Fix crashes in multi-threaded process teardown

mlx/backend/metal/device.cpp

@@ -805,8 +805,11 @@ void Device::set_residency_set(const MTL::ResidencySet* residency_set) {
 }
 
 Device& device(mlx::core::Device) {
-  static Device metal_device;
-  return metal_device;
+  // Leak singleton device intentionally, to avoid cases where a compute kernel
+  // returns and tries to access the object after it has been freed by the main
+  // thread teardown.
+  static Device* metal_device = new Device;
+  return *metal_device;
 }
 
 std::unique_ptr<void, std::function<void(void*)>> new_scoped_memory_pool() {

mlx/scheduler.cpp

@@ -61,16 +61,12 @@ namespace scheduler {
 
 /** A singleton scheduler to manage devices, streams, and task execution. */
 Scheduler& scheduler() {
-  // Leak the scheduler on Windows to avoid joining threads on exit, can be
-  // removed after Visual Studio fixes bug:
-  // https://developercommunity.visualstudio.com/t/1654756
-#ifdef _WIN32
+  // Intentionally leaked to avoid the "static destruction order fiasco":
+  // background threads (e.g. command buffer completion handlers) may
+  // reference this singleton after other static objects are destroyed
+  // during process teardown.
   static Scheduler* scheduler = new Scheduler;
   return *scheduler;
-#else
-  static Scheduler scheduler;
-  return scheduler;
-#endif
 }
 
 } // namespace scheduler

tests/CMakeLists.txt

@@ -42,3 +42,11 @@ target_link_options(tests PRIVATE ${SANITIZER_LINK_FLAGS})
 
 doctest_discover_tests(tests)
 add_test(NAME tests COMMAND tests)
+
+# Standalone test: verify clean exit when GPU work is in-flight during teardown.
+# (Cannot be a doctest case because the crash occurs during static destruction.)
+add_executable(test_teardown test_teardown.cpp)
+target_link_libraries(test_teardown PRIVATE mlx)
+target_compile_options(test_teardown PRIVATE ${SANITIZER_COMPILE_FLAGS})
+target_link_options(test_teardown PRIVATE ${SANITIZER_LINK_FLAGS})
+add_test(NAME teardown COMMAND test_teardown)

tests/test_teardown.cpp

@@ -0,0 +1,37 @@
+// Copyright © 2026 Apple Inc.
+//
+// Regression test for https://github.com/ml-explore/mlx/issues/3126
+// Verifies that the process exits cleanly when a background thread is
+// performing GPU work and the main thread exits.
+
+#include <chrono>
+#include <iostream>
+#include <thread>
+
+#include "mlx/mlx.h"
+
+namespace mx = mlx::core;
+
+int main() {
+  using namespace std::chrono_literals;
+
+  std::thread t([] {
+    auto a = mx::random::normal({2048, 2048});
+    std::cout << "START" << std::endl;
+    for (int i = 0; i < 1000; ++i) {
+      a = mx::matmul(a, a);
+      // Eval periodically to avoid building a huge graph
+      if (i % 10 == 0) {
+        mx::eval(a);
+        std::cout << "Step " << i << std::endl;
+      }
+    }
+    mx::eval(a);
+    std::cout << "Done: " << a.shape(0) << "x" << a.shape(1) << std::endl;
+  });
+
+  std::this_thread::sleep_for(1s);
+  t.detach();
+  std::cout << "Main thread exiting." << std::endl;
+  return 0;
+}