server : defer tasks when "slot unavailable" (#5018)

* server: defer task when no slot is available

* remove unnecessary log

---------

Co-authored-by: Xuan Son Nguyen <redacted>

diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 93f99929880f640c967602d4153cf5163b750789..0462fbd24739b7b214b0ca26565e835a48a89a38 100644 (file)
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -1558,6 +1558,7 @@ struct llama_server_context
     void process_tasks()
     {
         std::unique_lock<std::mutex> lock(mutex_tasks);
+        std::vector<task_server> deferred_tasks;
         while (!queue_tasks.empty())
         {
             task_server task = queue_tasks.front();
@@ -1568,9 +1569,8 @@ struct llama_server_context
                     llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
                     if (slot == nullptr)
                     {
-                        LOG_TEE("slot unavailable\n");
-                        // send error result
-                        send_error(task, "slot unavailable");
+                        // if no slot is available, we defer this task for processing later
+                        deferred_tasks.push_back(task);
                         break;
                     }
 
@@ -1616,6 +1616,12 @@ struct llama_server_context
             }
         }
 
+        // add all the deferred tasks back the the queue
+        for (task_server &task : deferred_tasks)
+        {
+            queue_tasks.push_back(task);
+        }
+
         // remove finished multitasks from the queue of multitasks, and add the corresponding result to the result queue
         std::vector<task_result> agg_results;
         auto queue_iterator = queue_multitasks.begin();