https://github.com/tky823/bitlinear158compression

https://github.com/tky823/bitlinear158compression

Last synced: 17 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/tky823/bitlinear158compression
• Owner: tky823
• Created: 2024-12-31T06:30:21.000Z (24 days ago)
• Default Branch: main
• Last Pushed: 2024-12-31T07:24:49.000Z (24 days ago)
• Last Synced: 2024-12-31T08:18:46.585Z (24 days ago)
• Language: Python
• Size: 5.86 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# BitLinear158Compression

In this library, we compare following compression models for inference by BitLinear158:

- `BitLinear158`: Ternary weight is managed by `float32` (naive implementation)

- `BitLinear158Int8Inference`: Ternary weight is managed by `int8`

- `BitLinear158Uint2Inference`: Ternary weight is managed by `uint2` defined in `torchao`

- `BitLinear158CppInference`: Ternary weight is managed by `int8` and forward pass is implemented by C++/CUDA. When weight is on CUDA, CUDA kernel is used under the hood.

## Installation

```sh

pip install .

```

## Test

```sh

python tests/test_bitlinear158_inference.py

```

## NOTE

There is room for improvement of computational speed by setting additional compling arguments in `setup.py`.