vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-10 17:14:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Modified RoPE with linear scaling

Browse source 
When the context size is greater than the maximum context size
during training, scale the position given to RoPE with
trainign context / n_ctx.
This commit is contained in:
Iwan Kawrakow 2023-06-27 15:00:22 +03:00
parent 0be54f75a6
commit cda30038e4
6 changed files with 34 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

4
llama.cpp
View file

@ -1491,11 +1491,11 @@ static bool llama_eval_internal(
offload_func_kq(tmpq);
ggml_set_name(tmpq, "tmpq");
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0, n_ctx);
offload_func_kq(Kcur);
ggml_set_name(Kcur, "Kcur");
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0, n_ctx);
offload_func_kq(Qcur);
ggml_set_name(Qcur, "Qcur");