Skip to content

fix(sdk): default llama_cpp thread counts to llama.cpp's own value #1367

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

Draft
Mengsheng Wu (mengshengwu) wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

fix(sdk): default llama_cpp thread counts to llama.cpp's own value #1367

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 6 additions & 5 deletions sdk/plugins/llama_cpp/include/params.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,11 +33,12 @@ Device classify_device(const char* device_id, int n_gpu_layers);

// Map a caller's config to llama params, filling each unset (0) field from the
// plugin defaults. build_context_params resolves n_ctx (default differs between
// LLM at 4096 and VLM at 16384) plus n_batch / n_ubatch / n_seq_max, and picks
// thread counts via the per-(platform, device) rule (offloaded inference pins
// the upstream-fixed count; pure CPU uses cores/2). build_model_params only
// reads n_gpu_layers. Device selection and tensor-buffer overrides stay at the
// call site.
// LLM at 4096 and VLM at 16384) plus n_batch / n_ubatch / n_seq_max, and defaults
// n_threads / n_threads_batch to common_cpu_get_num_math() regardless of device —
// matching upstream llama-bench/llama-cli, which never vary thread count by
// backend. Callers may still override via ModelConfig's n_threads/n_threads_batch.
// build_model_params only reads n_gpu_layers. Device selection and tensor-buffer
// overrides stay at the call site.
llama_model_params build_model_params(const geniex_ModelConfig& config, Device device);
llama_context_params build_context_params(const geniex_ModelConfig& config, int32_t n_ctx_default, Device device);

Expand Down
31 changes: 14 additions & 17 deletions sdk/plugins/llama_cpp/src/params.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,12 +3,14 @@

#include "params.h"

#include <algorithm>
#include <cstring>
#include <fstream>
#include <sstream>
#include <string>
#include <thread>

#include "common.h"
#include "logging.h"

namespace geniex {
Expand Down Expand Up @@ -37,18 +39,7 @@ Device classify_device(const char* device_id, int n_gpu_layers) {
return Device::CPU; // UNKOWN treat as CPU
}

int resolve_n_threads(int requested, Device device) {
static const int cpu_threads = static_cast<int32_t>(std::thread::hardware_concurrency());
static const int cpu_matrix[3][3] = {
{cpu_threads, 6, 6}, // Linux
{cpu_threads, 6, 6}, // Windows
{6, 6, 6} // Android
};

if (requested > 0) return requested;

return cpu_matrix[static_cast<int>(kHostPlatform)][static_cast<int>(device)];
}
int resolve_n_threads(int requested) { return requested > 0 ? requested : common_cpu_get_num_math(); }

llama_model_params build_model_params(const geniex_ModelConfig& config, Device device) {
static const bool mmap_matrix[3][3] = {
Expand Down Expand Up @@ -89,8 +80,8 @@ llama_context_params build_context_params(const geniex_ModelConfig& config, int3
cpar.n_batch = config.n_batch > 0 ? config.n_batch : 2048;
cpar.n_ubatch = config.n_ubatch > 0 ? config.n_ubatch : ubatch;
cpar.n_seq_max = config.n_seq_max > 0 ? config.n_seq_max : 1;
cpar.n_threads = resolve_n_threads(config.n_threads, device);
cpar.n_threads_batch = resolve_n_threads(config.n_threads_batch, device);
cpar.n_threads = resolve_n_threads(config.n_threads);
cpar.n_threads_batch = resolve_n_threads(config.n_threads_batch);
cpar.flash_attn_type = static_cast<llama_flash_attn_type>(fa);
cpar.no_perf = false;

Expand Down Expand Up @@ -122,13 +113,19 @@ ggml_threadpool_params build_threadpool_params(int n_threads, Device device) {
{1000, 1000, 1000} // Android
};

int reserved_cores = 2;
uint32_t poll = poll_matrix[static_cast<int>(kHostPlatform)][static_cast<int>(device)];
bool pin = pin_matrix[static_cast<int>(kHostPlatform)][static_cast<int>(device)];
uint32_t poll = poll_matrix[static_cast<int>(kHostPlatform)][static_cast<int>(device)];
bool pin = pin_matrix[static_cast<int>(kHostPlatform)][static_cast<int>(device)];

ggml_threadpool_params tpp = ggml_threadpool_params_default(n_threads);

if (pin) {
// Leave up to 2 cores free for the OS/other work, but only if n_threads leaves
// that slack; n_threads can now reach hardware_concurrency() (see
// resolve_n_threads), and reserving cores on top of an already-full thread
// count would either pin past the real core count or oversubscribe the
// remaining ones.
int hw_threads = static_cast<int>(std::thread::hardware_concurrency());
int reserved_cores = hw_threads > 0 ? std::max(0, std::min(2, hw_threads - n_threads)) : 2;
for (int i = 0; i < n_threads; ++i) {
tpp.cpumask[reserved_cores + i] = true;
}
Expand Down
Loading