skolai/fewbit
Compression schema for gradients of activations in backward pass
This project helps machine learning engineers train very large neural networks more efficiently by reducing the memory required during the backward pass. It optimizes activation functions and linear layers, taking in your existing PyTorch model architecture and outputting a memory-optimized version. The primary users are deep learning practitioners working with models that push the limits of GPU memory.
No commits in the last 6 months.
Use this if you are training large neural networks and frequently encounter out-of-memory errors or want to reduce GPU memory footprint to use larger batch sizes or more complex models.
Not ideal if you are working with small models or datasets where memory efficiency is not a primary concern, or if you need to use highly custom activation functions not included in the library.
Stars
45
Forks
6
Language
Python
License
BSD-3-Clause
Category
Last pushed
Jul 26, 2023
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/ml-frameworks/skolai/fewbit"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
open-mmlab/mmengine
OpenMMLab Foundational Library for Training Deep Learning Models
Xilinx/brevitas
Brevitas: neural network quantization in PyTorch
google/qkeras
QKeras: a quantization deep learning library for Tensorflow Keras
fastmachinelearning/qonnx
QONNX: Arbitrary-Precision Quantized Neural Networks in ONNX
tensorflow/model-optimization
A toolkit to optimize ML models for deployment for Keras and TensorFlow, including quantization...