Skip to content

Revert the LLM identification method to the previous version #33006

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

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
+19 −8
Open

Revert the LLM identification method to the previous version #33006

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8e4ca11
Revert the LLM identification method to the previous version
wangleis Nov 24, 2025
191ad6a
update for comment
wangleis Nov 26, 2025
f460262
update logic
wangleis Nov 26, 2025
28467c9
update logic
wangleis Nov 26, 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
27 changes: 19 additions & 8 deletions src/plugins/intel_cpu/src/cpu_streams_calculation.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
#include "openvino/core/any.hpp"
#include "openvino/core/except.hpp"
#include "openvino/core/model.hpp"
#include "openvino/core/type.hpp"
#include "openvino/runtime/intel_cpu/properties.hpp"
#include "openvino/runtime/properties.hpp"
#include "openvino/runtime/system_conf.hpp"
Expand All @@ -33,8 +34,10 @@
#endif
#include "cpu_map_scheduling.hpp"
#include "openvino/op/fake_quantize.hpp"
#include "openvino/op/paged_attention.hpp"
#include "openvino/runtime/threading/cpu_streams_info.hpp"
#include "openvino/runtime/threading/istreams_executor.hpp"
#include "transformations/cpu_opset/common/op/sdpa.hpp"
#include "transformations/utils/utils.hpp"
#include "utils/general_utils.h"

Expand Down Expand Up @@ -609,20 +612,28 @@ int get_model_prefer_threads(const int num_streams,
const std::shared_ptr<ov::Model>& model,
Config& config) {
bool int8_intensive = ov::op::util::has_op_with_type<ov::op::v0::FakeQuantize>(model);
auto is_paged_attention_model = false;
const auto is_LLM =
ov::op::util::is_large_language_model(*model, [&is_paged_attention_model](std::shared_ptr<ov::Node> node) {
if (ov::is_type<ov::op::PagedAttentionExtension>(node)) {
is_paged_attention_model = true;
return true;
} else if (ov::is_type<ov::intel_cpu::ScaledDotProductAttentionWithKVCache>(node)) {
return true;
}

return false;
});

auto default_prefer_threads_latency = [&]() {
bool llm_related = ov::op::util::is_large_language_model(*model);
const int int8_threshold = 4; // ~relative efficiency of the VNNI-intensive code for Big vs Little cores;
const int fp32_threshold = 2; // ~relative efficiency of the AVX2 fp32 code for Big vs Little cores;
// By default the latency case uses (faster) Big cores only, depending on the compute ratio
// But on MTL detected by ov::get_number_of_blocked_cores(), use Big and Little cores together in Big
// cores only cases except LLM.

bool use_all_cores =
proc_type_table[0][MAIN_CORE_PROC] <= (proc_type_table[0][EFFICIENT_CORE_PROC] /
(int8_intensive || llm_related ? int8_threshold : fp32_threshold));
bool use_big_and_little = !llm_related && (ov::get_number_of_blocked_cores() != 0);
(int8_intensive ? int8_threshold : fp32_threshold));

if (use_all_cores || use_big_and_little) {
if (use_all_cores && !is_LLM) {
config.modelPreferThreadsLatency =
proc_type_table[0][MAIN_CORE_PROC] + proc_type_table[0][EFFICIENT_CORE_PROC];
} else {
Expand Down Expand Up @@ -720,7 +731,7 @@ int get_model_prefer_threads(const int num_streams,
if ((proc_type_table[0][MAIN_CORE_PROC] < config.threads || config.threads == 0) &&
(ov::get_number_of_blocked_cores() || proc_type_table[0][LP_EFFICIENT_CORE_PROC] > 0) &&
proc_type_table[0][EFFICIENT_CORE_PROC] <= 2 * proc_type_table[0][MAIN_CORE_PROC]) {
if (ov::op::util::is_large_language_model(*model)) {
if (is_LLM) {
config.modelPreferThreadsLatency = proc_type_table[0][MAIN_CORE_PROC];
} else {
config.modelPreferThreadsLatency =
Expand Down
Loading