From 844a8b0bfa58d093b30fddd5b5387cafaa340f61 Mon Sep 17 00:00:00 2001
From: Kawrakow <iwankawrakow@gmail.com>
Date: Sat, 13 Dec 2025 18:40:49 +0100
Subject: [PATCH] Fix sync logic (#1064)

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
---
 ggml/src/ggml-backend.cpp | 21 ++++++++-------------
 1 file changed, 8 insertions(+), 13 deletions(-)

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 9ef38485..4006d6f5 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -1892,13 +1892,10 @@ static void ggml_backend_sched_copy_inputs(ggml_backend_sched_t sched, ggml_back
 
         if (input->flags & GGML_TENSOR_FLAG_INPUT) {
             // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
-            if (needs_sync[split_backend_id]) {
-                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                } else {
-                    ggml_backend_synchronize(split_backend);
-                }
-                //needs_sync[split_backend_id] = false;
+            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+            } else {
+                ggml_backend_synchronize(split_backend);
             }
             ggml_backend_tensor_copy(input, input_cpy);
         } else {
@@ -1909,7 +1906,7 @@ static void ggml_backend_sched_copy_inputs(ggml_backend_sched_t sched, ggml_back
                 } else {
                     ggml_backend_synchronize(split_backend);
                 }
-                //needs_sync[split_backend_id] = false;
+                needs_sync[split_backend_id] = false;
             }
 
             ggml_tensor * node = split->graph.nodes[0];
@@ -1923,7 +1920,6 @@ static void ggml_backend_sched_copy_inputs(ggml_backend_sched_t sched, ggml_back
                     last_input_backend = input_backend;
                 }
 
-                //printf("node: %s have %d inputs, processing input %d\n", node->name, split->n_inputs, j);
                 ggml_tensor * ids_tensor = node->op == GGML_OP_MUL_MAT_ID ? node->src[2] : node->src[3];
                 auto ids_backend = split_backend;
 
@@ -1945,7 +1941,7 @@ static void ggml_backend_sched_copy_inputs(ggml_backend_sched_t sched, ggml_back
                     ggml_backend_tensor_get_async(ids_backend, ids_tensor, ids.data(), 0, ggml_nbytes(ids_tensor));
 
                     ggml_backend_synchronize(ids_backend);
-                    //needs_sync[tensor_backend_id(ids_tensor)] = false;
+                    needs_sync[tensor_backend_id(ids_tensor)] = false;
 
                     unique_ids.resize((n_expert + 31)/32);
                     std::memset(unique_ids.data(), 0, unique_ids.size()*sizeof(uint32_t));
@@ -2005,7 +2001,7 @@ static void ggml_backend_sched_copy_inputs(ggml_backend_sched_t sched, ggml_back
                     int input_backend_id = tensor_backend_id(input);
                     if (needs_sync[input_backend_id]) {
                         ggml_backend_synchronize(input_backend);
-                        //needs_sync[input_backend_id] = false;
+                        needs_sync[input_backend_id] = false;
                     }
                     if (needs_sync[split_backend_id]) {
                         if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
@@ -2013,7 +2009,7 @@ static void ggml_backend_sched_copy_inputs(ggml_backend_sched_t sched, ggml_back
                         } else {
                             ggml_backend_synchronize(split_backend);
                         }
-                        //needs_sync[split_backend_id] = false;
+                        needs_sync[split_backend_id] = false;
                     }
                     ggml_backend_tensor_copy(input, input_cpy);
                 }
@@ -2034,7 +2030,6 @@ static ggml_status ggml_backend_sched_compute_splits_sm_graph(ggml_backend_sched
     for (int i = 0; i < sched->n_splits; ++i) {
         auto split_i = &splits[i];
         this_split.clear();
-        //auto& this_split = all_splits.emplace_back();
         this_split.push_back(split_i);
         for (int j = i+1; j < sched->n_splits; ++j) {
             auto split_j = &splits[j];