wassname/lora-lite
1
0
Fork 0
mirror of https://github.com/wassname/lora-lite.git synced 2026-06-27 14:00:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

lora_xs: fix docstring -- A=diag(Sr)Vhr has row norms Sr, not orthonormal

Browse Source 
External review (GPT-5.5) flagged 'two near-orthonormal bases' as inaccurate:
only B=Ur is orthonormal; A folds the singular values so its rows are scaled.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-06-18 20:01:59 +08:00
parent c792ad3e5f
commit 4e03f9c07f
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/lora_lite/variants/lora_xs.py
+4 -2
View File
@@ -13,8 +13,10 @@ the full W, and R (init normal(0, 1e-5)) starts the adapter at ~identity. So the
trainable tensor is r*r (e.g. r=32 -> 1024 params/layer), hence "extremely small".
The reference folds all singular values into A and leaves B as the raw left singular
vectors; R sits between two frozen, near-orthonormal bases. Their LLaMA math-tuning
config sets lora_alpha = r (scale = 1.0) and lr ~ 4e-3 (scripts/run_math_tuning.sh).
vectors. So R sits between B = Ur (orthonormal) and A = diag(Sr) Vhr (orthonormal rows
*scaled* by the singular values, so row norms = Sr, not unit) -- the asymmetry is the
reference's, not a bug. Their LLaMA math-tuning config sets lora_alpha = r (scale = 1.0)
and lr ~ 4e-3 (scripts/run_math_tuning.sh).
Refs:
- paper repo: https://github.com/MohammadrezaBanaei/LoRA-XS