llama: propagating the results of graph_compute to the user interface

Author: Xarbirus

src/llama.cpp

@@ -16612,7 +16612,7 @@ static void llama_output_reorder(struct llama_context * ctx) {
     }
 }
 
-static void llama_graph_compute(
+static enum ggml_status llama_graph_compute(
           llama_context & lctx,
             ggml_cgraph * gf,
                     int   n_threads,
@@ -16634,9 +16634,11 @@ static void llama_graph_compute(
     }
 #endif
 
-    ggml_backend_sched_graph_compute_async(lctx.sched, gf);
+    auto status = ggml_backend_sched_graph_compute_async(lctx.sched, gf);
 
     // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
+
+    return status;
 }
 
 // decode a batch of tokens by evaluating the transformer
@@ -16818,7 +16820,18 @@ static int llama_decode_internal(
 
         llama_set_inputs(lctx, ubatch);
 
-        llama_graph_compute(lctx, gf, n_threads, threadpool);
+        const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool);
+        switch (compute_status) {
+            case GGML_STATUS_SUCCESS:
+                break;
+            case GGML_STATUS_ABORTED:
+                return 2;
+            case GGML_STATUS_ALLOC_FAILED:
+                return -2;
+            case GGML_STATUS_FAILED:
+            default:
+                return -3;
+        }
 
         // update the kv ring buffer
         {
@@ -17038,7 +17051,18 @@ static int llama_encode_internal(
 
     llama_set_inputs(lctx, ubatch);
 
-    llama_graph_compute(lctx, gf, n_threads, threadpool);
+    const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool);
+    switch (compute_status) {
+        case GGML_STATUS_SUCCESS:
+            break;
+        case GGML_STATUS_ABORTED:
+            return 2;
+        case GGML_STATUS_ALLOC_FAILED:
+            return -2;
+        case GGML_STATUS_FAILED:
+        default:
+            return -3;
+    }
 
     // extract embeddings
     if (embd) {