vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

use lora->get_scale

Browse source
This commit is contained in:
Xuan Son Nguyen 2025-01-07 22:32:26 +01:00
parent e444b8e0c2
commit b37af1424a
1 changed files with 2 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

5
src/llama.cpp
View file

@ -2632,9 +2632,8 @@ static struct ggml_tensor * llm_build_lora_mm(
if (lora == nullptr) { if (lora == nullptr) {
continue; continue;
} }
const float alpha = it.first->alpha; const float adapter_scale = it.second;
const float rank = (float) lora->b->ne[0]; const float scale = lora->get_scale(it.first->alpha, adapter_scale);
const float scale = alpha ? it.second * alpha / rank : it.second;
struct ggml_tensor * ab_cur = ggml_mul_mat( struct ggml_tensor * ab_cur = ggml_mul_mat(
ctx0, lora->b, ctx0, lora->b,
ggml_mul_mat(ctx0, lora->a, cur) ggml_mul_mat(ctx0, lora->a, cur)