FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

作者： Tri Dao (2023)

arXiv： 2307.08691

领域

推理

TLDR（中文）

用更激进的 warp 级并行和 work partition 把 FlashAttention 再翻倍。今天 vLLM / SGLang / Megatron 训练后端基本都升级到 FA-2。

TLDR (English)

Uses more aggressive warp-level parallelism and work partitioning to double FlashAttention performance. Today vLLM/SGLang/Megatron training backends have all upgraded to FA-2.

出现在这些文章里

• 高效注意力：突破序列长度平方瓶颈
• Efficient Attention: Breaking the Quadratic Sequence Bottleneck

同被引用

这些论文与本文出现在同一篇文章中

• GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints Joshua Ainslie et al. (2023)
• FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness Tri Dao et al. (2022)
• DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model DeepSeek-AI (2024)
• FlashAttention-3: Fast and Accurate Attention with Asympotic IO Complexity Jay Shah et al. (2024)

相关论文

同一领域的其他论文

• DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model DeepSeek-AI (2024)
  架构推理
• FlashAttention-3: Fast and Accurate Attention with Asympotic IO Complexity Jay Shah et al. (2024)
  推理
• Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters Charlie Snell et al. (2024)
  推理能力推理
• GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints Joshua Ainslie et al. (2023)
  推理架构
• Accelerating Large Language Model Decoding with Speculative Sampling Charlie Chen et al. (2023)
  推理
• QLoRA: Efficient Finetuning of Quantized LLMs Tim Dettmers et al. (2023)
  对齐推理