vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-10 14:04:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix manully set layer window and gpu layers

Browse source
This commit is contained in:
Lizonghang 2025-01-18 16:44:26 +04:00
parent 70bd9db008
commit f9d16fbf71
1 changed files with 40 additions and 40 deletions
Download patch file Download diff file Expand all files Collapse all files

80
common/common.cpp
View file

@ -1408,63 +1408,63 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
} }
} }
// get device profile
LOG_INF("Start profiling this device, this may take some seconds ...\n");
device_info dev_info; device_info dev_info;
dev_info.rank = params.rank; uint32_t n_world = params.n_world;
llama_profile_device(&dev_info, model, ml, params.cuda_mem, params.n_predict, params.n_ctx, params.cpuparams.n_threads, params.flash_attn); uint32_t my_rank = params.rank;
bool auto_schedule = n_world == 1 || params.n_layer_window[0] == 0;
if (auto_schedule) {
// get device profile
LOG_INF("Start profiling this device, this may take some seconds ...\n");
dev_info.rank = params.rank;
llama_profile_device(&dev_info, model, ml, params.cuda_mem, params.n_predict, params.n_ctx, params.cpuparams.n_threads, params.flash_attn);
}
// create llama context // create llama context
struct llama_context_params cparams = llama_context_params_from_gpt_params(params); struct llama_context_params cparams = llama_context_params_from_gpt_params(params);
llama_context * lctx = llama_new_context_with_model(model, cparams); llama_context * lctx = llama_new_context_with_model(model, cparams);
uint32_t n_world = cparams.n_world;
uint32_t my_rank = cparams.rank;
// initialize sockets // initialize sockets
llama_init_sockets(lctx, n_world, my_rank); llama_init_sockets(lctx, n_world, my_rank);
// sychronize device profile to the master node if (auto_schedule) {
struct device_info * dev_info_set = nullptr; // sychronize device profile to the master node
if (my_rank == 0) { struct device_info * dev_info_set = nullptr;
dev_info_set = (struct device_info *)malloc(n_world * sizeof(struct device_info)); if (my_rank == 0) {
dev_info_set[0] = dev_info; dev_info_set = (struct device_info *)malloc(n_world * sizeof(struct device_info));
llama_gather_device_info(lctx, dev_info_set); dev_info_set[0] = dev_info;
} else { llama_gather_device_info(lctx, dev_info_set);
llama_send_device_info(lctx, &dev_info);
}
uint32_t n_layer_window[32] = {0}, n_gpu_layers[32] = {0};
if (my_rank == 0) {
if (n_world == 1 || params.n_layer_window[0] == 0) {
// automatically determine n_layer_window and n_gpu_layers
assign_device(n_world, my_rank, dev_info_set, n_layer_window, n_gpu_layers, model, cparams);
} else { } else {
// use manually set n_layer_window llama_send_device_info(lctx, &dev_info);
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), n_layer_window);
} }
// synchronize the new n_layer_window and n_gpu_layers to other nodes uint32_t n_layer_window[32] = {0}, n_gpu_layers[32] = {0};
llama_bcast_layer_setup(lctx, n_layer_window, n_gpu_layers); if (my_rank == 0) {
} else { // automatically determine n_layer_window and n_gpu_layers
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers); assign_device(n_world, my_rank, dev_info_set, n_layer_window, n_gpu_layers, model, cparams);
} // synchronize the new n_layer_window and n_gpu_layers to other nodes
llama_bcast_layer_setup(lctx, n_layer_window, n_gpu_layers);
} else {
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers);
}
// update n_layer_window and n_gpu_layers // update n_layer_window and n_gpu_layers
std::copy(std::begin(n_layer_window), std::end(n_layer_window), params.n_layer_window); std::copy(std::begin(n_layer_window), std::end(n_layer_window), params.n_layer_window);
std::copy(std::begin(n_layer_window), std::end(n_layer_window), cparams.n_layer_window); std::copy(std::begin(n_layer_window), std::end(n_layer_window), cparams.n_layer_window);
std::copy(std::begin(n_layer_window), std::end(n_layer_window), mparams.n_layer_window); std::copy(std::begin(n_layer_window), std::end(n_layer_window), mparams.n_layer_window);
std::copy(std::begin(n_layer_window), std::end(n_layer_window), llama_context_n_layer_window(lctx)); std::copy(std::begin(n_layer_window), std::end(n_layer_window), llama_context_n_layer_window(lctx));
params.n_gpu_layers = n_gpu_layers[my_rank]; params.n_gpu_layers = n_gpu_layers[my_rank];
cparams.n_gpu_layers = n_gpu_layers[my_rank]; cparams.n_gpu_layers = n_gpu_layers[my_rank];
mparams.n_gpu_layers = n_gpu_layers[my_rank]; mparams.n_gpu_layers = n_gpu_layers[my_rank];
llama_model_set_n_gpu_layers(model, n_gpu_layers[my_rank]); llama_model_set_n_gpu_layers(model, n_gpu_layers[my_rank]);
#ifdef LLAMA_DEBUG #ifdef LLAMA_DEBUG
device_print_props(dev_info_set, n_world, model, cparams); device_print_props(dev_info_set, n_world, model, cparams);
#endif #endif
}
LOG_INF("\nUsing window size: %d, GPU layers: %d\n\n", cparams.n_layer_window[my_rank], cparams.n_gpu_layers);
if (!mparams.vocab_only && llm_load_tensors(ml, model, mparams) < 0) { if (!mparams.vocab_only && llm_load_tensors(ml, model, mparams) < 0) {
LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str()); LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str());