vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-10 08:04:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

communication: use barrier instead of manually adding delay

Browse source
This commit is contained in:
Li, Zonghang 2025-06-26 17:30:47 +04:00
parent 3f27a25340
commit a05022c05a
2 changed files with 95 additions and 86 deletions
Download patch file Download diff file Expand all files Collapse all files

22
common/common.cpp
View file

@ -1778,7 +1778,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
dev_info_set[0] = dev_info;
llama_gather_device_info(lctx, dev_info_set.data());
device_print_props(dev_info_set.data(), n_world, model, cparams);
device_print_props (dev_info_set.data(), n_world, model, cparams);
// assign layers to devices and remove weak devices
if (!assign_layers_and_select_devices(n_world, dev_info_set, n_layer_window, n_gpu_layers, model, cparams)) {
@ -1788,8 +1788,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
return iparams;
}
llama_bcast_layer_setup(lctx, n_layer_window, n_gpu_layers);
std::this_thread::sleep_for(std::chrono::milliseconds(200)); // add a delay to avoid packet interleaving
llama_rebuild_topo(lctx, n_layer_window, dev_info_set.data(), &node_type, is_forwarder);
llama_rebuild_topo (lctx, n_layer_window, dev_info_set.data(), &node_type, is_forwarder);
} else {
// use the user-defined n_layer_window
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), n_layer_window);
@ -1797,12 +1796,11 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
}
} else {
if (auto_schedule){
llama_send_device_info(lctx, &dev_info);
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers);
std::this_thread::sleep_for(std::chrono::milliseconds(200)); // add a delay to avoid packet interleaving
llama_rebuild_topo (lctx, n_layer_window, nullptr, &node_type, is_forwarder);
llama_send_device_info (lctx, &dev_info);
llama_recv_layer_setup (lctx, n_layer_window, n_gpu_layers);
llama_rebuild_topo (lctx, n_layer_window, nullptr, &node_type, is_forwarder);
} else {
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers);
llama_recv_layer_setup (lctx, n_layer_window, n_gpu_layers);
}
}
@ -1823,17 +1821,13 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
if (n_layer_window[i] <= 0 && is_forwarder[i] == 0) {
continue;
}
if (i <= my_rank) {
update_rank++;
}
if (i <= my_rank) update_rank++;
update_n_world++;
n_layer_window_temp.push_back(n_layer_window[i]);
n_gpu_layers_temp.push_back(n_gpu_layers[i]);
if (n_layer_window[i] > 0) {
if (i <= my_rank) {
worker_rank++;
}
if (i <= my_rank) worker_rank++;
n_worker++;
}
}