mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2026-09-20 01:31:42 +02:00
merge new rope param nonsense
This commit is contained in:
+10
-2
@@ -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
|
||||
{
|
||||
|
||||
Reference in New Issue
Block a user