fix(grad): Nx.stack grad should remove the added axis (unbroadcast) (#…

…1536)

nx/lib/nx/defn/grad.ex

@@ -23,6 +23,7 @@ defmodule Nx.Defn.Grad do
     expr = to_grad |> fun.()
 
     transformed_expr = transform.(expr) |> validate_expr!() |> Nx.devectorize(keep_names: false)
+
     {parents, nodes} = parents_tree(transformed_expr, ids)
 
     to_grad_ids = {to_grad, ids}
@@ -623,7 +624,9 @@ defmodule Nx.Defn.Grad do
         current_limit = 1 + limit
         start = List.replace_at(zero_axes, axis, limit)
         len = List.replace_at(ans_shape_list, axis, 1)
-        {{t, Nx.slice(g, start, len)}, current_limit}
+        g = Nx.slice(g, start, len)
+        g = Nx.squeeze(g, axes: [axis])
+        {{t, g}, current_limit}
       end)
 
     pairs

nx/test/nx/defn/grad_test.exs

@@ -1773,6 +1773,24 @@ defmodule Nx.Defn.GradTest do
     end
   end
 
+  describe "stack" do
+    test "works on compound functions for more than 1 axis" do
+      # This is a test that ensures that the added axis from the
+      # stack operation is correctly squeezed back out by
+      # the gradient computation.
+      x = 2.0
+
+      assert grad(Nx.tensor([[x]]), fn t ->
+               a = Nx.pow(t, 2)
+               b = Nx.pow(t, 3)
+               c = Nx.pow(t, 4)
+
+               Nx.stack([a, b, c], axis: 1)
+               |> Nx.sum()
+             end) == Nx.tensor([[2 * x + 3 * x ** 2 + 4 * x ** 3]])
+    end
+  end
+
   describe "cholesky" do
     defn cholesky_grad(t) do
       grad(t, fn x -> x |> Nx.LinAlg.cholesky() |> Nx.sum() end)