From b163918b46ed67db86a8346a6bd8a4c2f31b1e32 Mon Sep 17 00:00:00 2001
From: Lizonghang <870644199@qq.com>
Date: Wed, 12 Feb 2025 00:17:33 +0400
Subject: [PATCH] disable prefetch in standalone mode

---
 src/llama.cpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/llama.cpp b/src/llama.cpp
index 9f613743..c41ebb8c 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -18108,9 +18108,9 @@ static int llama_decode_internal(
                 timer(manage_graph_tensors);
                 
                 int next_gf_id = (i + 1) % gf.size();
-                manage_graph_tensors(gf[next_gf_id], POSIX_MADV_WILLNEED, true);
+                manage_graph_tensors(gf[next_gf_id], POSIX_MADV_WILLNEED, n_world > 1);
                 if (my_rank == 0 && (is_last_l || (next_gf_id == (int)gf.size() - 1))) {
-                    manage_graph_tensors(gf[0], POSIX_MADV_WILLNEED, true);
+                    manage_graph_tensors(gf[0], POSIX_MADV_WILLNEED, n_world > 1);
                 }
 
                 if (cparams.unload && n_world > 1) {