mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 30 Packages Projects Releases Wiki Activity

mla : make the V tensor a view of K (#18986)

Browse Source 
* mla : pass V as a view of K to the FA op

* cuda : adjust mla logic to new layout

* kv-cache : fix rope shift

* tests : remove comment

* cuda : fix reusable_cutoff

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Georgi Gerganov
2026-01-22 22:09:01 +02:00
committed by GitHub
parent e2baf02162
commit a5eaa1d6a3
8 changed files with 39 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-kv-cache.cpp
+6 -2
View File
@@ -1594,6 +1594,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
const auto & n_embd_head_k = hparams.n_embd_head_k;
//const auto & n_embd_head_v = hparams.n_embd_head_v;
const auto & n_rot = hparams.n_rot;
const auto n_embd_nope = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
auto inp = std::make_unique<llm_graph_input_k_shift>(this);
inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, (int64_t) get_size()*n_stream);
@@ -1614,10 +1618,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
ggml_tensor * k =
ggml_view_3d(ctx, layer.k,
n_embd_head_k, n_head_kv, get_size()*n_stream,
n_rot, n_head_kv, get_size()*n_stream,
ggml_row_size(layer.k->type, n_embd_head_k),
ggml_row_size(layer.k->type, n_embd_k_gqa),
0);
ggml_row_size(layer.k->type, n_embd_nope));
ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);