vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

use posix_fadvise64(.., POSIX_FADV_SEQUENTIAL) to improve performance with mmap

Browse source
This commit is contained in:
slaren 2023-12-20 12:57:45 +01:00
parent c3678ca84f
commit 5241045819
1 changed files with 8 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

9
llama.cpp
View file

@ -33,6 +33,7 @@
#include <unistd.h> #include <unistd.h>
#if defined(_POSIX_MAPPED_FILES) #if defined(_POSIX_MAPPED_FILES)
#include <sys/mman.h> #include <sys/mman.h>
#include <fcntl.h>
#endif #endif
#if defined(_POSIX_MEMLOCK_RANGE) #if defined(_POSIX_MEMLOCK_RANGE)
#include <sys/resource.h> #include <sys/resource.h>
@ -840,6 +841,10 @@ struct llama_mmap {
// prefetch/readahead impairs performance on NUMA systems // prefetch/readahead impairs performance on NUMA systems
if (numa) { prefetch = 0; } if (numa) { prefetch = 0; }
#ifdef __linux__ #ifdef __linux__
if (posix_fadvise64(fd, 0, file->size, POSIX_FADV_SEQUENTIAL)) {
fprintf(stderr, "warning: fadvise(.., POSIX_FADV_SEQUENTIAL) failed: %s\n",
strerror(errno));
}
if (prefetch) { flags |= MAP_POPULATE; } if (prefetch) { flags |= MAP_POPULATE; }
#endif #endif
addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0); addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
@ -2314,7 +2319,9 @@ struct llama_model_loader {
} }
*/ */
// prefetch the whole file - all the data is needed anyway // prefetch the whole file - all the data is needed anyway
mapping.reset(new llama_mmap(&file, -1, ggml_is_numa())); if (use_mmap) {
mapping.reset(new llama_mmap(&file, -1, ggml_is_numa()));
}
} }
// for backwards compatibility only // for backwards compatibility only