Fix int32 overflow in elem_count for shapes with >INT32_MAX elements

Author: edenamram

emlx/c_src/emlx_nif.cpp

@@ -218,7 +218,7 @@ NIF(to_blob) {
 }
 
 uint64_t elem_count(std::vector<int> shape) {
-  return std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<>{});
+  return std::accumulate(shape.begin(), shape.end(), uint64_t{1}, std::multiplies<uint64_t>{});
 }
 
 NIF(from_blob) {

emlx/test/emlx_test.exs

@@ -170,4 +170,23 @@ defmodule EMLXTest do
       end
     end
   end
+
+  describe "large tensors (element count > INT32_MAX)" do
+    # Regression: elem_count overflowed int32 for shapes whose element count
+    # exceeds INT32_MAX, causing "Binary size is too small" on valid binaries.
+    @tag :large_tensor
+    test "from_binary accepts shape whose element count exceeds INT32_MAX" do
+      # Reshape on BinaryBackend first — Nx.from_binary creates a flat 1D tensor
+      # whose single dimension would also exceed INT32_MAX.
+      binary = :binary.copy(<<7>>, 2_147_483_648)
+
+      t =
+        Nx.from_binary(binary, :u8, backend: Nx.BinaryBackend)
+        |> Nx.reshape({2, 1_073_741_824})
+        |> Nx.backend_transfer(EMLX.Backend)
+
+      assert Nx.shape(t) == {2, 1_073_741_824}
+      assert Nx.to_number(t[0][0]) == 7
+    end
+  end
 end