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

remove ggml_allocr_free as suggested in issue #4791

Browse source
This commit is contained in:
Uzo Nweke 2024-01-19 13:12:27 -05:00
parent f783c5971f
commit 38f2e6e7c3
1 changed files with 6 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

14
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -263,7 +263,6 @@ static void init_model(struct my_llama_model * model) {
model->data.resize(size + tensor_alignment); model->data.resize(size + tensor_alignment);
alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment); alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment);
alloc_model(alloc, model); alloc_model(alloc, model);
ggml_allocr_free(alloc);
} }
static void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) { static void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) {
@ -1077,6 +1076,7 @@ int main(int argc, char ** argv) {
std::vector<uint8_t> mem_input_data; std::vector<uint8_t> mem_input_data;
std::vector<uint8_t> mem_compute_data; std::vector<uint8_t> mem_compute_data;
ggml_allocr * alloc = NULL;
// context for input tensors without their data // context for input tensors without their data
struct ggml_init_params ctx_input_params = { struct ggml_init_params ctx_input_params = {
@ -1098,9 +1098,9 @@ int main(int argc, char ** argv) {
// allocate input tensors // allocate input tensors
mem_input_data.resize(max_input_size); mem_input_data.resize(max_input_size);
ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment); alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
ggml_allocr_alloc(alloc_inps, tokens_input); ggml_allocr_alloc(alloc, tokens_input);
ggml_allocr_alloc(alloc_inps, target_probs); ggml_allocr_alloc(alloc, target_probs);
// context for compute tensors without their data // context for compute tensors without their data
const size_t estimated_compute_size_wo_data = ( const size_t estimated_compute_size_wo_data = (
@ -1127,7 +1127,7 @@ int main(int argc, char ** argv) {
// find best evaluation order // find best evaluation order
for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) { for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
ctx_compute = ggml_init(ctx_compute_params); ctx_compute = ggml_init(ctx_compute_params);
ggml_allocr_t alloc = ggml_allocr_new_measure(tensor_alignment); alloc = ggml_allocr_new_measure(tensor_alignment);
gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
gf->order = (enum ggml_cgraph_eval_order) order; gf->order = (enum ggml_cgraph_eval_order) order;
gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@ -1147,7 +1147,6 @@ int main(int argc, char ** argv) {
best_compute_size = max_compute_size; best_compute_size = max_compute_size;
best_order = gf->order; best_order = gf->order;
} }
ggml_allocr_free(alloc);
ggml_free(ctx_compute); ggml_free(ctx_compute);
} }
size_t max_compute_size = best_compute_size; size_t max_compute_size = best_compute_size;
@ -1160,7 +1159,7 @@ int main(int argc, char ** argv) {
// allocate compute tensors // allocate compute tensors
mem_compute_data.resize(max_compute_size); mem_compute_data.resize(max_compute_size);
ctx_compute = ggml_init(ctx_compute_params); ctx_compute = ggml_init(ctx_compute_params);
ggml_allocr_t alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment); alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
gf->order = best_order; gf->order = best_order;
gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@ -1175,7 +1174,6 @@ int main(int argc, char ** argv) {
params.common.use_flash, params.common.use_flash,
params.common.use_checkpointing params.common.use_checkpointing
); );
ggml_allocr_free(alloc);
std::vector<llama_token> train_tokens; std::vector<llama_token> train_tokens;
std::vector<size_t> train_samples_begin; std::vector<size_t> train_samples_begin;