fix(server): a slot action pops the next deferred task (pilotlight 114)

Author: Stratum

tools/server/server-context.cpp

@@ -39,6 +39,25 @@
 
 constexpr int HTTP_POLLING_SECONDS = 1;
 
+// KVFIX-114: a slot action (save, restore, erase) runs on the task loop
+// without launching its slot, so no release follows it and
+// slot.callback_on_release never pops the next deferred task. A completion
+// deferred alongside the action was then left in queue_tasks_deferred until
+// some other request released the slot: forever, on an idle server
+// (pilotlight item 114, spike 99 UPSTREAM.md). Constructed once the action
+// is committed to run (after the is_processing() deferral check), it pops
+// one deferred task on every exit, success or error. pop_deferred_task
+// still prefers a task naming this slot, so a chain of waiting slot actions
+// drains one after another and then one generation task starts.
+struct slot_action_drain {
+    server_queue & queue;
+    const int      id_slot;
+
+    ~slot_action_drain() {
+        queue.pop_deferred_task(id_slot);
+    }
+};
+
 static common_speculative_output_limits server_output_limits(const common_params & params) {
     if (params.embedding ||
             (params.pooling_type != LLAMA_POOLING_TYPE_UNSPECIFIED && params.pooling_type != LLAMA_POOLING_TYPE_NONE)) {
@@ -2905,6 +2924,9 @@ struct server_context_impl {
                         break;
                     }
 
+                    // KVFIX-114: drain one deferred task when this action ends
+                    const slot_action_drain drain_on_exit{queue_tasks, id_slot};
+
                     const int64_t t_start = ggml_time_us();
 
                     std::string filename = task.slot_action.filename;
@@ -2972,6 +2994,9 @@ struct server_context_impl {
                         break;
                     }
 
+                    // KVFIX-114: drain one deferred task when this action ends
+                    const slot_action_drain drain_on_exit{queue_tasks, id_slot};
+
                     const int64_t t_start = ggml_time_us();
 
                     std::string filename = task.slot_action.filename;
@@ -3055,6 +3080,9 @@ struct server_context_impl {
                         break;
                     }
 
+                    // KVFIX-114: drain one deferred task when this action ends
+                    const slot_action_drain drain_on_exit{queue_tasks, id_slot};
+
                     // Erase token cache
                     const size_t n_erased = slot->prompt.tokens.size();