diff --git a/parsec/mca/device/device_gpu.c b/parsec/mca/device/device_gpu.c
index 30e34789f..a5e11c322 100644
--- a/parsec/mca/device/device_gpu.c
+++ b/parsec/mca/device/device_gpu.c
@@ -1227,6 +1227,7 @@ parsec_default_gpu_stage_out(parsec_gpu_task_t        *gtask,
         if(flow_mask & (1U << i)){
             source = task->data[i].data_out;
             dest = source->original->device_copies[0];
+            if (NULL == dest) continue;
             dst_dev = (parsec_device_gpu_module_t*)parsec_mca_device_get(dest->device_index);
             src_dev = (parsec_device_gpu_module_t*)parsec_mca_device_get(source->device_index);
 
diff --git a/parsec/mca/device/transfer_gpu.c b/parsec/mca/device/transfer_gpu.c
index 986cda05e..54dd14613 100644
--- a/parsec/mca/device/transfer_gpu.c
+++ b/parsec/mca/device/transfer_gpu.c
@@ -241,6 +241,25 @@ parsec_gpu_create_w2r_task(parsec_device_gpu_module_t *gpu_device,
         parsec_atomic_lock( &gpu_copy->original->lock );
         /* get the next item before altering the next pointer */
         item = (parsec_list_item_t*)item->list_next;  /* conversion needed for volatile */
+
+        if (gpu_copy->original->device_copies[0] == NULL) {
+            /* no host-side copy available, release immediately if no readers */
+            if (gpu_copy->readers == 0) {
+              PARSEC_DEBUG_VERBOSE(10, parsec_gpu_output_stream,
+                                   "D2H[%d:%s] CPU data copy discarded, GPU data copy %p [%p] now available",
+                                   gpu_device->super.device_index, gpu_device->super.name, gpu_copy, gpu_copy->original);
+
+              parsec_list_item_ring_chop((parsec_list_item_t*)gpu_copy);
+              PARSEC_LIST_ITEM_SINGLETON(gpu_copy);
+              gpu_copy->coherency_state = PARSEC_DATA_COHERENCY_SHARED;
+              gpu_copy->data_transfer_status = PARSEC_DATA_STATUS_COMPLETE_TRANSFER;
+              parsec_list_push_back(&gpu_device->gpu_mem_lru, (parsec_list_item_t*)gpu_copy);
+              gpu_device->data_avail_epoch++;
+            }
+            parsec_atomic_unlock( &gpu_copy->original->lock );
+            continue;
+        }
+
         if( 0 == gpu_copy->readers ) {
             if( PARSEC_UNLIKELY(NULL == d2h_task) ) {  /* allocate on-demand */
                 d2h_task = (parsec_gpu_d2h_task_t*)parsec_thread_mempool_allocate(es->context_mempool);
@@ -284,6 +303,8 @@ parsec_gpu_create_w2r_task(parsec_device_gpu_module_t *gpu_device,
     w2r_task->stage_out        = &parsec_default_gpu_stage_out;
     w2r_task->complete_stage   = NULL;
 
+    parsec_atomic_fetch_inc_int32( &(gpu_device->mutex) );
+
     (void)es;
     return w2r_task;
 }
@@ -314,20 +335,22 @@ int parsec_gpu_complete_w2r_task(parsec_device_gpu_module_t *gpu_device,
 
         cpu_copy = original->device_copies[0];
 
-        if( cpu_copy->version < gpu_copy->version ) {
+        if( NULL != cpu_copy && cpu_copy->version < gpu_copy->version ) {
             /* the GPU version has been acquired by a new task that is waiting for submission */
             PARSEC_DEBUG_VERBOSE(10, parsec_gpu_output_stream,
                                  "D2H[%d:%s] task %p:%i GPU data copy %p [%p] has a backup in memory",
                                  gpu_device->super.device_index, gpu_device->super.name, (void*)task, i, gpu_copy, gpu_copy->original);
         } else {
             gpu_copy->coherency_state = PARSEC_DATA_COHERENCY_SHARED;
-            cpu_copy->coherency_state =  PARSEC_DATA_COHERENCY_SHARED;
-            cpu_copy->version = gpu_copy->version;
-            PARSEC_DEBUG_VERBOSE(10, parsec_gpu_output_stream,
-                                 "D2H[%d:%s]: CPU copy %p gets the same version %d as GPU copy %p at %s:%d",
-                                 gpu_device->super.device_index, gpu_device->super.name,
-                                 cpu_copy, cpu_copy->version, gpu_copy,
-                                 __FILE__, __LINE__);
+            if (NULL != cpu_copy) {
+                cpu_copy->coherency_state =  PARSEC_DATA_COHERENCY_SHARED;
+                cpu_copy->version = gpu_copy->version;
+                PARSEC_DEBUG_VERBOSE(10, parsec_gpu_output_stream,
+                                     "D2H[%d:%s]: CPU copy %p gets the same version %d as GPU copy %p at %s:%d",
+                                     gpu_device->super.device_index, gpu_device->super.name,
+                                     cpu_copy, cpu_copy->version, gpu_copy,
+                                     __FILE__, __LINE__);
+            }
             PARSEC_DEBUG_VERBOSE(10, parsec_gpu_output_stream,
                                  "D2H[%d:%s] task %p:%i GPU data copy %p [%p] now available",
                                  gpu_device->super.device_index, gpu_device->super.name, (void*)task, i, gpu_copy, gpu_copy->original);
@@ -338,5 +361,6 @@ int parsec_gpu_complete_w2r_task(parsec_device_gpu_module_t *gpu_device,
     parsec_thread_mempool_free(es->context_mempool, task);
     free(gpu_task);
     gpu_device->data_avail_epoch++;
+    parsec_atomic_fetch_dec_int32( &(gpu_device->mutex) );
     return 0;
 }