vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-06 15:39:28 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

rename arg --keep-inp-out-in-metal to --keep-out-in-metal

Browse source
This commit is contained in:
Lizonghang 2025-01-23 23:17:06 +04:00
parent 5fcf020cfb
commit 1c0087e919
5 changed files with 14 additions and 14 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
common/arg.cpp
View file

@ -739,10 +739,10 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
).set_env("LLAMA_ARG_CUDA_MEM"));
#ifdef GGML_USE_METAL
add_opt(llama_arg(
{"--keep-inp-out-in-metal"},
format("whether to keep input and output weight in metal (default: %s)", params.keep_inp_out_in_metal ? "true" : "false"),
{"--keep-out-in-metal"},
format("whether to keep output weights in metal memory (default: %s)", params.keep_out_in_metal ? "true" : "false"),
[](gpt_params & params) {
params.keep_inp_out_in_metal = true;
params.keep_out_in_metal = true;
}
).set_env("LLAMA_ARG_KEEP_INP_OUT_IN_METAL"));
#endif

6
common/common.cpp
View file

@ -1162,7 +1162,7 @@ static void assign_device(
if (dev.gpu_support.cuda || dev.gpu_support.metal) {
float reserved_mem = 0.1f; // reserved shared memory to avoid potential OOM, set to 100 MiB by default
vec_z_gpu[m] = (double)((dev.gpu_props.memory_free - reserved_mem) * GIGABYTE - c_gpu[m]) / (double)(n_layer * b_prime);
if (dev.gpu_support.metal && m == 0 && cparams.keep_inp_out_in_metal) {
if (dev.gpu_support.metal && m == 0 && cparams.keep_out_in_metal) {
vec_z_gpu[m] -= (double)(bi + bo) / (double)(n_layer * b_prime);
}
dev_gpu[m] = 1;
@ -1624,7 +1624,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
mparams.use_mmap = params.use_mmap;
mparams.use_mlock = params.use_mlock;
mparams.check_tensors = params.check_tensors;
mparams.keep_inp_out_in_metal = params.keep_inp_out_in_metal;
mparams.keep_out_in_metal = params.keep_out_in_metal;
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), mparams.n_layer_window);
if (params.kv_overrides.empty()) {
mparams.kv_overrides = NULL;
@ -1671,7 +1671,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.n_world = params.n_world;
cparams.rank = params.rank;
cparams.unload = params.unload;
cparams.keep_inp_out_in_metal = params.keep_inp_out_in_metal;
cparams.keep_out_in_metal = params.keep_out_in_metal;
cparams.n_gpu_layers = params.n_gpu_layers;
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), cparams.n_layer_window);

2
common/common.h
View file

@ -148,7 +148,7 @@ struct gpt_params {
std::string master_ip = "localhost"; // ip address of the master node
std::string next_node_ip = "localhost"; // ip address of my next node
bool unload = false; // unload layer weights after use or not
bool keep_inp_out_in_metal = false; // whether to keep input/output weight in metal, not by default
bool keep_out_in_metal = false; // whether to keep output weights in metal memory, not by default
int32_t gpu_mem = 999.0; // gpu memory to use, in GiB
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size

4
include/llama.h
View file

@ -312,7 +312,7 @@ extern "C" {
bool use_mmap; // use mmap if possible
bool use_mlock; // force system to keep model in RAM
bool check_tensors; // validate model tensor data
bool keep_inp_out_in_metal; // whether to keep input/output weight in metal
bool keep_out_in_metal; // whether to keep output weights in metal memory
};
// NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
@ -323,7 +323,7 @@ extern "C" {
uint32_t n_layer_window[32];// number of layers to process in each compute
uint32_t n_gpu_layers; // number of layers to process on GPU
bool unload; // whether to unload layer weights after use
bool keep_inp_out_in_metal; // whether to keep input/output weight in metal
bool keep_out_in_metal; // whether to keep output weights in metal memory
char * master_ip; // ip address of the master node
char * next_node_ip; // ip address of the next node
uint32_t n_ctx; // text context, 0 = from model

10
src/llama.cpp
View file

@ -7397,7 +7397,7 @@ static bool llm_load_tensors_impl(
enum llama_split_mode split_mode,
int main_gpu,
bool use_mlock,
bool keep_inp_out_in_metal,
bool keep_out_in_metal,
llama_progress_callback progress_callback,
void * progress_callback_user_data) {
auto & hparams = model.hparams;
@ -9283,7 +9283,7 @@ static bool llm_load_tensors_impl(
void * addr = nullptr;
auto & ranges = ctx_buffer_ranges[idx];
ml.get_mapping_ranges(ranges, &addr, idx, ctx, keep_inp_out_in_metal ? cpu_ctx : nullptr);
ml.get_mapping_ranges(ranges, &addr, idx, ctx, keep_out_in_metal ? cpu_ctx : nullptr);
for (const auto & range : ranges) {
size_t first = range.first;
@ -9407,7 +9407,7 @@ int llm_load_tensors(
try {
if (!llm_load_tensors_impl(
*ml, *model, params.n_world, params.rank, params.n_layer_window, params.n_gpu_layers, params.split_mode,
params.main_gpu, params.use_mlock, params.keep_inp_out_in_metal, params.progress_callback, params.progress_callback_user_data
params.main_gpu, params.use_mlock, params.keep_out_in_metal, params.progress_callback, params.progress_callback_user_data
)) {
return -2;
}
@ -19784,7 +19784,7 @@ struct llama_model_params llama_model_default_params() {
/*.use_mmap =*/ true,
/*.use_mlock =*/ false,
/*.check_tensors =*/ false,
/*.keep_inp_out_in_metal =*/ false,
/*.keep_out_in_metal =*/ false,
};
#ifdef GGML_USE_METAL
@ -19802,7 +19802,7 @@ struct llama_context_params llama_context_default_params() {
/*.n_layer_window =*/ {32},
/*.n_gpu_layers =*/ 0,
/*.unload =*/ false,
/*.keep_inp_out_in_metal =*/ false,
/*.keep_out_in_metal =*/ false,
/*.master_ip =*/ nullptr,
/*.next_node_ip =*/ nullptr,
/*.n_ctx =*/ 512,