Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-23 01:49:18 +01:00
Code Issues Packages Projects Releases Wiki Activity

llama : use n_embd_head_v when reshaping kqv (#7327)

Browse Source 
* llama : use n_embd_head_v instead of n_embd_head_k when reshaping kqv

* llama : use n_embd_v_gqa and n_embd_head_v instead of n_embd_k_gqa and n_embd_head_k when making a view of cached value vectors.

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
This commit is contained in:
fairydreaming 2024-05-17 13:24:38 +02:00 committed by GitHub
parent 29c60d8cdd
commit 27b040691c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
llama.cpp
View File

@ -6622,6 +6622,7 @@ static struct ggml_tensor * llm_build_kqv(
const int64_t n_embd_head_k = hparams.n_embd_head_k;
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
const int64_t n_embd_head_v = hparams.n_embd_head_v;
const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3);
cb(q, "q", il);
@ -6644,8 +6645,8 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * v =
ggml_view_3d(ctx, kv.v_l[il],
n_embd_head_v, n_kv, n_head_kv,
ggml_row_size(kv.v_l[il]->type, n_embd_k_gqa),
ggml_row_size(kv.v_l[il]->type, n_embd_head_k),
ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa),
ggml_row_size(kv.v_l[il]->type, n_embd_head_v),
0);
cb(v, "v", il);
@ -6655,7 +6656,7 @@ static struct ggml_tensor * llm_build_kqv(
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
}
cur = ggml_reshape_2d(ctx, cur, n_embd_head_k*n_head, n_tokens);
cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);
} else {
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
cb(kq, "kq", il);
@ -6700,7 +6701,7 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3);
cb(kqv_merged, "kqv_merged", il);
cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens);
cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_v*n_head, n_tokens);
cb(cur, "kqv_merged_cont", il);
}