MLA 结构代码实现及优化

# 从 LlamaAttention 修改而来，适配 DeepseekV2 模型的注意力模块，简单版本不带 kv cache
class DeepseekV2MLA(nn.Module):
    def __init__(self, config: DeepseekV2Config):
        super().__init__()
        # MHA 初始化相关
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.v_head_dim = config.v_head_dim

        self.o_proj = nn.Linear(
            self.v_head_dim * self.num_heads, 
            self.hidden_size,
            bias=config.attention_bias,
        )

        self.attention_dropout = config.attention_dropout
        self.training = False
        self.qk_nope_head_dim = config.qk_nope_head_dim
        self.qk_rope_head_dim = config.qk_rope_head_dim

        # MLA 相关 part1: 压缩
        self.q_lora_rank = config.q_lora_rank
        self.kv_lora_rank = config.kv_lora_rank

        self.q_down_proj = nn.Linear(self.hidden_size, self.q_lora_rank)
        self.q_down_rmsnorm = DeepseekV2RMSNorm(self.q_lora_rank)
        
        self.kv_down_proj = nn.Linear(
            self.hidden_size, 
            self.kv_lora_rank + config.qk_rope_head_dim
        )
        self.kv_down_rmsnorm = DeepseekV2RMSNorm(self.kv_lora_rank)
        
        # MLA 相关 part2: 解压缩. # W^{WQ} 和 W^{QR} 权重是合并再一起的。
        self.qk_head_dim = self.qk_nope_head_dim  + self.qk_rope_head_dim
        self.q_up_proj = nn.Linear(
            self.q_lora_rank, 
            self.num_heads * self.qk_head_dim,
            bias=False,
        )
        
        self.kv_up_proj = nn.Linear(
            self.kv_lora_rank, 
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
        )
        
        # MLA 相关 part3: 切片 q k 张量，以及 rope 旋转位置编码
        self.rotary_emb = DeepseekV2RotaryEmbedding(
            config.qk_rope_head_dim,
            config.max_position_embeddings,
            config.rope_theta,
        ) 

    def forward(self, hidden_states, position_ids, casual_mask=None):
        batch_size, q_len, hidden_size = hidden_states.shape

        # 1，q 压缩和解压缩，以及 split to q_nope, q_rope
        q = self.q_up_proj(
            self.q_down_rmsnorm(self.q_down_proj(hidden_states))
        )

        q = q.view(batch_size, q_len, self.num_heads, self.qk_head_dim).transpose(1,2)
        q_nope, q_rope = torch.split(
            q,
            [self.qk_nope_head_dim, self.qk_rope_head_dim],
            dim = -1,
        )

        # 2, kv 压缩和解压缩
        kv_down = self.kv_down_proj(hidden_states)
        
        # compressed_kv 压缩后的 kv 张量
        compressed_kv, k_rope = torch.split(
            kv_down,
            [self.kv_lora_rank, self.qk_rope_head_dim],
            dim = -1,
        )
        # num_heads = 1 后续广播其它 heads 上
        k_rope = k_rope.view(batch_size, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)

        # 对 compressed_kv 解压缩
        kv = (
            self.kv_up_proj(self.kv_down_rmsnorm(compressed_kv))
            .view(batch_size, q_len, self.num_heads, self.qk_nope_head_dim + self.v_head_dim)
            .transpose(1, 2)
        )

        k_nope, value_states = torch.split(
            kv,
            [self.qk_nope_head_dim, self.v_head_dim],
            dim = -1,
        )

        # 3, 计算 cos 和 sin，并应用 rope 旋转位置编码
        kv_seq_len = value_states.shape[-2] # shape (b, nums_head, seq_len, v_head_dim)
        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
        
        q_rope, k_rope = apply_rotary_pos_emb(q_rope, k_rope, cos, sin, position_ids)

        # 4, 执行 self-attention 计算
        query_states = torch.concat([q_nope, q_rope], dim=-1)
        key_states = torch.concat(
            [k_nope, k_rope.expand(-1, self.num_heads, -1, -1)], 
            dim=-1
        )
        # qk^t
        scores = (
            torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.qk_head_dim)
        )

        if casual_mask is not None:
            scores = scores.masked_fill(casual_mask == 0, float('-inf'))
        
        attn_weights = F.softmax(scores, dim=-1).to(query_states.dtype)
        attn_weights = F.dropout(
            attn_weights, p=self.attention_dropout, training=self.training
        ) # attn_weights shape: [bs, num_heads, seq_len, seq_len]
        
        attn_output = torch.matmul(attn_weights, value_states) # shape: [bs, num_heads, seq_len, head_dim]
        attn_output = attn_output.transpose(1, 2).contiguous().reshape(batch_size, q_len, self.num_heads * self.v_head_dim)

        # 5, MLA 输出映射
        output = self.o_proj(attn_output)

        return output, attn_weights

# (1) 生成多头 Value：v_t ∈ ℝ^{b×l×h×d}
v_t = einsum('hdc, blc -> blhd', W_UV, c_t_KV)

# (2) 按注意力权重加权求和：o ∈ ℝ^{b×q×h×d}
o   = einsum('bqhl, blhd -> bqhd', a, v_t)

# (3) 输出投影：u ∈ ℝ^{b×h×D}
u   = einsum('hdD, bhqd -> bhD', W_o, o)

# 将上述三式合并，得到总的计算过程
u   = einsum('hdc,blc,bqhl,hdD->bhD', W_UV, c_t_KV, attn_weights, W_o)

# (4) 先把注意力权重直接乘到原始上下文，得到 o_ ∈ ℝ^{b×h×q×c}
o_  = einsum('bqhl, blc -> bhqc', attn_weights, c_t_KV)

# (5) 再做 Value 投影：o ∈ ℝ^{b×h×q×d}
o   = einsum('bhqc, hdc->bhqd', o_, W_UV)  # (5)

# (6) 最终输出：u ∈ ℝ^{b×h×D}
u   = einsum('hdD, bhqd -> bhD', W_o, o)

import torch
import time

# 配置参数
b, q, l, h, d, c, D = 32, 64, 128, 64, 64, 128, 256  # 将 h 调整为 64
n_warmup = 10   # 预热次数
n_trials = 100  # 正式测试次数

# 初始化张量（GPU）
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

W_UV = torch.randn(h, d, c, device=device)
c_t_KV = torch.randn(b, l, c, device=device)
attn_weights = torch.randn(b, q, h, l, device=device)
W_o = torch.randn(h, d, D, device=device)  # h 维度为 64

# 预热 GPU
for _ in range(n_warmup):
    _ = torch.einsum('hdc,blc->blhd', W_UV, c_t_KV)
    _ = torch.einsum('bqhl,blhd->bqhd', attn_weights, _)
    _ = torch.einsum('hdD,bhqd->bhD', W_o, _)

# 原始分步实现
def original_method():
    v_t = torch.einsum('hdc,blc->blhd', W_UV, c_t_KV)
    o = torch.einsum('bqhl,blhd->bqhd', attn_weights, v_t)
    u = torch.einsum('hdD,bhqd->bhD', W_o, o.permute(0, 2, 1, 3))

# 优化后实现
def optimized_method():
    o_ = torch.einsum('bhql,blc->bhqc', attn_weights.permute(0, 2, 1, 3), c_t_KV)
    o = torch.einsum('bhqc,hdc->bhqd', o_, W_UV)
    u = torch.einsum('hdD,bhqd->bhD', W_o, o)

# 测量时间
def benchmark(func):
    times = []
    for _ in range(n_trials):
        start = time.time()
        func()
        end = time.time()
        times.append(end - start)
    return sum(times) / n_trials

# 执行测试
time_original = benchmark(original_method) * 1000  # 转换为毫秒
time_optimized = benchmark(optimized_method) * 1000

# 打印结果
print(f"原始方法平均时间: {time_original:.3f} ms")
print(f"优化方法平均时间: {time_optimized:.3f} ms")
print(f"速度提升: {time_original / time_optimized - 1:.1%}")

# 验证等价性
def validate_equivalence():
    v_t_orig = torch.einsum('hdc,blc->blhd', W_UV, c_t_KV)
    o_orig = torch.einsum('bqhl,blhd->bqhd', attn_weights, v_t_orig)
    u_orig = torch.einsum('hdD,bhqd->bhD', W_o, o_orig.permute(0, 2, 1, 3))
    
    v_t_opt = torch.einsum('hdc,blc->blhd', W_UV, c_t_KV)
    o_opt = torch.einsum('bqhl,blhd->bqhd', attn_weights, v_t_opt)
    u_opt = torch.einsum('hdD,bhqd->bhD', W_o, o_opt.permute(0, 2, 1, 3))
    
    # 检查是否等价
    assert torch.allclose(u_orig, u_opt, atol=1e-4), "两种方法结果不一致！"
    print("两种方法结果一致，验证通过。")

# 调用验证函数
validate_equivalence()

"""
原始方法平均时间: 28.649 ms
优化方法平均时间: 20.378 ms
速度提升: 40.6%
两种方法结果一致，验证通过。
"""