Search Results for "bit"

...The project’s focus on extreme quantization dramatically reduces memory footprint and energy consumption compared with traditional 16-bit or 32-bit LLMs, making it practical to deploy advanced language understanding and generation models on everyday machines. BitNet is built to scale across architectures, with configurable kernels and tiling strategies that adapt to different hardware, and it supports large models with impressive throughput even on modest resources.

BitNet is a machine learning research implementation that explores extremely low-precision neural network architectures designed to dramatically reduce the computational cost of large language models. The project implements the BitNet architecture described in research on scaling transformer models using extremely low-bit quantization techniques. In this approach, neural network weights are quantized to approximately one bit per parameter, allowing models to operate with far lower memory usage than traditional 16-bit or 32-bit neural networks. The architecture introduces specialized layers such as BitLinear, which replace standard linear projections in transformer networks with quantized operations. ...

FastDeploy is an open-source inference and deployment toolkit designed to simplify the process of running and serving deep learning models across a wide range of hardware platforms. Developed within the PaddlePaddle ecosystem, the toolkit focuses on providing high-performance deployment capabilities for modern AI models including large language models and vision-language systems. The platform enables developers to deploy trained models quickly using optimized inference pipelines that support...

TurboDiffusion is an advanced open-source framework designed to dramatically accelerate video diffusion model generation, aiming for performance improvements on the order of 100–200× compared with traditional implementations while retaining high output quality. It achieves this by combining a suite of algorithmic and engineering optimizations, including attention acceleration techniques, efficient step distillation methods, and quantization strategies that reduce computational overhead. The...

SageAttention is an open-source optimization library designed to accelerate the attention mechanism used in transformer-based neural networks. Since attention operations are often the most computationally expensive component of modern AI models, SageAttention introduces quantization techniques that significantly reduce computational overhead while preserving model accuracy. The system achieves this by using low-precision numerical formats such as INT4, FP8, or INT8 to represent key matrices...

...In experiments on LJSpeech, HiFi-GAN was shown to achieve mean opinion scores close to human recordings while synthesizing 22.05 kHz audio up to ~168× faster than real time on an NVIDIA V100 GPU. A smaller configuration trades a bit of quality for even higher speed and can run more than 13× faster than real time on CPU, making it suitable for deployment scenarios without powerful GPUs.

Linux kernel fork-off with cloud and machine learning features