fix: fixes Nx.slice with tensor start_indices which affected defn_whi…

…le and Nx.Random PRNG (mimiquate#32)

lib/candlex/backend.ex

@@ -754,7 +754,11 @@ defmodule Candlex.Backend do
 
   defp narrow(t, [start | starts], [length | lengths], axis, shape) do
     dim = elem(shape, axis)
-    start = min(start, dim - length)
+
+    start =
+      start
+      |> Nx.to_number()
+      |> min(dim - length)
 
     if start == 0 and length == dim do
       # Nothing to narrow at this step

lib/candlex/native.ex

@@ -4,7 +4,10 @@ defmodule Candlex.Native do
   mix_config = Mix.Project.config()
   version = mix_config[:version]
   source_url = mix_config[:package][:links]["GitHub"]
-  mode = if Mix.env() in [:dev, :test], do: :debug, else: :release
+  # We can't run on :debug mode until we find a workaround to
+  # ignore integer overflows when running Nx.Random Threefry PRNG.
+  # mode = if Mix.env() in [:dev, :test], do: :debug, else: :release
+  mode = :release
 
   use RustlerPrecompiled,
     otp_app: :candlex,

test/candlex_test.exs

@@ -1824,6 +1824,18 @@ defmodule CandlexTest do
       #     [5, 6]
       #   ]
       # ))
+
+      t([0, 1])
+      |> Nx.slice([t(0)], [1])
+      |> assert_equal(t([0]))
+
+      t([0, 1])
+      |> Nx.slice([t(1)], [1])
+      |> assert_equal(t([1]))
+
+      t([[1, 2, 3], [4, 5, 6]])
+      |> Nx.slice([t(0), t(1)], [1, 1])
+      |> assert_equal(t([[2]]))
     end
 
     test "squeeze" do

test/defn_test.exs

@@ -21,4 +21,90 @@ defmodule Candlex.DefnTest do
     |> TG.tanh_grad()
     |> assert_close(Nx.tensor(0.41997432708740234))
   end
+
+  describe "while/3" do
+    defmodule Mod do
+      import Nx.Defn
+
+      defn upto10(x) do
+        while x, Nx.less(x, 10) do
+          x + 1
+        end
+      end
+
+      defn factorial_tuple(x) do
+        factorial = Nx.tensor(1, type: Nx.type(x))
+
+        {factorial, _} =
+          while {factorial, x}, Nx.greater(x, 1) do
+            {factorial * x, x - 1}
+          end
+
+        factorial
+      end
+
+      defn factorial_map(x) do
+        factorial = Nx.tensor(1, type: Nx.type(x))
+
+        %{factorial: factorial} =
+          while map = %{factorial: factorial, x: x}, Nx.greater(map.x, 1) do
+            %{map | factorial: map.factorial * map.x, x: map.x - 1}
+          end
+
+        factorial
+      end
+
+      defn factorial_map_input(map) do
+        %{factorial: factorial} =
+          while map, Nx.greater(map.x, 1) do
+            %{map | factorial: map.factorial * map.x, x: map.x - 1}
+          end
+
+        factorial
+      end
+
+      defn tensor_generator_sum() do
+        while x = 0, r <- Nx.tensor([0, 1, 2]) do
+          x + r
+        end
+      end
+    end
+
+    test "simple" do
+      Mod.upto10(0)
+      |> assert_equal(Nx.tensor(10))
+
+      Mod.upto10(5)
+      |> assert_equal(Nx.tensor(10))
+    end
+
+    test "factorial tuple" do
+      Mod.factorial_tuple(5)
+      |> assert_equal(Nx.tensor(120))
+
+      Mod.factorial_tuple(10.0)
+      |> assert_equal(Nx.tensor(3_628_800.0))
+    end
+
+    test "factorial map" do
+      Mod.factorial_map(5)
+      |> assert_equal(Nx.tensor(120))
+
+      Mod.factorial_map(10.0)
+      |> assert_equal(Nx.tensor(3_628_800.0))
+    end
+
+    test "factorial map input" do
+      Mod.factorial_map_input(%{factorial: 1, x: 5})
+      |> assert_equal(Nx.tensor(120))
+
+      Mod.factorial_map_input(%{factorial: 1.0, x: 10.0})
+      |> assert_equal(Nx.tensor(3_628_800.0))
+    end
+
+    test "tensor generator sum" do
+      Mod.tensor_generator_sum()
+      |> assert_equal(Nx.tensor(3))
+    end
+  end
 end

test/random_test.exs

@@ -0,0 +1,32 @@
+defmodule Candlex.RandomTest do
+  use Nx.Case, async: true
+
+  test "key/1" do
+    Nx.Random.key(42)
+    |> assert_equal(Nx.tensor([0, 42]))
+  end
+
+  test "uniform/1" do
+    {normal, new_key} =
+      Nx.Random.key(42)
+      |> Nx.Random.uniform()
+
+    normal
+    |> assert_close(Nx.tensor(0.9145736694335938))
+
+    new_key
+    |> assert_equal(Nx.tensor([2_465_931_498, 3_679_230_171]))
+  end
+
+  test "normal/1" do
+    {normal, new_key} =
+      Nx.Random.key(42)
+      |> Nx.Random.normal()
+
+    normal
+    |> assert_close(Nx.tensor(1.3694695234298706))
+
+    new_key
+    |> assert_equal(Nx.tensor([2_465_931_498, 3_679_230_171]))
+  end
+end