Playing games with the scheduler

This change tricks it into doing the right thing^TM.
Still quite a bit slower than split mode layer for the 8B LlaMA model.
But for the 70B LlaMA it now beats split mode layer for TG:
28 t/s vs 24.4 t/s. PP is 627 t/s vs 744 t/s.
In comparison, split mode "row" in mainline gets
484 t/s PP and 19.3 t/s TG.

ggml/src/ggml-backend.cpp

@@ -1632,7 +1632,10 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;
-            if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
+            if (node->op == GGML_OP_ADD && node->op_params[0] == 0xff) {
+                need_new_split = true;
+            }
+            else if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
                 for (int j = 0; j < GGML_MAX_SRC; j++) {
                     struct ggml_tensor * src = node->src[j];
                     if (src == NULL) {

src/llama-build-context.cpp

@@ -680,6 +680,7 @@ ggml_tensor * llm_build_context::llm_build_ffn(
             cur = ggml_add(ctx, cur, ffn[id]);
             cb(cur, "combine_ffn", il);
         }
+        cur->op_params[0] = 0xff;
         return cur;
     }
 
@@ -9088,6 +9089,7 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
                 cur = ggml_add(ctx0, cur, attn[id]);
                 cb(cur, "combine_attn", il);
             }
+            cur->op_params[0] = 0xff;
             return cur;
         }
     }