Be able to force synchronization when copying graph ginputs

ggml/CMakeLists.txt

@@ -108,6 +108,7 @@ endif()
 # ggml core
 set(GGML_SCHED_MAX_COPIES  "1" CACHE STRING "ggml: max input copies for pipeline parallelism")
 set(GGML_MAX_CONTEXTS       "" CACHE STRING "ggml: max model contexts (override only; defaults to 64 in the code)")
+set(GGML_FORCE_SYNC         "" CACHE STRING "ggml-backend: control device copy synchronization")
 
 # 3rd party libs / backends
 option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)

ggml/src/CMakeLists.txt

@@ -1462,6 +1462,9 @@ add_library(ggml
 if (EMSCRIPTEN)
     set_target_properties(ggml PROPERTIES COMPILE_FLAGS "-msimd128")
 endif()
+if (GGML_FORCE_SYNC)
+    set_source_files_properties(ggml-backend.cpp PROPERTIES COMPILE_FLAGS -DGGML_FORCE_SYNC)
+endif()
 
 target_compile_definitions(ggml PUBLIC  ${GGML_CDEF_PUBLIC})
 target_compile_definitions(ggml PRIVATE $<$<CONFIG:Release>:NDEBUG>)

ggml/src/ggml-backend.cpp

@@ -1974,7 +1974,11 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 static void ggml_backend_sched_copy_inputs(ggml_backend_sched_t sched, ggml_backend_sched_split * split, std::array<bool, GGML_SCHED_MAX_BACKENDS> & needs_sync,
         std::vector<int32_t> & ids, std::vector<uint32_t> & unique_ids, ggml_tensor * last_ids_tensor) {
     if (split->n_inputs < 1) return;
+#ifdef GGML_FORCE_SYNC
+    constexpr bool k_set_sync = true;
+#else
     constexpr bool k_set_sync = false;
+#endif
     int split_backend_id = split->backend_id;
     ggml_backend_t split_backend = sched->backends[split_backend_id];
     ggml_backend_t last_input_backend = nullptr;