什么是GQA？

多个head的Query共用一组K和V。llama模型就用到该技术。

需要明确几点：

1.group有几组

2.每个group对应几个head

3.q以head为单位 k,v以group为单位 每个head/group特征维度都是head_dim

代码实现

import torch.nn as nn
import torch
import math# 自注意力
class GroupQueryAttention(nn.Module):def __init__(self, d_model, n_heads, n_groups):super().__init__()self.d_model = d_modelself.n_heads = n_headsself.n_groups = n_groupsassert d_model % n_heads == 0self.head_dim = d_model // n_heads# 每个group对应几个headself.heads_per_group = n_heads // n_groups# 以head为单位self.w_q = nn.Linear(d_model, n_heads*self.head_dim)  # n_heads*head_dim=d_model# 以group为单位self.w_k = nn.Linear(d_model, n_groups*self.head_dim)self.w_v = nn.Linear(d_model, n_groups*self.head_dim)self.w_combine = nn.Linear(d_model, d_model)self.softmax = nn.Softmax(dim=-1)# 给k,v进行复制，假设每个组对应3个head，那就要把每个组的数据复制3遍def expand(self, data): # data:k/v [b, group, seq_len, head_dim]b,_,seq_len,_ = data.shapedata = data[:,:,None,:,:].expand(b, self.n_groups, self.heads_per_group, seq_len, self.head_dim)data = data.contiguous().view(b, -1, seq_len, self.head_dim)return data  # [b, group*heads_per_group, seq_len, head_dim]def forward(self, x, use_mask=False): # x: [b, seq_len, d_model]b, seq_len, _ = x.shapeq,k,v = self.w_q(x), self.w_k(x), self.w_v(x)q = q.view(b, seq_len, self.n_heads, self.head_dim).permute(0,2,1,3)  # 以head为单位k = k.view(b, seq_len, self.n_groups, self.head_dim).permute(0,2,1,3) # 以group为单位v = v.view(b, seq_len, self.n_groups, self.head_dim).permute(0,2,1,3)# 复制k,v = self.expand(k), self.expand(v)score = q @ k.transpose(-1,-2) / math.sqrt(self.head_dim)if use_mask:mask = torch.tril(torch.ones(seq_len, seq_len))score = score.masked_fill(mask==0, float('-inf'))score = self.softmax(score) @ vscore = score.permute(0,2,1,3).contiguous().view(b, seq_len, self.n_heads*self.head_dim)out = self.w_combine(score)return outx = torch.rand(2,100,384)
model = GroupQueryAttention(384, 4, 2)
out = model(x)
print(out.shape)