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

add comment explaining why finetune checkpoints are allocated in one block

Browse source
This commit is contained in:
xaedes 2023-08-29 00:57:28 +02:00
parent 9a28bce29a
commit 49af7fbe12
No known key found for this signature in database
GPG key ID: 30030EDD817EA2B1
1 changed files with 2 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

2
examples/finetune/finetune.cpp
View file

@ -932,6 +932,8 @@ struct ggml_tensor * llama_build_lora_finetune_graphs(
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one)); ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
} }
// allocating checkpoints in one block to reduce memory fragmentation
// note: they will be freed in reverse order
for (int i = 0; i < checkpoints.size(); ++i) { for (int i = 0; i < checkpoints.size(); ++i) {
if (checkpoints[i]->data == NULL && !ggml_is_view(checkpoints[i])) { if (checkpoints[i]->data == NULL && !ggml_is_view(checkpoints[i])) {
ggml_allocr_alloc(alloc, checkpoints[i]); ggml_allocr_alloc(alloc, checkpoints[i]);