Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

llama : refactor rope_freq_base/scale_swa conversion and init (#18553)

Browse Source 
* refactor rope_freq_base/scale_swa conversion and init

* safe defaults for unknowns

* update relevant models

* grammar

* add get_rope_freq_scale to modern-bert

* const

* const

* log swa info
This commit is contained in:
Sigbjørn Skjæret
2026-01-05 09:14:04 +01:00
committed by GitHub
parent 67e3f6f601
commit eadc4184ca
10 changed files with 94 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/llama-iswa.cpp
+6 -2
View File
@@ -25,8 +25,12 @@ llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
const float freq_base_l = model.get_rope_freq_base (cparams, il);
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
ggml_tensor * inpSA = inpL;
// This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
(il + 1) % hparams.n_no_rope_layer_step != 0;
@@ -67,13 +71,13 @@ llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_
if (use_rope) {
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow
);
} else if (inp_attn_scale) {