flash attention jax

闪光注意 - Jax

在 Jax 中实现 Flash Attention。由于缺乏精细内存管理的能力，它的性能可能不如官方 CUDA 版本。但仅出于教育目的以及了解 XLA 编译器有多聪明（或不聪明）。

安装

$ pip install flash-attention-jax

用法

 from jax import random
from flash_attention_jax import flash_attention

rng_key = random . PRNGKey ( 42 )

q = random . normal ( rng_key , ( 1 , 2 , 131072 , 512 ))  # (batch, heads, seq, dim)
k = random . normal ( rng_key , ( 1 , 2 , 131072 , 512 ))
v = random . normal ( rng_key , ( 1 , 2 , 131072 , 512 ))
mask = random . randint ( rng_key , ( 1 , 131072 ,), 0 , 2 ) # (batch, seq)

out , _ = flash_attention ( q , k , v , mask )

out . shape  # (1, 2, 131072, 512) - (batch, heads, seq, dim)

快速健全性检查

 from flash_attention_jax import plain_attention , flash_attention , value_and_grad_difference

diff , ( dq_diff , dk_diff , dv_diff ) = value_and_grad_difference (
    plain_attention ,
    flash_attention ,
    seed = 42
)

print ( 'shows differences between normal and flash attention for output, dq, dk, dv' )
print ( f'o: { diff } ' )       # < 1e-4
print ( f'dq: { dq_diff } ' )   # < 1e-6
print ( f'dk: { dk_diff } ' )   # < 1e-6
print ( f'dv: { dv_diff } ' )   # < 1e-6

Autoregressive Flash Attention - 类似 GPT 的解码器注意力

 from jax import random
from flash_attention_jax import causal_flash_attention

rng_key = random . PRNGKey ( 42 )

q = random . normal ( rng_key , ( 131072 , 512 ))
k = random . normal ( rng_key , ( 131072 , 512 ))
v = random . normal ( rng_key , ( 131072 , 512 ))

out , _ = causal_flash_attention ( q , k , v )

out . shape  # (131072, 512) 

托多

• 因果闪光注意变体的主要维度

• 找出 jit 和 static argnums 的问题

• 参考论文算法和解释进行评论

• 确保它可以工作单向键/值，就像在 PaLM 中一样

引文

 @article { Dao2022FlashAttentionFA ,
    title   = { FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness } ,
    author  = { Tri Dao and Daniel Y. Fu and Stefano Ermon and Atri Rudra and Christopher R'e } ,
    journal = { ArXiv } ,
    year    = { 2022 } ,
    volume  = { abs/2205.14135 }
}

 @article { Rabe2021SelfattentionDN ,
    title   = { Self-attention Does Not Need O(n2) Memory } ,
    author  = { Markus N. Rabe and Charles Staats } ,
    journal = { ArXiv } ,
    year    = { 2021 } ,
    volume  = { abs/2112.05682 }
}