ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 15:44:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

FlashMLA-2: slightly smaller computer buffer size

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-03-12 15:06:31 +02:00
parent 3db0f1cc7f
commit 765c03d09b
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama.cpp
View File

@@ -13661,6 +13661,9 @@ struct llm_build_context {
auto k_nope = ggml_cast(ctx0, k_nope_f32, kv_self.kv_l[il]->type);
cb(k_nope, "k_nope", il);
ggml_build_forward_expand(gf, k_nope);
ggml_build_forward_expand(gf, v);
auto kv_cache_rope = ggml_view_3d(ctx0, kv_self.kv_l[il], n_embd_head_qk_rope, n_kv, 1,
kv_self.kv_l[il]->nb[1], kv_self.kv_l[il]->nb[2], ggml_row_size(kv_self.kv_l[il]->type, kv_lora_rank));
@@ -13673,7 +13676,6 @@ struct llm_build_context {
cb(k, "k", il);
ggml_build_forward_expand(gf, k);
ggml_build_forward_expand(gf, v);
}
else {
// Hahaha, we need to convert the KV cache for this layer to f32 because the general purpose ML library ggml does not