vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

remove unused train params: mem_compute1_gb & mem_compute2_gb

Browse source 
mem_compute_gb is used for compute when automatic memory allocator is not enabled, otherwise it can be very small to only hold the tensor definitions
mem_compute0_gb is used for automatic memory allocator (as long as measurement of max required size is not implemented)
This commit is contained in:
xaedes 2023-08-14 18:44:42 +02:00
parent 6f161c784b
commit 3794dceb7f
No known key found for this signature in database
GPG key ID: 30030EDD817EA2B1
1 changed files with 1 additions and 25 deletions
Download patch file Download diff file Expand all files Collapse all files

26
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -2217,9 +2217,6 @@ struct train_params {
int mem_model_gb; int mem_model_gb;
int mem_compute_gb; int mem_compute_gb;
int mem_compute0_gb; int mem_compute0_gb;
int mem_compute1_gb;
int mem_compute2_gb;
int mem_compute3_gb;
}; };
struct train_params get_default_train_params() { struct train_params get_default_train_params() {
@ -2278,8 +2275,6 @@ struct train_params get_default_train_params() {
params.mem_model_gb = 2; params.mem_model_gb = 2;
params.mem_compute_gb = 24; params.mem_compute_gb = 24;
params.mem_compute0_gb = 8; params.mem_compute0_gb = 8;
params.mem_compute1_gb = 1;
params.mem_compute2_gb = 2;
return params; return params;
} }
@ -2336,9 +2331,7 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p
fprintf(stderr, " --lbfgs-iter N Maximum number of LBFGS optimization iterations for each batch (default %d)\n", params->lbfgs_n_iter); fprintf(stderr, " --lbfgs-iter N Maximum number of LBFGS optimization iterations for each batch (default %d)\n", params->lbfgs_n_iter);
fprintf(stderr, " --mem-model N Memory to allocate for model and cache in gigabytes. (default %d)\n", params->mem_model_gb); fprintf(stderr, " --mem-model N Memory to allocate for model and cache in gigabytes. (default %d)\n", params->mem_model_gb);
fprintf(stderr, " --mem-compute N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute_gb); fprintf(stderr, " --mem-compute N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute_gb);
fprintf(stderr, " --mem-compute0 N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute0_gb); fprintf(stderr, " --mem-compute0 N Memory to allocate for automatic memory allocator in gigabytes. (default %d)\n", params->mem_compute0_gb);
fprintf(stderr, " --mem-compute1 N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute1_gb);
fprintf(stderr, " --mem-compute2 N Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute2_gb);
fprintf(stderr, "\n"); fprintf(stderr, "\n");
} }
@ -2604,18 +2597,6 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) {
break; break;
} }
params->mem_compute0_gb = std::stoi(argv[i]); params->mem_compute0_gb = std::stoi(argv[i]);
} else if (arg == "--mem-compute1") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->mem_compute1_gb = std::stoi(argv[i]);
} else if (arg == "--mem-compute2") {
if (++i >= argc) {
invalid_param = true;
break;
}
params->mem_compute2_gb = std::stoi(argv[i]);
} else if (arg == "-h" || arg == "--help") { } else if (arg == "-h" || arg == "--help") {
train_print_usage(argc, argv, &default_params); train_print_usage(argc, argv, &default_params);
exit(0); exit(0);
@ -2839,11 +2820,7 @@ int main(int argc, char ** argv) {
uint8_t * compute_addr = new uint8_t[compute_size]; uint8_t * compute_addr = new uint8_t[compute_size];
size_t size_buf_0 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute0_gb); size_t size_buf_0 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute0_gb);
size_t size_buf_1 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute1_gb);
size_t size_buf_2 = 1024ll*1024ll*1024ll*((size_t) params.mem_compute2_gb);
uint8_t * compute_buf_0 = new uint8_t[size_buf_0]; uint8_t * compute_buf_0 = new uint8_t[size_buf_0];
uint8_t * compute_buf_1 = new uint8_t[size_buf_1];
uint8_t * compute_buf_2 = new uint8_t[size_buf_2];
ggml_allocr * alloc = NULL; ggml_allocr * alloc = NULL;
if (params.use_alloc) { if (params.use_alloc) {
@ -3090,7 +3067,6 @@ int main(int argc, char ** argv) {
delete[] compute_addr; delete[] compute_addr;
delete[] compute_buf_0; delete[] compute_buf_0;
delete[] compute_buf_1;
ggml_free(model.ctx); ggml_free(model.ctx);
llama_free(lctx); llama_free(lctx);
llama_free_model(lmodel); llama_free_model(lmodel);