tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 04:15:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: Move build_inp_pos to the top of the graph section for build_granite (#13538)

Browse Source 
This matches how others do it, but will still avoid the extra
initialization when rope is disabled.

Branch: GraniteFour

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart
2025-05-14 06:53:59 -06:00
committed by GitHub
parent 053174436f
commit 5e7d95e22e
1 changed files with 3 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/llama-model.cpp
View File

@ -12218,6 +12218,9 @@ struct llm_build_granite : public llm_graph_context {
// inp_pos - built only if rope enabled
ggml_tensor * inp_pos = nullptr;
if (use_rope) {
inp_pos = build_inp_pos();
}
auto * inp_attn = build_attn_inp_kv_unified();
@ -12260,10 +12263,6 @@ struct llm_build_granite : public llm_graph_context {
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
if (use_rope) {
if (!inp_pos) {
inp_pos = build_inp_pos();
}
ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, rope_factors,