Metadata-Version: 2.5
Name: loragression
Version: 0.0.2
Summary: LoRA fine-tuning, but more aggressively
Author: Jenny Luo
Classifier: Development Status :: 1 - Planning
Classifier: Programming Language :: Python :: 3
Requires-Python: >=3.10
Description-Content-Type: text/markdown

# LoRAgression

LoRA fine-tuning, but more aggressively.

Experimental utilities for parameter-efficient fine-tuning.

Very early development.

## Current functionality

Utilities for inspecting trainable model parameters and estimating LoRA adapter sizes.

```python
from loragression import lora_parameter_count

lora_parameter_count(in_features=768, out_features=768, rank=16)  # 24576
```

`lora_parameter_count` counts the two adapter matrices for one dense linear layer.
It excludes the base layer, biases, and task head; no model needs to be loaded.
`trainable_parameter_counts(model)` reports trainable and total counts and their fraction.

More aggression forthcoming.
