vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-05 20:20:39 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix some bugs

Browse source 
Signed-off-by: DeEMO <yzzxrx@gmail.com>
This commit is contained in:
DeEMO 2025-05-15 06:25:12 +00:00
parent cc46aa9828
commit 4b36aef157
2 changed files with 20 additions and 14 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
common/common.cpp
View file

@ -1547,7 +1547,7 @@ static bool tune_layer_allocation(
dev_infos_temp.clear();
n_layer_windows_temp.clear();
n_gpu_layers_temp.clear();
for(auto i=0; i<n_world; i++) {
for(uint32_t i=0; i<n_world; i++) {
if (n_layer_windows_[i] > 1 || i==0 ) {
dev_infos_temp.push_back(dev_infos_[i]);
n_layer_windows_temp.push_back(n_layer_windows_[i]);
@ -1561,7 +1561,7 @@ static bool tune_layer_allocation(
n_world = dev_infos_temp.size();
}
int i =0 , j =0;
uint32_t i =0 , j =0;
while(j < n_world) {
if(dev_infos[i].rank == dev_infos_temp[j].rank){
n_layer_window[i] = n_layer_windows_temp[j];
@ -1701,13 +1701,19 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers);
}
}
if(n_layer_window[my_rank]<=0){
LOG_INF("%s: info: rank %d has no layers to run, skipping\n", __func__, my_rank);
llama_free(lctx);
llama_free_model(model);
exit(0);
}
//update rank and n_world for consistency
uint32_t update_rank = 0;
uint32_t update_n_world = 1;
std::vector<uint32_t> n_layer_window_temp = {n_layer_window[0]};
std::vector<uint32_t> n_gpu_layers_temp = {n_gpu_layers[0]};
for(auto i=1; i<n_world; i++) {
for(uint32_t i=1; i<n_world; i++) {
if(n_layer_window[i] <= 0 ){
continue;
}
@ -1720,7 +1726,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
}
memset(n_layer_window, 0, n_world * sizeof(uint32_t));
memset(n_gpu_layers, 0, n_world * sizeof(uint32_t));
for (auto i=0; i<update_n_world; i++) {
for (uint32_t i=0; i<update_n_world; i++) {
n_layer_window[i] = n_layer_window_temp[i];
n_gpu_layers[i] = n_gpu_layers_temp[i];
}

20
src/llama.cpp
View file

@ -20263,7 +20263,7 @@ int llama_send_device_info(struct llama_context * ctx, struct device_info * dev_
}
LLAMA_API int llama_bcast_startup_args(llama_context *ctx, uint32_t rank, startup_args *args) {
int32_t n_world = ctx->cparams.n_world;
auto n_world = ctx->cparams.n_world;
if (n_world == 1) {
return 0;
}
@ -20343,14 +20343,14 @@ LLAMA_API int llama_rebuild_topo(llama_context *ctx,
}
dev_info_ptr = new device_info[n_world];
for (size_t i = 0; i < msgs.size(); i++) {
deserialize((const char *)msgs[i].data(), &dev_info_set[i]);
deserialize((const char *)msgs[i].data(), &dev_info_ptr[i]);
}
}else{
char * buffer = nullptr;
for(size_t i = 0; i < n_world; i++) {
size_t buffer_size = serialize(&dev_info_set[i], &buffer);
msgs.emplace_back(buffer, buffer_size);
free(buffer);
}
dev_info_ptr = dev_info_set;
@ -20361,9 +20361,9 @@ LLAMA_API int llama_rebuild_topo(llama_context *ctx,
// notify next rank
auto next_rank = (my_rank + 1) % n_world;
if(n_layer_window[next_rank] <= 0){
if(n_layer_window[next_rank] <= 0 && next_rank != 0){
try {
ctx->send_socket->setsockopt(ZMQ_LINGER, 3500);
ctx->send_socket->set(zmq::sockopt::linger, 3500);
zmq::send_multipart(*ctx->send_socket, msgs);
} catch (const zmq::error_t& e) {
LLAMA_LOG_INFO("Failed to send data: %s\n", e.what());
@ -20382,7 +20382,7 @@ LLAMA_API int llama_rebuild_topo(llama_context *ctx,
auto current_rank = my_rank;
while(next_rank!=my_rank){
if(n_layer_window[next_rank] > 0){
next_ip = dev_info_ptr[next_rank].next_ip;
next_ip = dev_info_ptr[current_rank].next_ip;
break;
}
next_rank = (next_rank + 1) % n_world;
@ -20402,6 +20402,9 @@ LLAMA_API int llama_rebuild_topo(llama_context *ctx,
}
return -1;
}
}else{
// only one node
ctx->next_node_ip = "";
}
}
if(!dev_info_set){
@ -20409,10 +20412,7 @@ LLAMA_API int llama_rebuild_topo(llama_context *ctx,
}
socket_to_close->close();
delete socket_to_close;
if(n_layer_window[my_rank]<=0){
exit(0);
}
return true;
return 0;
}
int llama_recv_layer_setup(struct llama_context * ctx, uint32_t * n_layer_window, uint32_t * n_gpu_layers) {