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

* 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
+9 -5
View File
@@ -22,8 +22,15 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
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;
// dual attention normalization (pre)
cur = build_norm(inpL,
model.layers[il].attn_norm, NULL,
@@ -56,19 +63,16 @@ llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_para
cb(Qcur, "Qcur_normed", il);
cb(Kcur, "Kcur_normed", il);
// RoPE only for sliding_attention layers
const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
((il + 1) % hparams.n_no_rope_layer_step) != 0;
if (use_rope) {
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
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);
cb(Qcur, "Qcur_rope", il);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
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);
cb(Kcur, "Kcur_rope", il);
}