diff --git a/mlx/backend/metal/allocator.cpp b/mlx/backend/metal/allocator.cpp
index 07f50299..3debc474 100644
--- a/mlx/backend/metal/allocator.cpp
+++ b/mlx/backend/metal/allocator.cpp
@@ -29,6 +29,7 @@ BufferCache::BufferCache(MTL::Device* device)
     : device_(device), head_(nullptr), tail_(nullptr), pool_size_(0) {}
 
 BufferCache::~BufferCache() {
+  auto thread_pool = metal::new_scoped_memory_pool();
   clear();
 }
 
@@ -166,6 +167,8 @@ Buffer MetalAllocator::malloc(size_t size, bool allow_swap /* = false */) {
       return Buffer{nullptr};
     }
 
+    auto thread_pool = metal::new_scoped_memory_pool();
+
     // If we have a lot of memory pressure, check if we can reclaim some memory
     // from the cache
     if (device_->currentAllocatedSize() + size >= gc_limit_) {
diff --git a/mlx/backend/metal/device.cpp b/mlx/backend/metal/device.cpp
index c48f2908..6d7528f8 100644
--- a/mlx/backend/metal/device.cpp
+++ b/mlx/backend/metal/device.cpp
@@ -19,16 +19,14 @@ namespace mlx::core::metal {
 
 namespace {
 
-// Catch things related to the main-thread static variables
-static std::shared_ptr<void> global_memory_pool = new_scoped_memory_pool();
-
 // TODO nicer way to set this or possibly expose as an environment variable
 static constexpr int MAX_BUFFERS_PER_QUEUE = 12;
 
 static constexpr const char* default_mtllib_path = METAL_PATH;
 
 auto load_device() {
-  MTL::Device* device = MTL::CreateSystemDefaultDevice();
+  auto devices = MTL::CopyAllDevices();
+  auto device = static_cast<MTL::Device*>(devices->object(0));
   if (!device) {
     throw std::runtime_error("Failed to load device");
   }
@@ -120,6 +118,7 @@ Device::Device() {
 }
 
 Device::~Device() {
+  auto pool = new_scoped_memory_pool();
   for (auto& q : queue_map_) {
     q.second->release();
   }
@@ -139,6 +138,8 @@ Device::~Device() {
 }
 
 void Device::new_queue(int index) {
+  auto thread_pool = metal::new_scoped_memory_pool();
+  
   // Multiple threads can ask the device for queues
   // We lock this as a critical section for safety
   const std::lock_guard<std::mutex> lock(mtx_);
diff --git a/mlx/scheduler.h b/mlx/scheduler.h
index 150cc96d..755efeaf 100644
--- a/mlx/scheduler.h
+++ b/mlx/scheduler.h
@@ -35,8 +35,7 @@ struct StreamThread {
   }
 
   void thread_fn() {
-    auto thread_pool = metal::new_scoped_memory_pool();
-    metal::new_stream(stream);
+    bool initialized = false;
     while (true) {
       std::function<void()> task;
       {
@@ -48,6 +47,16 @@ struct StreamThread {
         task = std::move(q.front());
         q.pop();
       }
+
+      // thread_fn may be called from a static initializer and we cannot
+      // call metal-cpp until all static initializers have completed. waiting
+      // for a task to arrive means that user code is running so metal-cpp
+      // can safely be called.
+      if (!initialized) {
+        initialized = true;
+        metal::new_stream(stream);
+      }
+
       task();
     }
   }