fix: stop_grad for logsumexp (#1470)

Author: polvalente

nx/lib/nx.ex

@@ -17332,6 +17332,16 @@ defmodule Nx do
     axes = opts[:axes]
     keep_axes = opts[:keep_axes]
     max = reduce_max(tensor, axes: axes, keep_axes: true)
+
+    max =
+      case max do
+        %T{data: %Nx.Defn.Expr{}} = t ->
+          Nx.Defn.Kernel.stop_grad(t)
+
+        t ->
+          t
+      end
+
     infinity_mask = is_infinity(max)
     max = select(infinity_mask, Nx.tensor(0, type: type), max)
     exponentials = tensor |> subtract(max) |> exp()

nx/test/nx/defn/expr_test.exs

@@ -188,6 +188,16 @@ defmodule Nx.Defn.ExprTest do
       named = Nx.tensor([4], names: [:dim])
       assert %T{type: {:f, 32}, names: [:dim]} = Nx.multiply(Expr.tensor(named), Expr.tensor(1.0))
     end
+
+    test "logsumexp" do
+      expr = Nx.logsumexp(Expr.tensor(Nx.tensor([1, 2, 3, 4, 5, 6])))
+
+      assert inspect(expr) =~ """
+               tensor a                                       s64[6]
+               b = reduce_max a, axes: [0], keep_axes: true   s64[1]
+               c = metadata b, :stop_grad                     s64[1]
+             """
+    end
   end
 
   describe "inspect" do