@@ -11,7 +11,6 @@
 #include <atomic>
 #include <cstddef>
 #include <cstring>
-#include <exception>
 #include <memory>
 #include <mutex>
 #include <stdexcept>
@@ -207,33 +206,6 @@ set_deallocation_stream(const DevicePtrHandle &h, const StreamHandle &h_stream)
     return CUDA_SUCCESS;
 }
 
-// Build the box for a freshly acquired device pointer. Binding the deallocation
-// stream and allocating the box can fail: make_deallocation_stream returns
-// false (err set) when no context is current, and either step can throw
-// std::bad_alloc. On any failure run undo() to release the pointer and return
-// nullptr; a thrown exception is left in `failure` for the caller to rethrow
-// once it holds no lock (undo may not report under a lock, see
-// deviceptr_import_ipc). Never throws.
-template <typename Undo>
-static DevicePtrBox *
-make_deviceptr_box(CUdeviceptr ptr, const StreamHandle &h_stream, Undo &&undo, std::exception_ptr &failure) noexcept
-{
-    DevicePtrBox *box = nullptr;
-    try {
-        DeallocationStream ds;
-        if (make_deallocation_stream(h_stream, ds)) {
-            box = new DevicePtrBox{ptr, std::move(ds)};
-        }
-    }
-    catch (...) {
-        failure = std::current_exception();
-    }
-    if (!box) {
-        undo();
-    }
-    return box;
-}
-
 DevicePtrHandle
 deviceptr_alloc_from_pool(size_t size, const MemoryPoolHandle &h_pool, const StreamHandle &h_stream)
 {
@@ -243,19 +215,15 @@ deviceptr_alloc_from_pool(size_t size, const MemoryPoolHandle &h_pool, const Str
         return {};
     }
 
-    std::exception_ptr failure;
-    DevicePtrBox *raw =
-        make_deviceptr_box(ptr, h_stream, [&]() noexcept { pw_cuMemFreeAsync(ptr, as_cu(h_stream)); }, failure);
-    if (!raw) {
-        if (failure) {
-            std::rethrow_exception(failure);
-        }
+    DeallocationStream ds;
+    if (!make_deallocation_stream(h_stream, ds)) {
+        pw_cuMemFreeAsync(ptr, as_cu(h_stream));
         return {};
     }
 
     // clang-format off
     auto box = std::shared_ptr<DevicePtrBox>(
-        raw,
+        new DevicePtrBox{ptr, std::move(ds)},
         [h_pool](DevicePtrBox *b) {
             GILReleaseGuard gil;
             const DeallocationStream& stream = b->deallocation;
@@ -281,19 +249,15 @@ deviceptr_alloc_async(size_t size, const StreamHandle &h_stream)
         return {};
     }
 
-    std::exception_ptr failure;
-    DevicePtrBox *raw =
-        make_deviceptr_box(ptr, h_stream, [&]() noexcept { pw_cuMemFreeAsync(ptr, as_cu(h_stream)); }, failure);
-    if (!raw) {
-        if (failure) {
-            std::rethrow_exception(failure);
-        }
+    DeallocationStream ds;
+    if (!make_deallocation_stream(h_stream, ds)) {
+        pw_cuMemFreeAsync(ptr, as_cu(h_stream));
         return {};
     }
 
     // clang-format off
     auto box = std::shared_ptr<DevicePtrBox>(
-        raw,
+        new DevicePtrBox{ptr, std::move(ds)},
         [](DevicePtrBox *b) {
             GILReleaseGuard gil;
             const DeallocationStream& stream = b->deallocation;
@@ -587,7 +551,6 @@ deviceptr_import_ipc(const MemoryPoolHandle &h_pool, const void *export_data, co
         GILReleaseGuard gil;
         CUresult discard_status = CUDA_SUCCESS;
         CUdeviceptr discarded = 0;
-        std::exception_ptr failure;
         {
             std::lock_guard<std::mutex> lock(ipc_import_mutex);
 
@@ -600,22 +563,11 @@ deviceptr_import_ipc(const MemoryPoolHandle &h_pool, const void *export_data, co
                 return {};
             }
 
-            // No deallocation stream could be recorded (or the box could not be
-            // allocated): discard the import with the raw call (a pw_ report would
-            // acquire the GIL under the mutex).
-            DevicePtrBox *raw = make_deviceptr_box(
-                ptr,
-                h_stream,
-                [&]() noexcept {
-                    discard_status = p_cuMemFreeAsync(ptr, as_cu(h_stream));  // raw: under ipc_import_mutex
-                    discarded = ptr;
-                },
-                failure
-            );
-            if (raw) {
+            DeallocationStream ds;
+            if (make_deallocation_stream(h_stream, ds)) {
                 // clang-format off
                 auto box = std::shared_ptr<DevicePtrBox>(
-                    raw,
+                    new DevicePtrBox{ptr, std::move(ds)},
                     [h_pool, key](DevicePtrBox *b) {
                         // Release the GIL first (the GIL is the outermost lock), then hold the
                         // mutex across unregister + free: a concurrent import that finds this
@@ -643,6 +595,11 @@ deviceptr_import_ipc(const MemoryPoolHandle &h_pool, const void *export_data, co
                 ipc_ptr_cache.register_handle(key, h);
                 return h;
             }
+
+            // No deallocation stream could be recorded: discard the import with
+            // the raw call (a pw_ report would acquire the GIL under the mutex).
+            discard_status = p_cuMemFreeAsync(ptr, as_cu(h_stream));  // raw: under ipc_import_mutex
+            discarded = ptr;
         }
         if (discard_status != CUDA_SUCCESS) {
             char operation[160];
@@ -654,9 +611,6 @@ deviceptr_import_ipc(const MemoryPoolHandle &h_pool, const void *export_data, co
                 "deallocation stream; the mapping leaked"
             );
         }
-        if (failure) {
-            std::rethrow_exception(failure);
-        }
         return {};
     }
     else {
@@ -666,19 +620,15 @@ deviceptr_import_ipc(const MemoryPoolHandle &h_pool, const void *export_data, co
             return {};
         }
 
-        std::exception_ptr failure;
-        DevicePtrBox *raw =
-            make_deviceptr_box(ptr, h_stream, [&]() noexcept { pw_cuMemFreeAsync(ptr, as_cu(h_stream)); }, failure);
-        if (!raw) {
-            if (failure) {
-                std::rethrow_exception(failure);
-            }
+        DeallocationStream ds;
+        if (!make_deallocation_stream(h_stream, ds)) {
+            pw_cuMemFreeAsync(ptr, as_cu(h_stream));
             return {};
         }
 
         // clang-format off
         auto box = std::shared_ptr<DevicePtrBox>(
-            raw,
+            new DevicePtrBox{ptr, std::move(ds)},
             [h_pool](DevicePtrBox *b) {
                 GILReleaseGuard gil;
                 const DeallocationStream& stream = b->deallocation;