llama_modeling/diff_attn.py

from flash_attn import flash_attn_func
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat

from .extact import xATGLU
from .liger_rope import LigerRopeFunction
from .config import LlamaConfig

# The four-flash attn strategy comes from here:
# https://github.com/microsoft/unilm/blob/master/Diff-Transformer/multihead_flashdiff_2.py

class DifferentialAttention(nn.Module):
    def __init__(self, config: LlamaConfig, layer_num):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.num_kv_heads = config.num_key_value_heads
        self.n_rep = self.num_heads // self.num_kv_heads
        self.head_dim = self.hidden_size // (2 * self.num_heads)
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.scaling = self.head_dim ** -0.5

        self.q_proj = nn.Linear(self.hidden_size, 2 * self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, 2 * self.num_kv_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, 2 * self.num_kv_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(2 * self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.lambda_init = 0.8 - 0.6 * math.exp(-0.3 * layer_num)
        self.lambda_q1 = nn.Parameter(torch.zeros(self.head_dim).normal_(0, 0.1))
        self.lambda_k1 = nn.Parameter(torch.zeros(self.head_dim).normal_(0, 0.1))
        self.lambda_q2 = nn.Parameter(torch.zeros(self.head_dim).normal_(0, 0.1))
        self.lambda_k2 = nn.Parameter(torch.zeros(self.head_dim).normal_(0, 0.1))

        self.subln = nn.LayerNorm(2 * self.head_dim, elementwise_affine=False)

        self.register_buffer(
            "cos_cached",
            self._compute_rope_embeddings(
                self.max_position_embeddings,
                self.head_dim,
                self.rope_theta,
                dtype=torch.float32,
                device=self.q_proj.weight.device,
            )[0],
            persistent=False,
        )
        self.register_buffer(
            "sin_cached",
            self._compute_rope_embeddings(
                self.max_position_embeddings,
                self.head_dim,
                self.rope_theta,
                dtype=torch.float32,
                device=self.q_proj.weight.device,
            )[1],
            persistent=False,
        )

    def _compute_rope_embeddings(self, max_position_embeddings, head_dim, base=10000, dtype=None, device=None):
        inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2, device=device).float() / head_dim))
        t = torch.arange(max_position_embeddings, device=device, dtype=torch.float32)
        freqs = torch.einsum("i,j->ij", t, inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        cos = emb.cos().to(dtype)
        sin = emb.sin().to(dtype)
        return cos.unsqueeze(0), sin.unsqueeze(0)

    def forward(
        self,
        hidden_states,
        attention_mask,
        position_ids,
    ) -> torch.Tensor:
        bsz, seq_len, embed_dim = hidden_states.size()
        
        if position_ids is None:
            position_ids = torch.arange(seq_len, device=hidden_states.device)
            position_ids = repeat(position_ids, 'l -> b l', b=bsz)

        q = self.q_proj(hidden_states)
        k = self.k_proj(hidden_states)
        v = self.v_proj(hidden_states)

        q = rearrange(q, 'b s (h d) -> b s h d', h=2*self.num_heads, d=self.head_dim)
        k = rearrange(k, 'b s (h d) -> b s h d', h=2*self.num_kv_heads, d=self.head_dim)
        
        # Reshaped for GQA
        v = rearrange(v, 'b s (h g d) -> b s h g d', h=self.num_kv_heads, g=2, d=self.head_dim)

        # Apply rotary embeddings using LigerRopeFunction            
        cos = self.cos_cached[:, position_ids]  # [1, bsz, seq_len, dim]
        sin = self.sin_cached[:, position_ids]  # [1, bsz, seq_len, dim]
        q, k = LigerRopeFunction.apply(q, k, cos, sin, position_ids)
        
        # Rearrange into GQA style
        q = rearrange(q, 'b s (h g) d -> b s h g d', h=self.num_heads, g=2)
        k = rearrange(k, 'b s (h g) d -> b s h g d', h=self.num_kv_heads, g=2)
        
        q1, q2 = q[:, :, :, 0], q[:, :, :, 1]
        k1, k2 = k[:, :, :, 0], k[:, :, :, 1]
        v1, v2 = v[:, :, :, 0], v[:, :, :, 1]

        # First attention group on q1/k1 and the v's
        attn11 = flash_attn_func(
            q1,
            k1,
            v1,
            dropout_p=0.0, # @Z TODO::
            causal=attention_mask is None
        ) 
        attn12 = flash_attn_func(
            q1,
            k1,
            v2,
            dropout_p=0.0,
            causal=attention_mask is None
        )
        attn1 = torch.cat([attn11, attn12], dim=-1)

        # Second attention group on q2/k2 and the v's
        attn21 = flash_attn_func(
            q2,
            k2,
            v1,
            dropout_p=0.0,
            causal=attention_mask is None
        ) 
        attn22 = flash_attn_func(
            q2,
            k2,
            v2,
            dropout_p=0.0,
            causal=attention_mask is None
        )
        attn2 = torch.cat([attn21, attn22], dim=-1)

        lambda_1 = torch.exp(torch.sum(self.lambda_q1 * self.lambda_k1, dim=-1).float()).type_as(q)
        lambda_2 = torch.exp(torch.sum(self.lambda_q2 * self.lambda_k2, dim=-1).float()).type_as(q)
        lambda_full = lambda_1 - lambda_2 + self.lambda_init
        attn = attn1 - lambda_full * attn2

        attn = self.subln(attn)
        attn = attn * (1 - self.lambda_init)

        attn_output = rearrange(attn, "b s h d -> b s (h d)")
        return self.o_proj(attn_output)