Fixed mmap prefetch for GPU offloading (#2529)

diff --git a/llama-util.h b/llama-util.h
index 3fc03ce28273ed4ed2a09647160152c68e31ea8b..6e9e39ddb6f5804e7dd706404c19d8a50d0f062f 100644 (file)
--- 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 839739870eb3e97a72fd56e68f4377a3639ac4b9..39aefd499dd0c50cc9b52361bd8b7d652b837e12 100644 (file)
--- 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;
             }
         }