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

llama : fix llm_build_k_shift to use n_head_kv instead of n_head

Browse source
This commit is contained in:
Georgi Gerganov 2023-10-29 21:17:46 +02:00
parent 5990861938
commit 31a12f3d03
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
llama.cpp
View file

@ -3253,8 +3253,8 @@ static void llm_build_k_shift(
const auto & hparams = model.hparams; const auto & hparams = model.hparams;
const int64_t n_head = hparams.n_head;
const int64_t n_layer = hparams.n_layer; const int64_t n_layer = hparams.n_layer;
const int64_t n_head_kv = hparams.n_head_kv;
const int64_t n_embd_gqa = hparams.n_embd_gqa(); const int64_t n_embd_gqa = hparams.n_embd_gqa();
const int64_t n_embd_head = hparams.n_embd_head(); const int64_t n_embd_head = hparams.n_embd_head();
@ -3281,7 +3281,7 @@ static void llm_build_k_shift(
// we rotate only the first n_rot dimensions // we rotate only the first n_rot dimensions
ggml_rope_custom_inplace(ctx, ggml_rope_custom_inplace(ctx,
ggml_view_3d(ctx, kv_self.k, ggml_view_3d(ctx, kv_self.k,
n_rot, n_head, n_ctx, n_rot, n_head_kv, n_ctx,
ggml_element_size(kv_self.k)*n_embd_head, ggml_element_size(kv_self.k)*n_embd_head,
ggml_element_size(kv_self.k)*n_embd_gqa, ggml_element_size(kv_self.k)*n_embd_gqa,
ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il),