From 875a1e111eaa9db3fd51be8c3b3288291ec2f1d2 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Mon, 26 Jun 2023 20:27:24 +0300
Subject: [PATCH] llama : avoid ggml include in llama-util.h

---
 llama-util.h | 18 ++++++++++--------
 llama.cpp    |  4 ++--
 2 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/llama-util.h b/llama-util.h
index d709319df..042ebe43c 100644
--- a/llama-util.h
+++ b/llama-util.h
@@ -16,8 +16,6 @@
 #include <vector>
 #include <stdexcept>
 
-#include "ggml.h"
-
 #ifdef __has_include
     #if __has_include(<unistd.h>)
         #include <unistd.h>
@@ -174,12 +172,12 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */) {
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
         // prefetch/readahead impairs performance on NUMA systems
-        if (ggml_is_numa()) { prefetch = 0; }
+        if (numa) { prefetch = 0; }
 #ifdef __linux__
         if (prefetch) { flags |= MAP_POPULATE; }
 #endif
@@ -195,7 +193,7 @@ struct llama_mmap {
                         strerror(errno));
             }
         }
-        if (ggml_is_numa()) {
+        if (numa) {
             // advise the kernel not to use readahead
             // (because the next page might not belong on the same node)
             if (madvise(addr, file->size, MADV_RANDOM)) {
@@ -211,7 +209,9 @@ struct llama_mmap {
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true) {
+    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
+        (void) numa;
+
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
@@ -256,8 +256,10 @@ struct llama_mmap {
 #else
     static constexpr bool SUPPORTED = false;
 
-    llama_mmap(struct llama_file *, bool prefetch = true) {
-        (void)prefetch;
+    llama_mmap(struct llama_file *, bool prefetch = true, bool numa = false) {
+        (void) prefetch;
+        (void) numa;
+
         throw std::runtime_error(std::string("mmap not supported"));
     }
 #endif
diff --git a/llama.cpp b/llama.cpp
index e932636fc..1a15844bc 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -774,7 +774,7 @@ struct llama_model_loader {
         }
 
         if (use_mmap) {
-            mapping.reset(new llama_mmap(&file_loaders.at(0)->file, prefetch_size));
+            mapping.reset(new llama_mmap(&file_loaders.at(0)->file, prefetch_size, ggml_is_numa()));
             if (lmlock) {
                 lmlock->init(mapping->addr);
             }
@@ -2903,7 +2903,7 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
 
         // maybe this should in llama_model_loader
         if (model_loader->use_mmap) {
-            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ 0));
+            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ 0, ggml_is_numa()));
         }
     }