ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-20 22:49:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

Discard unnecessary changes in llama-build-context.cpp

Browse Source
This commit is contained in:
Kawrakow
2026-02-13 06:57:12 +00:00
parent 10f125754e
commit edadb80ed6
1 changed files with 4 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/llama-build-context.cpp
View File

@@ -557,12 +557,12 @@ void llm_build_context::llm_build_kv_store(
struct ggml_tensor * v_cache_view = nullptr;
if (!kv.v_trans) {
if (cparams.flash_attn) {
v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa,
(kv_head)*ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa));
lctx.cache_copies[2*il+1].step = ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa);
} else {
// note: the V cache is transposed for legacy non-FA layouts
// note: the V cache is transposed when not using flash attention
v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa,
( n_ctx)*ggml_element_size(kv.v_l[il]),
(kv_head)*ggml_element_size(kv.v_l[il]));
@@ -1502,21 +1502,12 @@ static ggml_tensor * llm_build_kqv(
} else {
// split cached v into n_head heads
struct ggml_tensor * v;
if (kv.v_trans) {
v = ggml_view_3d(ctx, kv.v_l[il],
struct ggml_tensor * v =
ggml_view_3d(ctx, kv.v_l[il],
n_kv, n_embd_head_v, n_head_kv,
ggml_element_size(kv.v_l[il])*n_ctx,
ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v,
0);
} else {
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_v_gqa),
ggml_row_size(kv.v_l[il]->type, n_embd_head_v),
0);
v = ggml_cont(ctx, ggml_transpose(ctx, v));
}
cb(v, "v", il);
auto kq_size = k->ne[1]*q->ne[1]*q->ne[2]*sizeof(float)/(1024*1024);