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

train : fix context size calculations

Browse source
This commit is contained in:
Georgi Gerganov 2023-11-03 14:40:45 +02:00
parent dc22db734b
commit b1592ea054
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
2 changed files with 6 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

8
examples/finetune/finetune.cpp
View file

@ -1742,11 +1742,9 @@ int main(int argc, char ** argv) {
ggml_allocr_free(alloc);
// context for compute tensors without their data
size_t estimated_compute_size_wo_data = (
ggml_tensor_overhead()*LLAMA_TRAIN_MAX_NODES*2
+ (GGML_OBJECT_SIZE+ggml_graph_overhead())*(
params.common.use_checkpointing ? 3 : 2
)
const size_t estimated_compute_size_wo_data = (
2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() +
(params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true))
);
struct ggml_init_params ctx_compute_params = {
estimated_compute_size_wo_data, // mem_size

8
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -1109,11 +1109,9 @@ int main(int argc, char ** argv) {
ggml_allocr_free(alloc);
// context for compute tensors without their data
size_t estimated_compute_size_wo_data = (
ggml_tensor_overhead()*LLAMA_TRAIN_MAX_NODES*2
+ (GGML_OBJECT_SIZE+ggml_graph_overhead())*(
params.common.use_checkpointing ? 3 : 2
)
const size_t estimated_compute_size_wo_data = (
2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() +
(params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true))
);
struct ggml_init_params ctx_compute_params = {
estimated_compute_size_wo_data, // mem_size