diff --git a/Makefile b/Makefile
index f3bda7b9f..163971525 100644
--- a/Makefile
+++ b/Makefile
@@ -150,8 +150,8 @@ ifndef LLAMA_NO_ACCELERATE
 endif # LLAMA_NO_ACCELERATE
 
 ifdef LLAMA_MPI
-	CFLAGS += -DGGML_USE_MPI
-	CXXFLAGS += -DGGML_USE_MPI
+	CFLAGS += -DGGML_USE_MPI -Wno-cast-qual -Wno-int-to-void-pointer-cast -Wno-void-pointer-to-int-cast
+	CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp
index 2d913cebb..57a0fb7c5 100644
--- a/examples/simple/simple.cpp
+++ b/examples/simple/simple.cpp
@@ -173,6 +173,8 @@ int main(int argc, char ** argv)
     llama_free( ctx );
     llama_free_model( model );
 
+    llama_finalize_backend();
+
     return 0;
 }
 
diff --git a/llama.cpp b/llama.cpp
index c7de0bc60..a4435897e 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -1581,17 +1581,17 @@ static bool llama_eval_internal(
     // used at the end to optionally extract the embeddings
     struct ggml_tensor * embeddings = NULL;
 
-#ifdef GGML_USE_MPI
-    cur = ggml_send_tensor(ctx0, cur, (lctx.mpi_rank+1)%lctx.mpi_size);
-    ggml_set_name(cur, "send");
-#endif
+    if (lctx.mpi_size > 1) {
+        cur = ggml_send_tensor(ctx0, cur, (lctx.mpi_rank+1)%lctx.mpi_size);
+        ggml_set_name(cur, "send");
+    }
     if (lctx.mpi_rank == 0) {
-#ifdef GGML_USE_MPI
-        cur = ggml_recv_tensor(ctx0, cur,
-                ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N),
-                lctx.mpi_size-1);
-        ggml_set_name(cur, "recv");
-#endif
+        if (lctx.mpi_size > 1) {
+            cur = ggml_recv_tensor(ctx0, cur,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N),
+                    lctx.mpi_size-1);
+            ggml_set_name(cur, "recv");
+        }
         // norm
         {
             cur = ggml_rms_norm(ctx0, cur);