GQA分组注意力机制

一、目录

1. 定义
2. demo

二、实现

1. 定义
   grouped query attention（GQA）
   1 GQA 原理与优点：将query 进行分组，每组query 参数共享一份key,value, 从而使key, value 矩阵变小。
   2. 优点： 降低内存读取模型权重的时间开销：由于Key矩阵和Value矩阵数量变少了，因此权重参数量也减少了，需要读取到内存的数量量少了，因此减少了读取权重的等待时间。
   3. 效果(并未降低模型性能)：GQA通过设置合适的分组大小，可以和MQA的推理性能几乎相等，同时逼近MHA的模型性能。
2. llama3 分组数为4， chatglm2 分组数为2 .
   
   
   参考：https://zhuanlan.zhihu.com/p/693928854
   demo

import torch
import torch.nn as nn
import math

#GQA
bs=3
seq_len =5
hidden_size= 32
n_heads=4
n_kv_heads = 2
head_dim = hidden_size//n_heads #
groups = n_heads//n_kv_heads # 4/2
print("groups=",groups)
x=torch.randn((bs,seq_len,hidden_size))
print("x:", x.shape)
wq = nn.Linear(hidden_size,n_heads*head_dim,bias=False)
wk = nn.Linear(hidden_size, n_kv_heads * head_dim, bias=False)
wv = nn.Linear(hidden_size, n_kv_heads * head_dim, bias=False)
xq,xk,xv=wq(x),wk(x),wv(x)
xq = xq.view(bs,seq_len, n_heads, head_dim).transpose(1, 2)
xk = xk.view(bs,seq_len, n_kv_heads, head_dim).transpose(1, 2)
xv = xv.view(bs,seq_len, n_kv_heads, head_dim).transpose(1, 2)
print("xq:",xq.shape) #[bs,n_heads,seq_len, head_dim]
print("xk:", xk.shape)#[bs,n_kv_heads,seq_len, head_dim]
print("xv:", xv.shape)#[bs,n_kv_heads,seq_len, head_dim]
def repeat_kv(keys: torch.Tensor, values: torch.Tensor, repeats: int, dim: int):
    keys = torch.repeat_interleave(keys, repeats=repeats, dim=dim)
    values = torch.repeat_interleave(values, repeats=repeats, dim=dim)
    return keys, values
#复制kv head
key,val = repeat_kv(xk,xv, groups,dim=1)
print("key:", key.shape)
print("val:", val.shape)
attn_weights = torch.matmul(xq, key.transpose(2, 3)) / math.sqrt(head_dim)
print("attn_weights:", attn_weights.shape) #[bs,n_heads,seq_len,seq_len]
attn_output = torch.matmul(attn_weights, val)
print("attn_output:", attn_output.shape)  # [bs,n_heads,seq_len,head_dim]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37