train : make KQ_pos memory buffer permanent via dummy scale op
This commit is contained in:
parent
54206962c7
commit
2f3a46fccf
1 changed files with 2 additions and 0 deletions
|
@ -795,6 +795,8 @@ struct ggml_tensor * llama_build_train_graphs(
|
|||
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
|
||||
// input gradient
|
||||
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
|
||||
// KQ_pos
|
||||
ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
|
||||
GGML_ASSERT(t36->grad->data == NULL && !ggml_is_view(t36->grad));
|
||||
ggml_allocr_alloc(alloc, t36->grad);
|
||||
// gradient tensors (will be set to zero by ggml_graph_reset)
|
||||
|
|
Loading…
Add table
Add a link
Reference in a new issue