diff --git a/dory/Hardware_targets/PULP/Common/Templates/network.c.t b/dory/Hardware_targets/PULP/Common/Templates/network.c.t
index c81189f7..698eb346 100644
--- a/dory/Hardware_targets/PULP/Common/Templates/network.c.t
+++ b/dory/Hardware_targets/PULP/Common/Templates/network.c.t
@@ -104,7 +104,7 @@ void ${prefix}execute_layer_fork(void *args) {
   if (pi_core_id() == 0) pmsis_l1_malloc_free(layer_args->L1_buffer, ${l1_buffer});
 }
 
-void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""})
+struct ${prefix}network_run_token ${prefix}network_run_async(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""})
 {
   struct pi_device cluster_dev = {0};
   struct pi_cluster_conf conf;
@@ -132,12 +132,24 @@ void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final
   cluster_task.stack_size = ${master_stack};
   cluster_task.slave_stack_size = ${slave_stack};
   pi_cluster_send_task_to_cl(&cluster_dev, &cluster_task);
-  pi_cluster_close(&cluster_dev);
+  return (struct ${prefix}network_run_token) {
+    .cluster_dev = cluster_dev
+  };
+}
+
+void ${prefix}network_run_wait(struct ${prefix}network_run_token token)
+{
+  pi_cluster_close(&token.cluster_dev);
   % if 'Perf_final' in verbose_level:
   print_perf("Final", ${prefix}cycle_network_execution, ${MACs});
   % endif
 }
 
+void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""})
+{
+  ${prefix}network_run_wait(network_run_async(l2_buffer, l2_buffer_size, l2_final_output, exec${", L2_input_h" if not l3_supported else ""}));
+}
+
 void ${prefix}network_run_cluster(void *args) {
   unsigned int * real_args = (unsigned int *) args;
   void * l2_buffer = (void *) real_args[0];
diff --git a/dory/Hardware_targets/PULP/Common/Templates/network.h.t b/dory/Hardware_targets/PULP/Common/Templates/network.h.t
index bbb15d56..595aebaf 100644
--- a/dory/Hardware_targets/PULP/Common/Templates/network.h.t
+++ b/dory/Hardware_targets/PULP/Common/Templates/network.h.t
@@ -20,9 +20,6 @@
 #ifndef __${prefix.upper()}NETWORK_H__
 #define __${prefix.upper()}NETWORK_H__
 
-% if sdk == 'gap_sdk':
-#include "pulp.h"
-% endif
 <%
    l3_supported = DORY_HW_graph[0].HW_description['memory']['levels'] > 2
    single_input = n_inputs==1
@@ -31,6 +28,12 @@
 #include "${prefix}weights_definition.h"
 % endif
 #include <stddef.h>
+#include "pmsis.h"
+
+
+struct ${prefix}network_run_token {
+  struct pi_device cluster_dev;
+};
 
 
 % if l3_supported:
@@ -38,6 +41,8 @@ void ${prefix}network_terminate();
 void ${prefix}network_initialize();
 % endif
 void ${prefix}network_run_cluster(void * args);
+struct ${prefix}network_run_token ${prefix}network_run_async(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""});
+void network_run_wait(struct ${prefix}network_run_token token);
 void ${prefix}network_run(void *l2_buffer, size_t l2_buffer_size, void *l2_final_output, int exec${", void *L2_input_h" if not l3_supported else ""});
 void ${prefix}execute_layer_fork(void *arg);