server: improve slots scheduling for n_cmpl (#18789)

* server : make sure children tasks are scheduled to launch with parent

* fix

* add comment pointing to this PR

* fix

* clean up

* more debug messages

* add pop_deferred_task with specific ID version

* improve the logic

* simple approach

* no double move

* correct return type of launch_slots_with_parent_task

tools/server/server-task.h

@@ -121,8 +121,10 @@ struct server_task {
     int id_slot   = -1;
 
     // used by parallel sampling (multiple completions from same prompt)
-    int n_children =  0; // number of tasks reusing this prompt
     int id_parent  = -1;
+    // temporary store of child tasks for scheduling
+    // note: accessing to elements is invalid after the task is moved to server_slot
+    std::vector<server_task> child_tasks;
 
     // used by SERVER_TASK_TYPE_INFERENCE
     task_params   params;
@@ -197,11 +199,14 @@ struct server_task {
         std::unordered_set<int> ids(tasks.size());
         for (size_t i = 0; i < tasks.size(); i++) {
             ids.insert(tasks[i].id);
+            for (auto & child : tasks[i].child_tasks) {
+                ids.insert(child.id);
+            }
         }
         return ids;
     }
 
-    server_task create_child(int id_parent, int id_child) const {
+    void add_child(int id_parent, int id_child) {
         server_task copy;
 
         copy.id        = id_child;
@@ -209,8 +214,15 @@ struct server_task {
         copy.params    = params;
         copy.type      = type;
         copy.tokens    = tokens.clone();
+        copy.id_slot   = -1; // child tasks cannot specify slot
 
-        return copy;
+        // use different sampling seed for each child
+        // note: https://github.com/ggml-org/llama.cpp/pull/18700#discussion_r2675115723
+        if (copy.params.sampling.seed != LLAMA_DEFAULT_SEED) {
+            copy.params.sampling.seed += (uint32_t)child_tasks.size() + 1;
+        }
+
+        child_tasks.push_back(std::move(copy));
     }
 
     // the task will be moved into queue, then onto slots
@@ -218,6 +230,14 @@ struct server_task {
     task_result_state create_state() const {
         return task_result_state(params.oaicompat_chat_syntax);
     }
+
+    bool is_parent() const {
+        return child_tasks.size() > 0;
+    }
+
+    bool is_child() const {
+        return id_parent != -1;
+    }
 };
 
 struct result_timings {