Fix singleton lifetime issues at process exit

mlx/backend/cpu/compiled.cpp

@@ -41,8 +41,11 @@ struct CompilerCache {
 };
 
 static CompilerCache& cache() {
-  static CompilerCache cache_;
-  return cache_;
+  // Leak - see Scheduler singleton comment in scheduler.cpp.
+  // DLib destructors call dlclose() which unmaps JIT .so files;
+  // StreamThreads may still be executing that code at exit.
+  static CompilerCache* cache_ = new CompilerCache;
+  return *cache_;
 };
 
 // GPU compile is always available if the GPU is available and since we are in

mlx/io/load.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023-2024 Apple Inc.
+// Copyright © 2023-2026 Apple Inc.
 #include <algorithm>
 #include <cstring>
 #include <fstream>
@@ -335,13 +335,15 @@ array load(std::string file, StreamOrDevice s) {
 namespace io {
 
 ThreadPool& thread_pool() {
-  static ThreadPool pool_{4};
-  return pool_;
+  // Leak - see Scheduler singleton comment in scheduler.cpp.
+  static ThreadPool* pool_ = new ThreadPool{4};
+  return *pool_;
 }
 
 ThreadPool& ParallelFileReader::thread_pool() {
-  static ThreadPool thread_pool{4};
-  return thread_pool;
+  // Leak - see Scheduler singleton comment in scheduler.cpp.
+  static ThreadPool* thread_pool = new ThreadPool{4};
+  return *thread_pool;
 }
 
 void ParallelFileReader::read(char* data, size_t n) {

mlx/scheduler.cpp

@@ -1,4 +1,4 @@
-// Copyright © 2023 Apple Inc.
+// Copyright © 2023-2026 Apple Inc.
 
 #include "mlx/scheduler.h"
 #include "mlx/backend/cpu/eval.h"
@@ -58,12 +58,12 @@ void Scheduler::enqueue(Stream s, std::function<void()> task) {
   st->enqueue(std::move(task));
 }
 
-/** A singleton scheduler to manage devices, streams, and task execution. */
+// Leak the scheduler singleton on all platforms. During static destruction,
+// worker threads may still be executing JIT-compiled code that has been
+// unmapped, causing SIGSEGV (macOS/Linux) or join() deadlocks (Windows/MSVC
+// CRT).
+// The OS reclaims all resources at process exit anyway.
 Scheduler& scheduler() {
-  // 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;
 }