ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Make sure we can fuse Q and K RoPE for DeepSeek models (#985)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-11-19 13:43:08 +01:00
committed by GitHub
parent 054c31cf8f
commit 047a519771
1 changed files with 4 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama-build-context.cpp
View File

@@ -6054,6 +6054,8 @@ ggml_cgraph * llm_build_context::build_deepseek2() {
cb(k_rope, "k_rope", il);
cb(kv_compressed, "kv_compressed", il);
ggml_build_forward_expand(gf, q_rope);
ggml_build_forward_expand(gf, k_rope);
if (rope_cache) {
q_rope = ggml_rope_fast(ctx0, q_rope, rope_cache);
k_rope = ggml_rope_fast(ctx0, k_rope, rope_cache);
@@ -6066,6 +6068,8 @@ ggml_cgraph * llm_build_context::build_deepseek2() {
}
cb(q_rope, "q_rope", il);
cb(k_rope, "k_rope", il);
ggml_build_forward_expand(gf, q_rope);
ggml_build_forward_expand(gf, k_rope);
kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams, model.layers[il].attn_kv_a_norm, NULL, LLM_NORM_RMS, cb, il);
cb(kv_compressed, "kv_compressed", il);