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

Fix issue with alloc causing max_compute_size to be calculated

Browse source
This commit is contained in:
Uzo Nweke 2024-01-18 21:40:19 -05:00
parent 57e2a7a52a
commit f783c5971f
1 changed files with 5 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

12
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -1077,7 +1077,6 @@ 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 = {
@ -1099,10 +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);
alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment); ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
ggml_allocr_alloc(alloc, tokens_input); ggml_allocr_alloc(alloc_inps, tokens_input);
ggml_allocr_alloc(alloc, target_probs); ggml_allocr_alloc(alloc_inps, target_probs);
ggml_allocr_free(alloc);
// 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 = (
@ -1129,7 +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);
alloc = ggml_allocr_new_measure(tensor_alignment); ggml_allocr_t 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);
@ -1162,7 +1160,7 @@ 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);
alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment); ggml_allocr_t 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);