diff --git a/ggml.c b/ggml.c index f6dca255f..92c10e995 100644 --- a/ggml.c +++ b/ggml.c @@ -17398,8 +17398,12 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * } for (int i = 0; i < GGML_MAX_SRC; ++i) { - if (node->src[i]) { - ggml_visit_parents(cgraph, node->src[i]); + const int k = + (cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i : + (cgraph->order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? (GGML_MAX_SRC-1-i) : + /* unknown order, just fall back to using i*/ i; + if (node->src[k]) { + ggml_visit_parents(cgraph, node->src[k]); } } @@ -17458,6 +17462,7 @@ struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) { /*.grads =*/ { NULL }, /*.leafs =*/ { NULL }, /*.hash_table =*/ { NULL }, + /*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT, /*.perf_runs =*/ 0, /*.perf_cycles =*/ 0, /*.perf_time_us =*/ 0, @@ -17531,6 +17536,7 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { /*.grads =*/ { NULL }, /*.leafs =*/ { NULL }, /*.hash_table =*/ { NULL }, + /*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT, /*.perf_runs =*/ 0, /*.perf_cycles =*/ 0, /*.perf_time_us =*/ 0, diff --git a/ggml.h b/ggml.h index 20e67f4fd..8eafa67a0 100644 --- a/ggml.h +++ b/ggml.h @@ -516,6 +516,12 @@ extern "C" { // #define GGML_GRAPH_HASHTABLE_SIZE 16411 #define GGML_GRAPH_HASHTABLE_SIZE 32771 + enum ggml_cgraph_eval_order { + GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0, + GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT, + GGML_CGRAPH_EVAL_ORDER_COUNT + }; + // computation graph struct ggml_cgraph { int n_nodes; @@ -527,6 +533,8 @@ extern "C" { void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE]; + enum ggml_cgraph_eval_order order; + // performance int perf_runs; int64_t perf_cycles;