +68,8 @@ struct LpMaxFunctor {
     T vals[kILP];
     T r_x[kILP];
     for (int64_t i = 0; i < kILP; i++) {
-      vals[i] = T(-INFINITY);
-      r_x[i] = T(-INFINITY);
+      vals[i] = T(std::numeric_limits<T>::lowest());
+      r_x[i] = T(std::numeric_limits<T>::lowest());
     }
 
     if (n % kILP == 0 && (chunk_size & kILP) == 0 && is_aligned(x)) {
@@ -96,7 +96,7 @@ struct LpMaxFunctor {
       }
     }
 
-    auto val = T(-INFINITY);
+    auto val = T(std::numeric_limits<T>::lowest());
     for (int i = 0; i < kILP; i++) {
       val = max_propagate_nan(val, vals[i]);
     }
@@ -118,7 +118,7 @@ __global__ void lpmax_cleanup(
   __shared__ T vals[512];
   const T* output_this_tensor =
       output_per_tensor + blockIdx.x * max_chunks_per_tensor;
-  T val = -INFINITY;
+  T val = std::numeric_limits<T>::lowest();
   for (size_t i = threadIdx.x; i < max_chunks_per_tensor; i += blockDim.x) {
     val = max_propagate_nan(val, output_this_tensor[i]);
   }
@@ -130,21 +130,11 @@ __global__ void lpmax_cleanup(
 
 std::vector<Tensor> foreach_tensor_max_cuda(TensorList tensors) {
   check_foreach_api_restrictions(tensors);
-  // we currently use -INF as the identity value to compare against, which
-  // does not work for int8, int16, nor bool. Fall back to slow path here.
-  const bool has_small_int_or_bool =
-      std::any_of(tensors.begin(), tensors.end(), [](const auto& t) {
-        const auto scalar_type = t.scalar_type();
-        return scalar_type == at::ScalarType::Short ||
-            scalar_type == at::ScalarType::Char ||
-            scalar_type == at::ScalarType::Bool;
-      });
-  if (!can_use_fast_route(tensors) || has_small_int_or_bool) {
+  if (!can_use_fast_route(tensors)) {
     return foreach_tensor_max_slow(tensors);
   }
 
-  // for parity with max in ReduceAllOps.cpp, though I think max(empty) should
-  // eventually be allowed.
+  // for parity with max in ReduceAllOps.cpp, as max(empty) is ???
   TORCH_CHECK(
       std::all_of(
           tensors.begin(),

aten/src/ATen/native/cuda/block_reduce.cuh

@@ -103,7 +103,7 @@ __inline__ __device__ T BlockReduceMax(T val, T* shared) {
     shared[wid] = val;
   }
   __syncthreads();
-  val = (tid < B::Warps()) ? shared[lid] : T(-INFINITY);
+  val = (tid < B::Warps()) ? shared[lid] : T(std::numeric_limits<T>::lowest());
   if (wid == 0) {
     val = WarpReduceMax(val);
   }

test/test_foreach.py

@@ -1015,7 +1015,7 @@ class TestForeach(TestCase):
     def test_foreach_reduce_large_input(self, device, dtype, op):
         # test inputs larger than kChunkSize = 65536
         N = 65536 * 2
-        disable_fastpath = dtype in (torch.int8, torch.int16, torch.bool)
+        disable_fastpath = False
         kwargs = {}
         if op.name == "_foreach_norm":
             ord = 2

torch/testing/_internal/common_methods_invocations.py

@@ -9380,7 +9380,7 @@ class foreach_max_sample_func(foreach_inputs_sample_func):
         return []
 
     def _should_disable_fastpath(self, opinfo, rightmost_arg, rightmost_arg_type, dtype):
-        return dtype in (torch.int8, torch.int16, torch.bool)
+        return False
 
 
 class foreach_norm_sample_func(foreach_inputs_sample_func):
@@ -11125,6 +11125,7 @@ foreach_reduce_op_db: List[ForeachFuncInfo] = [
         supports_inplace_autograd=True,
         supports_forward_ad=True,
         decorators=(
+            # no complex support for ordering ops like max
             DecorateInfo(
                 unittest.expectedFailure,
                 "TestForeach",