Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Numa #1556

Merged
merged 20 commits into from
Jun 26, 2023
Merged

Numa #1556

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
6fc5f17
detect NUMA systems and pin work threads to nodes (linux)
zrm May 21, 2023
0d23f8c
disable mmap prefetch/readahead for NUMA systems
zrm May 21, 2023
9d058c2
avoid sending finalize op to thread pool if it does nothing
zrm May 21, 2023
2c1b5ae
silence robot
zrm May 23, 2023
8502d51
fix args
zrm May 23, 2023
bf83dcb
make --numa a param
zrm Jun 17, 2023
b71dfe6
recommendation that n_nodes evenly divide n_threads did not warrant s…
zrm Jun 17, 2023
adaad10
lower synchronization overhead
zrm Jun 18, 2023
c31d51d
statically allocate
zrm Jun 18, 2023
2f5bb46
move numa state to g_state
zrm Jun 18, 2023
4b94582
add description for --numa
zrm Jun 18, 2023
90a0e65
Merge branch 'master' into HEAD
ggerganov Jun 19, 2023
d0e3596
ggml : minor style changes
ggerganov Jun 19, 2023
67ba34e
ggml : minor style + try fix sanitizer build
ggerganov Jun 19, 2023
8f98035
Merge branch 'master' into HEAD
ggerganov Jun 26, 2023
0fe4b00
llama : allow to initialize backend with NUMA support
ggerganov Jun 26, 2023
875a1e1
llama : avoid ggml include in llama-util.h
ggerganov Jun 26, 2023
4a555b4
ggml : style / formatting
ggerganov Jun 26, 2023
81a40e9
ggml : fix handling of ops with n_threads > n_tasks > 1
ggerganov Jun 26, 2023
9aec2b7
server : utilize numa parameter
ggerganov Jun 26, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
llama : avoid ggml include in llama-util.h
  • Loading branch information
@ggerganov
ggerganov committed Jun 26, 2023
commit 875a1e111eaa9db3fd51be8c3b3288291ec2f1d2
18 changes: 10 additions & 8 deletions llama-util.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,8 +16,6 @@
#include <vector>
#include <stdexcept>

#include "ggml.h"

#ifdef __has_include
#if __has_include(<unistd.h>)
#include <unistd.h>
Expand Down Expand Up @@ -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; }
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes again. I want to understand why prefetch hurts performance with NUMA node ? (with paged memory or pin memory)

#ifdef __linux__
if (prefetch) { flags |= MAP_POPULATE; }
#endif
Expand All @@ -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)) {
Expand All @@ -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));
Expand Down Expand Up @@ -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
Expand Down
4 changes: 2 additions & 2 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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);
}
Expand Down Expand Up @@ -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()));
}
}

Expand Down