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

WA to fix TTFT in CB for Linux #1557

Merged
merged 4 commits into from
Jan 16, 2025
Merged

WA to fix TTFT in CB for Linux #1557

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
b4b528c
Use mmap for mem allocation.
popovaan Jan 15, 2025
8b3a789
Minor correction.
popovaan Jan 15, 2025
1e656e5
Minor fix.
popovaan Jan 15, 2025
b871ace
Added custom allocator.
popovaan Jan 16, 2025
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
24 changes: 21 additions & 3 deletions src/cpp/src/cache_manager.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,11 +5,14 @@

#include <vector>
#include <list>

#include "openvino/runtime/tensor.hpp"

#include "device_config.hpp"

#ifndef _WIN32
#include <sys/mman.h>
#include "openvino/core/shape.hpp"
#endif

namespace ov::genai {
class CacheManager {
DeviceConfig m_device_config;
Expand Down Expand Up @@ -56,8 +59,22 @@ class CacheManager {
for (size_t decoder_layer_id = 0; decoder_layer_id < m_device_config.get_num_layers(); ++decoder_layer_id) {
ov::Shape value_cache_shape = set_first_dim_and_make_static(m_device_config.get_value_cache_shape(decoder_layer_id), num_kv_blocks);
ov::Shape key_cache_shape = set_first_dim_and_make_static(m_device_config.get_key_cache_shape(decoder_layer_id), num_kv_blocks);
#ifdef _WIN32
ov::Tensor key_cache(m_device_config.get_cache_precision(), key_cache_shape);
ov::Tensor value_cache(m_device_config.get_cache_precision(), value_cache_shape);
#else
auto key_size = ov::shape_size(key_cache_shape) * m_device_config.get_cache_precision().size();
auto value_size = ov::shape_size(value_cache_shape) * m_device_config.get_cache_precision().size();

// allocate memory filled with zeros
auto key_ptr = mmap(NULL, key_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
auto value_ptr = mmap(NULL, value_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
OPENVINO_ASSERT(key_ptr != MAP_FAILED);
OPENVINO_ASSERT(value_ptr != MAP_FAILED);
ov::Tensor key_cache = ov::Tensor(m_device_config.get_cache_precision(), key_cache_shape, key_ptr);
ov::Tensor value_cache = ov::Tensor(m_device_config.get_cache_precision(), value_cache_shape, value_ptr);
ilya-lavrenov marked this conversation as resolved.
Show resolved Hide resolved

#endif

auto key_cache_roi_end = static_cast<unsigned char*>(key_cache.data());
auto value_cache_roi_end = static_cast<unsigned char*>(value_cache.data());
Expand All @@ -82,13 +99,14 @@ class CacheManager {

}

#ifdef _WIN32
// Some optimizations like AVX2, AVX512, AMX require a minimal shape and
// perform multiplying by zero on the excess data. Uninitialized tensor data contain NAN's,
// so NAN * 0 returns non-zero invalid data.
// So we need to set zeros to all newly allocated tensors data.
std::memset(key_cache_roi_end, 0, key_cache.get_byte_size() - key_roi_size_byte);
std::memset(value_cache_roi_end, 0, value_cache.get_byte_size() - value_roi_size_byte);
#endif
// set new cache tensors
if (m_key_cache.size() > decoder_layer_id) {
m_key_cache[decoder_layer_id] = key_cache;
Expand Down
Loading