From 3d9a55181603e85a26378a850a14068034e5002d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= <johannesg@5d6.de>
Date: Mon, 7 Aug 2023 10:09:40 +0200
Subject: [PATCH] Fixed mmap prefetch for GPU offloading (#2529)

---
 llama-util.h | 2 +-
 llama.cpp    | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/llama-util.h b/llama-util.h
index 3fc03ce..6e9e39d 100644
--- a/llama-util.h
+++ b/llama-util.h
@@ -219,7 +219,7 @@ struct llama_mmap {
         // prefetch/readahead impairs performance on NUMA systems
         if (numa) { prefetch = 0; }
 #ifdef __linux__
-        if (prefetch) { flags |= MAP_POPULATE; }
+        if (prefetch >= file->size) { flags |= MAP_POPULATE; }
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
diff --git a/llama.cpp b/llama.cpp
index 8397398..39aefd4 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -747,12 +747,12 @@ struct llama_model_loader {
 
     void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
         size_t data_size = 0;
-        size_t prefetch_size = 0;
+        size_t prefetch_size = file_loader->file.size;
         size_t lock_size = 0;
         for (const llama_load_tensor & lt : tensors_map.tensors) {
             data_size += lt.size;
-            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
-                prefetch_size += lt.size;
+            if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) {
+                prefetch_size -= lt.size;
             }
         }