mirror of
https://github.com/wassname/lora-lite.git
synced 2026-06-27 18:05:16 +08:00
wassname
90b5199ed9
Replace the stale single AntiPaSTO row (was 35.8K params from the removed rotation version, described block-Cayley which no longer exists) with the real 5000-step Qwen3-0.6B numbers and a family breakdown: corda 61.9% 14.3K (best: covariance-oriented basis) plain 61.4% 14.3K rot 61.4% 35.8K (the rotation this replaces) ablate 61.0% 14.4K arrow 60.5% 17.5K Headline: ~320x fewer trainable params than LoRA at ~97% of its accuracy. Rotation buys nothing (rot matches plain to 3 s.f. at 2.5x params, +20% wall-time, plus a per-forward Cayley solve), confirming the drop. Co-Authored-By: Claudypoo <noreply@anthropic.com>
98 lines
4.8 KiB
Markdown
98 lines
4.8 KiB
Markdown
# lora-lite
|
||
|
||
Hackable PyTorch adapters for LoRA-family and small PEFT experiments.
|
||
|
||
## Hackable code
|
||
|
||
|
||
To keep it simple and hackable we make these choices:
|
||
|
||
- Simple forward hooks, no module replacement or custom modules.
|
||
- Simple code over fast performance
|
||
- No merge/unmerge
|
||
- Single test where we train on MetaMathQA and test on GSM8K for each variant
|
||
|
||
Take a look at [lora.py](src/lora_lite/variants/lora.py)
|
||
|
||
## Install
|
||
|
||
```bash
|
||
pip install -e git+https://github.com/wassname/lora-lite.git#egg=lora-lite
|
||
```
|
||
|
||
## Quickstart
|
||
|
||
```python
|
||
import torch, lora_lite as ll
|
||
|
||
model = MyTransformer()
|
||
cfg = ll.LoRAConfig(r=8, alpha=16, dtype=torch.bfloat16)
|
||
ll.attach(model, cfg)
|
||
|
||
opt = torch.optim.AdamW([p for p in model.parameters() if p.requires_grad], lr=1e-4)
|
||
# train...
|
||
|
||
ll.save(model, "adapter.safetensors")
|
||
ll.detach(model)
|
||
ll.load(model, "adapter.safetensors")
|
||
```
|
||
|
||
## Does it work?
|
||
|
||
```bash
|
||
just check # pytest + smoke + package build + metadata check
|
||
just bnb-smoke # required CUDA bitsandbytes 4bit/8bit smoke
|
||
just qwen-probe # Qwen/Qwen3-0.6B train/save-load probe
|
||
```
|
||
|
||
## Variants
|
||
|
||
| Variant | 4bit/8bit | GSM8K % | Params | Peak GPU (GB) |
|
||
| --------------------------------------------- | --------- | ------- | ---------- | ------------- |
|
||
| [LoRA](https://arxiv.org/abs/2106.09685) | yes | 63.2% | 4.59M | 11.3 |
|
||
| [PiSSA](https://arxiv.org/abs/2404.02948) | no | 63.2% | 4.59M | 11.3 |
|
||
| [DoRA](https://arxiv.org/abs/2402.09353) | no | 62.4% | 4.67M | 11.3 |
|
||
| [DeLoRA](https://arxiv.org/abs/2503.18225) | yes | 61.5% | 4.59M | 11.3 |
|
||
| [AntiPaSTO](https://arxiv.org/abs/2601.07473) | no | 61.4% | 14.3K | 11.3 |
|
||
| [IA3-FF](https://arxiv.org/pdf/2205.05638) | yes | 61.4% | 86K | 11.4 |
|
||
| [EVA](https://arxiv.org/abs/2410.07170) | no | 60.3% | 4.59M | 11.3 |
|
||
| [IA3](https://arxiv.org/pdf/2205.05638) | yes | 60.0% | 57K | 11.4 |
|
||
| [HRA](https://arxiv.org/abs/2405.17484) | yes | 61.6% | 1.84M | 11.3 |
|
||
|
||
Params = trainable adapter params. Peak GPU = peak CUDA memory during train+eval (logged from this run onward; older runs predate the column).
|
||
|
||
Setup: Qwen3-0.6B-Base, MetaMathQA train (5k steps, batch 4 = 20k samples unless noted), r=32, all q/v targets, GSM8K test (1319 examples). HRA used batch 2 (10k samples) due to memory. The AntiPaSTO family used r=256 (default for these variants).
|
||
|
||
Reference: PEFT reports LoRA at 49.0% on Llama-3.2-3B (different model, different sample count). Our numbers are not directly comparable but suggest the adapters work.
|
||
|
||
### AntiPaSTO family
|
||
|
||
AntiPaSTO learns a per-direction gain on the frozen top-r SVD basis (`S_eff = S * (1 + ELU(coeff*g))`), so it rescales existing singular directions rather than creating new ones, hence ~320x fewer trainable params than LoRA at ~97% of its accuracy. All variants share the diagonal gain; they differ only in the basis they steer in or the extra structure on the top directions. All have `base_grad_leaks=0` (the frozen residual weight gets no gradient).
|
||
|
||
| Variant | GSM8K % | Params | Basis / extra structure |
|
||
| ------------------ | ------- | ------ | ------------------------------------------------------------------- |
|
||
| antipasto_corda | 61.9% | 14.3K | covariance-oriented input projector `P = Vh·C^{-1/2}` (best of family) |
|
||
| antipasto | 61.4% | 14.3K | plain weight-SVD basis, diagonal gain only |
|
||
| antipasto_rot | 61.4% | 35.8K | + block-Cayley rotation of the basis (the version this replaces) |
|
||
| antipasto_ablate | 61.0% | 14.4K | contractive output ablation `(I - α ĉĉᵀ)diag(S)`, can't amplify |
|
||
| antipasto_arrow | 60.5% | 17.5K | dense b×b mixing block on the top-b directions + diagonal tail |
|
||
|
||
The rotation buys nothing here: `antipasto_rot` matches plain `antipasto` to 3 s.f. (61.4%) at 2.5x the params and +20% wall-time, while paying a per-forward Cayley solve. Dropping it (the current default) is free. CorDA's data-oriented basis is the only structure that helps on this capability task; the ablation/arrow cores are aimed at steering and suppression, where the diagonal-only basis can't reach off-axis behavior, so they don't pay off for raw GSM8K accuracy.
|
||
|
||
|
||
## Developer docs
|
||
|
||
See [docs/developer_guide.md](docs/developer_guide.md) for the variant API, data-calibrated init, and save/load format.
|
||
|
||
## Citation
|
||
|
||
```bibtex
|
||
@misc{wassname2026loralite,
|
||
title = {LoRA-Lite: A Hackable Adapter Library for Research},
|
||
author = {Michael J. Clark},
|
||
year = {2026},
|
||
url = {https://github.com/wassname/lora-lite/}
|
||
}
|
||
```
|
||
|