CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

tests : fix test-thread-safety when compiling with multiple backends (#16699)

Browse Source 
* run one test per backend/device (even if it's the same device)
This commit is contained in:
Acly
2025-10-22 12:01:22 +02:00
committed by GitHub
parent 9285325ce0
commit d8eaa26e4d
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/test-thread-safety.cpp
View File

@@ -3,6 +3,7 @@
// - Creates n_parallel (--parallel) contexts per model
// - Runs inference in parallel on each context
#include <array>
#include <thread>
#include <vector>
#include <atomic>
@@ -38,13 +39,14 @@ int main(int argc, char ** argv) {
cparams.n_seq_max = 1;
int dev_count = ggml_backend_dev_count();
int gpu_dev_count = 0;
std::vector<std::array<ggml_backend_dev_t, 2>> gpus;
for (int i = 0; i < dev_count; ++i) {
auto * dev = ggml_backend_dev_get(i);
if (dev && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
gpu_dev_count++;
gpus.push_back({dev, nullptr});
}
}
const int gpu_dev_count = (int)gpus.size();
const int num_models = gpu_dev_count + 1 + 1; // GPUs + 1 CPU model + 1 layer split
//const int num_models = std::max(1, gpu_dev_count);
const int num_contexts = std::max(1, params.n_parallel);
@@ -58,12 +60,12 @@ int main(int argc, char ** argv) {
if (m < gpu_dev_count) {
mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
mparams.main_gpu = m;
mparams.devices = gpus[m].data();
} else if (m == gpu_dev_count) {
mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
mparams.main_gpu = -1; // CPU model
} else {
mparams.split_mode = LLAMA_SPLIT_MODE_LAYER;;
mparams.split_mode = LLAMA_SPLIT_MODE_LAYER;
}
llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);