vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-06 21:49:02 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix: reset -ngl to 0 when GPU is not used and reformat code

Browse source
This commit is contained in:
Lizonghang 2025-05-14 13:27:20 +04:00
parent b44187e3af
commit 2fbc0c8da3
4 changed files with 12 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
common/common.cpp
View file

@ -1527,6 +1527,12 @@ static bool assign_layers_to_device(
// //
struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
#if !(defined(GGML_USE_METAL) || defined(GGML_USE_CUDA))
// reset n_gpu_layers to 0 if GPU is not used
params.n_gpu_layers = 0;
#endif
llama_init_result iparams; llama_init_result iparams;
auto mparams = llama_model_params_from_gpt_params(params); auto mparams = llama_model_params_from_gpt_params(params);
@ -1582,6 +1588,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
if (n_world == 1) { if (n_world == 1) {
uint32_t n_layers = llama_model_n_layers(model); uint32_t n_layers = llama_model_n_layers(model);
// assign all layers to this device
params.n_layer_window[0] = n_layers; params.n_layer_window[0] = n_layers;
cparams.n_layer_window[0] = n_layers; cparams.n_layer_window[0] = n_layers;
mparams.n_layer_window[0] = n_layers; mparams.n_layer_window[0] = n_layers;
@ -1594,7 +1601,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
// broadcast startup args // broadcast startup args
struct startup_args args; struct startup_args args;
if (my_rank==0){ if (my_rank == 0){
args.should_profile = auto_schedule; args.should_profile = auto_schedule;
} }
llama_bcast_startup_args(lctx, my_rank, &args); llama_bcast_startup_args(lctx, my_rank, &args);

1
common/profiler.cpp
View file

@ -350,7 +350,6 @@ float device_inp_embd_delay(struct llama_model * model, enum ggml_type src0t, in
return 0.0f; return 0.0f;
} }
size_t QK_K = 0;
switch (src0t) { switch (src0t) {
case GGML_TYPE_F32: { case GGML_TYPE_F32: {
matrix_B = malloc(embd_size * sizeof(float)); matrix_B = malloc(embd_size * sizeof(float));

2
include/llama.h
View file

@ -453,7 +453,7 @@ extern "C" {
LLAMA_API void llama_free_sockets (struct llama_context * ctx, char ** msg); LLAMA_API void llama_free_sockets (struct llama_context * ctx, char ** msg);
LLAMA_API int llama_gather_device_info(struct llama_context * ctx, struct device_info * dev_info_set); LLAMA_API int llama_gather_device_info(struct llama_context * ctx, struct device_info * dev_info_set);
LLAMA_API int llama_send_device_info (struct llama_context * ctx, struct device_info * dev_info); LLAMA_API int llama_send_device_info (struct llama_context * ctx, struct device_info * dev_info);
LLAMA_API int llama_bcast_startup_args(struct llama_context * ctx, uint32_t rank, struct startup_args * args); LLAMA_API int llama_bcast_startup_args(struct llama_context * ctx, uint32_t rank, struct startup_args * args);
LLAMA_API int llama_bcast_layer_setup (struct llama_context * ctx, uint32_t * n_layer_window, uint32_t * n_gpu_layers); LLAMA_API int llama_bcast_layer_setup (struct llama_context * ctx, uint32_t * n_layer_window, uint32_t * n_gpu_layers);
LLAMA_API int llama_recv_layer_setup (struct llama_context * ctx, uint32_t * n_layer_window, uint32_t * n_gpu_layers); LLAMA_API int llama_recv_layer_setup (struct llama_context * ctx, uint32_t * n_layer_window, uint32_t * n_gpu_layers);

6
src/llama.cpp
View file

@ -20262,13 +20262,13 @@ int llama_send_device_info(struct llama_context * ctx, struct device_info * dev_
return 0; return 0;
} }
LLAMA_API int llama_bcast_startup_args(llama_context *ctx, uint32_t rank, startup_args *args) { int llama_bcast_startup_args(llama_context * ctx, uint32_t rank, startup_args * args) {
int32_t n_world = ctx->cparams.n_world; int32_t n_world = ctx->cparams.n_world;
if (n_world == 1) { if (n_world == 1) {
return 0; return 0;
} }
GGML_ASSERT(ctx != nullptr && ctx->send_socket != nullptr); GGML_ASSERT(ctx != nullptr && ctx->send_socket != nullptr);
if (rank==0){ if (rank == 0){
// send // send
try { try {
std::vector<zmq::message_t> send_msgs; std::vector<zmq::message_t> send_msgs;
@ -20289,7 +20289,7 @@ LLAMA_API int llama_bcast_startup_args(llama_context *ctx, uint32_t rank, startu
GGML_ASSERT(recv_msgs[1].size() == sizeof(bool)); GGML_ASSERT(recv_msgs[1].size() == sizeof(bool));
bool should_profile = *static_cast<bool*>(recv_msgs[1].data()); bool should_profile = *static_cast<bool*>(recv_msgs[1].data());
args->should_profile = should_profile; args->should_profile = should_profile;
if (rank != n_world-1){ if ((int)rank != (int)n_world - 1){
// send // send
try { try {
zmq::send_multipart(*ctx->send_socket, recv_msgs); zmq::send_multipart(*ctx->send_socket, recv_msgs);