LLM-Inference/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-08-12 22:29:39 +04:00
Code Issues Packages Projects Releases Wiki Activity

llama: pass rope freq factors to build_std_attention (#2291)

Browse Source 
build_llama passes nullptr for the rope_factors_in argument, so
rope_freqs.weight never reaches ggml_rope_ext and llama3 rope frequency
scaling is not applied. The LLAMA_SPLIT_MODE_GRAPH branch inside
build_std_attention falls back to model.layers[il].rope_freqs, so only
the standard path is affected.
This commit is contained in:
mb8565
2026-08-11 08:17:25 +02:00
committed by GitHub
parent c8772a8429
commit b382ebd848
1 changed files with 2 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphs/build_llama.cpp
+2 -4
View File
@@ -46,18 +46,16 @@ ggml_cgraph * llm_build_context::build_llama() {
int this_n_swa = this_KQ_mask == KQ_mask_swa ? hparams.n_swa : 0;
// rope freq factors for llama3; may return nullptr for llama2 and other models
//auto rope_factors = build_rope_factors(il);
auto rope_factors = build_rope_factors(il);
// self-attention
if (use_rope) {
cur = build_std_attention(gf, model.layers[il].attn_norm, inpL,
inp_pos, il == n_layer - 1 && n_tokens > 1 ? inp_out_ids : nullptr, nullptr,
inp_pos, il == n_layer - 1 && n_tokens > 1 ? inp_out_ids : nullptr, rope_factors,
this_KQ_mask, nullptr, nullptr, kq_scale, hparams.f_attention_scale, this_n_swa, il, true, false, true);
}
else {
auto rope_factors = build_rope_factors(il);
// norm
cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);