The easiest way to run an open model on your own machine: one command downloads and starts it, plus a local REST API and libraries for Python and JavaScript.
All repositories
Open models & inferenceDao-AILab
flash-attention
A fast, memory-efficient implementation of attention that large-model training and inference on GPUs lean on. The result is exact, with no approximation.
Who it is for
For people speeding up transformer training and inference on NVIDIA GPUs.
How to start
- Install
packaging,psutilandninja, which the build needs. - Install:
pip install flash-attn --no-build-isolation. - Call
flash_attn_func(q, k, v, causal=True)instead of standard attention; on Hopper and Blackwell there ispip install flash-attn-4.
Steps are taken from the README. Check the current version in the repository before running them.
Stars over the last 30 days
+269Sep 6 — Oct 5
Author's description
Fast and memory-efficient exact attention
More in «Open models & inference»
The core library for working with models: one codebase for text, vision, audio and multimodal tasks, for both inference and training. Most new open models ship through it.
A self-hosted ChatGPT-style interface that connects to Ollama and any OpenAI-compatible API. Installs with a single Docker command and runs on your own server.
A C/C++ engine that runs language models on ordinary hardware: laptops, phones, GPU-less servers. Much of local AI, Ollama included, is built on it.
Repository of the large open DeepSeek-V3 mixture-of-experts model: description, benchmark results and run instructions. It showed an open model can stand next to closed ones.
The foundation under most modern neural networks: GPU-accelerated tensors and automatic differentiation, wrapped in ordinary Python. Transformers, Llama and nearly every open model are written on it.