Download
FlashAttention is a high-performance deep learning optimization library that reimplements the attention mechanism used in transformer models to be significantly faster and more memory-efficient than standard implementations. It achieves this by using IO-aware algorithms that minimize memory reads and writes, reducing the quadratic memory overhead typically associated with attention operations. The project provides implementations of FlashAttention, FlashAttention-2, and newer iterations optimized for modern GPU architectures such as NVIDIA Hopper and AMD accelerators. By improving both forward and backward pass efficiency, it enables training and inference of large language models with longer sequence lengths and higher throughput. The library integrates with PyTorch and supports various attention configurations, including causal masking, multi-query attention, and rotary embeddings.
Features
- Memory-efficient attention with linear scaling instead of quadratic overhead
- High-performance GPU kernels optimized for CUDA and ROCm
- Support for FlashAttention-2 and newer optimized implementations
- Integration with PyTorch and modern transformer architectures
- Support for causal masking, rotary embeddings, and advanced attention types
- Enables faster training and inference for large-scale AI models
Project Samples
