vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-10 23:34:46 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

topo rebuild: add a delay to avoid packet interleaving

Browse source
This commit is contained in:
Li, Zonghang 2025-06-26 14:47:34 +04:00
parent 50807fd4e1
commit 729870fcd7
3 changed files with 40 additions and 21 deletions
Download patch file Download diff file Expand all files Collapse all files

2
common/common.cpp
View file

@ -1788,6 +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);
} else {
// use the user-defined n_layer_window
@ -1798,6 +1799,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
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);
} else {
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers);