vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-12 18:09:42 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

merge new rope param nonsense

Browse source
This commit is contained in:
Concedo 2023-09-30 11:33:30 +08:00
commit b84e210f0d
38 changed files with 2811 additions and 764 deletions
Download patch file Download diff file Expand all files Collapse all files

12
otherarch/gptj_v3.cpp
View file

@ -486,8 +486,16 @@ bool gptj_eval(
// self-attention
{
struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, n_ctx, freq_base, freq_scale);
struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, n_ctx, freq_base, freq_scale);
struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
{
int * data = (int *) KQ_pos->data;
for (int i = 0; i < N; ++i) {
data[i] = n_past + i;
}
}
struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, freq_base, freq_scale);
struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), KQ_pos, n_rot, 0, n_ctx, freq_base, freq_scale);
// store key and value to memory
{