理解Attention,MHA、MQA、GQA理论知识和代码实现

理论知识链接:理解Attention:从起源到MHA,MQA和GQA | Linsight

现有模型升级方法:https://blog.nghuyong.top/2023/09/10/NLP/llm-attention/

pytorch代码实现:

复制代码
class BaseAttention(torch.nn.Module):
    def __init__(self):
        super(BaseAttention, self).__init__()
        self.softmax = torch.nn.Softmax(dim=-1)

    def attention(self, q, k, v, mask=None, dropout=None):
        attn = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(q.shape[-1])

        if mask is not None:
            attn = attn + mask
        
        attn = self.softmax(attn)
        if dropout is not None:
            attn = dropout(attn)
        output = torch.matmul(attn, v)
        return output


class Attention(BaseAttention):

    def __init__(self, hidden_size, dropout=None):
        super(Attention, self).__init__()
        self.q_proj = torch.nn.Linear(hidden_size, hidden_size)
        self.k_proj = torch.nn.Linear(hidden_size, hidden_size)
        self.v_proj = torch.nn.Linear(hidden_size, hidden_size)
        self.softmax = torch.nn.Softmax(dim=-1)
        
        if dropout is not None:
            self.dropout = torch.nn.Dropout(p=dropout)
        else:
            self.dropout = None
    
    def forward(self, x, mask=None):
        q = self.q_proj(x)
        k = self.k_proj(x)
        v = self.v_proj(x)
        output = self.attention(q, k, v, mask, self.dropout)
        return output


class MHAttention(BaseAttention):

    def __init__(self, hidden_size, num_heads=32, dropout=None):
        super(MHAttention, self).__init__()
        self.num_heads = num_heads
        self.softmax = torch.nn.Softmax(dim=-1)
        self.q_proj = torch.nn.Linear(hidden_size, hidden_size)
        self.k_proj = torch.nn.Linear(hidden_size, hidden_size)
        self.v_proj = torch.nn.Linear(hidden_size, hidden_size)
        
        if dropout is not None:
            self.dropout = torch.nn.Dropout(p=dropout)
    
    def forward(self, x, mask=None):
        bs, seq_len, hidden_size = x.shape

        q = self.q_proj(x).view(bs, seq_len, self.num_heads, -1).transpose(1, 2)
        k = self.k_proj(x).view(bs, seq_len, self.num_heads, -1).transpose(1, 2)
        v = self.v_proj(x).view(bs, seq_len, self.num_heads, -1).transpose(1, 2)
        output = self.attention(q, k, v, mask, self.dropout)
        output = output.view(bs, seq_len, hidden_size)
        return output


class MQAttention(BaseAttention):

    def __init__(self, hidden_size, num_heads=32, dropout=None):
        super(MQAttention, self).__init__()
        self.num_heads = num_heads
        self.softmax = torch.nn.Softmax(dim=-1)
        assert hidden_size % num_heads == 0
        self.q_proj = torch.nn.Linear(hidden_size, hidden_size)
        self.k_proj = torch.nn.Linear(hidden_size, hidden_size // num_heads)
        self.v_proj = torch.nn.Linear(hidden_size, hidden_size // num_heads)
        
        if dropout is not None:
            self.dropout = torch.nn.Dropout(p=dropout)
    
    def forward(self, x, mask=None):
        bs, seq_len, hidden_size = x.shape

        q = self.q_proj(x).view(bs, seq_len, self.num_heads, -1).transpose(1, 2)
        k = self.k_proj(x).view(bs, seq_len, -1, hidden_size // self.num_heads).transpose(1, 2)
        v = self.v_proj(x).view(bs, seq_len, -1, hidden_size // self.num_heads).transpose(1, 2)
        output = self.attention(q, k, v, mask, self.dropout)
        output = output.view(bs, seq_len, hidden_size)
        return output


class GQAttention(BaseAttention):

    def __init__(self, hidden_size, num_heads=32, num_kv_heads=8, dropout=None):
        super(GQAttention, self).__init__()
        assert hidden_size % num_heads == 0 and num_heads % num_kv_heads == 0

        self.num_heads = num_heads
        self.num_kv_heads = num_kv_heads
        self.num_group = num_heads // num_kv_heads
        self.softmax = torch.nn.Softmax(dim=-1)
        self.q_proj = torch.nn.Linear(hidden_size, hidden_size)
        self.k_proj = torch.nn.Linear(hidden_size, hidden_size // num_heads * num_kv_heads)
        self.v_proj = torch.nn.Linear(hidden_size, hidden_size // num_heads * num_kv_heads)
        
        if dropout is not None:
            self.dropout = torch.nn.Dropout(p=dropout)
    
    def repeat_kv(self, feature, num_group): #llama2源码
        bs, num_kv_heads, seq_len, head_dims = feature.shape
        if num_group == 1:
            return feature
        feature = feature[:, :, None, :, :].expand(bs, num_kv_heads, num_group, seq_len, head_dims)
        return feature.reshape(bs, num_kv_heads * num_group, seq_len, head_dims)

    def forward(self, x, mask=None):
        bs, seq_len, hidden_size = x.shape

        q = self.q_proj(x).view(bs, seq_len, self.num_heads, -1).transpose(1, 2)
        k = self.k_proj(x).view(bs, seq_len, -1, hidden_size // self.num_heads).transpose(1, 2)
        v = self.v_proj(x).view(bs, seq_len, -1, hidden_size // self.num_heads).transpose(1, 2)
        k, v = self.repeat_kv(k, self.num_group), self.repeat_kv(v, self.num_group)
        output = self.attention(q, k, v, mask, self.dropout)
        output = output.view(bs, seq_len, hidden_size)
        return output
        

model = Attention(hidden_size=4096, dropout=0.1)
model = MHAttention(hidden_size=4096, num_heads=32, dropout=0.1)
model = MQAttention(hidden_size=4096, num_heads=32, dropout=0.1)
model = GQAttention(hidden_size=4096, num_heads=32, num_kv_heads=4, dropout=0.1)
input_data = torch.randn(1, 20, 4096)
output = model(input_data)
print()
相关推荐
羑悻的小杀马特24 分钟前
OpenCV 引擎:驱动实时应用开发的科技狂飙
人工智能·科技·opencv·计算机视觉
guanshiyishi3 小时前
ABeam 德硕 | 中国汽车市场(2)——新能源车的崛起与中国汽车市场机遇与挑战
人工智能
极客天成ScaleFlash4 小时前
极客天成NVFile:无缓存直击存储性能天花板,重新定义AI时代并行存储新范式
人工智能·缓存
Uzuki4 小时前
AI可解释性 II | Saliency Maps-based 归因方法(Attribution)论文导读(持续更新)
深度学习·机器学习·可解释性
澳鹏Appen5 小时前
AI安全:构建负责任且可靠的系统
人工智能·安全
蹦蹦跳跳真可爱5895 小时前
Python----机器学习(KNN:使用数学方法实现KNN)
人工智能·python·机器学习
视界宝藏库6 小时前
多元 AI 配音软件,打造独特音频体验
人工智能
xinxiyinhe6 小时前
GitHub上英语学习工具的精选分类汇总
人工智能·deepseek·学习英语精选
byxdaz6 小时前
PyTorch中Linear全连接层
pytorch
Start_Present7 小时前
Pytorch 第十二回:循环神经网络——LSTM模型
pytorch·rnn·神经网络·数据分析·lstm