import torch import torch.nn as nn from typing import Tuple, Union, Optional try: import flash_attn if hasattr(flash_attn, '__version__') and int(flash_attn.__version__[0]) == 2: from flash_attn.flash_attn_interface import flash_attn_kvpacked_func from flash_attn.modules.mha import FlashSelfAttention, FlashCrossAttention else: from flash_attn.flash_attn_interface import flash_attn_unpadded_kvpacked_func from flash_attn.modules.mha import FlashSelfAttention, FlashCrossAttention except Exception as e: print(f'flash_attn import failed: {e}') def reshape_for_broadcast(freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], x: torch.Tensor, head_first=False): """ Reshape frequency tensor for broadcasting it with another tensor. This function reshapes the frequency tensor to have the same shape as the target tensor 'x' for the purpose of broadcasting the frequency tensor during element-wise operations. Args: freqs_cis (Union[torch.Tensor, Tuple[torch.Tensor]]): Frequency tensor to be reshaped. x (torch.Tensor): Target tensor for broadcasting compatibility. head_first (bool): head dimension first (except batch dim) or not. Returns: torch.Tensor: Reshaped frequency tensor. Raises: AssertionError: If the frequency tensor doesn't match the expected shape. AssertionError: If the target tensor 'x' doesn't have the expected number of dimensions. """ ndim = x.ndim assert 0 <= 1 < ndim if isinstance(freqs_cis, tuple): # freqs_cis: (cos, sin) in real space if head_first: assert freqs_cis[0].shape == (x.shape[-2], x.shape[-1]), f'freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}' shape = [d if i == ndim - 2 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] else: assert freqs_cis[0].shape == (x.shape[1], x.shape[-1]), f'freqs_cis shape {freqs_cis[0].shape} does not match x shape {x.shape}' shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] return freqs_cis[0].view(*shape), freqs_cis[1].view(*shape) else: # freqs_cis: values in complex space if head_first: assert freqs_cis.shape == (x.shape[-2], x.shape[-1]), f'freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}' shape = [d if i == ndim - 2 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] else: assert freqs_cis.shape == (x.shape[1], x.shape[-1]), f'freqs_cis shape {freqs_cis.shape} does not match x shape {x.shape}' shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)] return freqs_cis.view(*shape) def rotate_half(x): x_real, x_imag = x.float().reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, S, H, D//2] return torch.stack([-x_imag, x_real], dim=-1).flatten(3) def apply_rotary_emb( xq: torch.Tensor, xk: Optional[torch.Tensor], freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]], head_first: bool = False, ) -> Tuple[torch.Tensor, torch.Tensor]: """ Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings to the given query 'xq' and key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting tensors contain rotary embeddings and are returned as real tensors. Args: xq (torch.Tensor): Query tensor to apply rotary embeddings. [B, S, H, D] xk (torch.Tensor): Key tensor to apply rotary embeddings. [B, S, H, D] freqs_cis (Union[torch.Tensor, Tuple[torch.Tensor]]): Precomputed frequency tensor for complex exponentials. head_first (bool): head dimension first (except batch dim) or not. Returns: Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings. """ xk_out = None if isinstance(freqs_cis, tuple): cos, sin = reshape_for_broadcast(freqs_cis, xq, head_first) # [S, D] cos, sin = cos.to(xq.device), sin.to(xq.device) xq_out = (xq.float() * cos + rotate_half(xq.float()) * sin).type_as(xq) if xk is not None: xk_out = (xk.float() * cos + rotate_half(xk.float()) * sin).type_as(xk) else: xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2)) # [B, S, H, D//2] freqs_cis = reshape_for_broadcast(freqs_cis, xq_, head_first).to(xq.device) # [S, D//2] --> [1, S, 1, D//2] xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3).type_as(xq) if xk is not None: xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2)) # [B, S, H, D//2] xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3).type_as(xk) return xq_out, xk_out class FlashSelfMHAModified(nn.Module): """ Use QK Normalization. """ def __init__(self, dim, num_heads, qkv_bias=True, qk_norm=False, attn_drop=0.0, proj_drop=0.0, device=None, dtype=None, norm_layer=nn.LayerNorm, ): factory_kwargs = {'device': device, 'dtype': dtype} super().__init__() self.dim = dim self.num_heads = num_heads assert self.dim % num_heads == 0, "self.kdim must be divisible by num_heads" self.head_dim = self.dim // num_heads assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8" self.Wqkv = nn.Linear(dim, 3 * dim, bias=qkv_bias, **factory_kwargs) # TODO: eps should be 1 / 65530 if using fp16 self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.inner_attn = FlashSelfAttention(attention_dropout=attn_drop) self.out_proj = nn.Linear(dim, dim, bias=qkv_bias, **factory_kwargs) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x, freqs_cis_img=None): """ Parameters ---------- x: torch.Tensor (batch, seqlen, hidden_dim) (where hidden_dim = num heads * head dim) freqs_cis_img: torch.Tensor (batch, hidden_dim // 2), RoPE for image """ b, s, d = x.shape qkv = self.Wqkv(x) qkv = qkv.view(b, s, 3, self.num_heads, self.head_dim) # [b, s, 3, h, d] q, k, v = qkv.unbind(dim=2) # [b, s, h, d] q = self.q_norm(q).half() # [b, s, h, d] k = self.k_norm(k).half() # Apply RoPE if needed if freqs_cis_img is not None: qq, kk = apply_rotary_emb(q, k, freqs_cis_img) assert qq.shape == q.shape and kk.shape == k.shape, f'qq: {qq.shape}, q: {q.shape}, kk: {kk.shape}, k: {k.shape}' q, k = qq, kk qkv = torch.stack([q, k, v], dim=2) # [b, s, 3, h, d] context = self.inner_attn(qkv) out = self.out_proj(context.view(b, s, d)) out = self.proj_drop(out) out_tuple = (out,) return out_tuple class FlashCrossMHAModified(nn.Module): """ Use QK Normalization. """ def __init__(self, qdim, kdim, num_heads, qkv_bias=True, qk_norm=False, attn_drop=0.0, proj_drop=0.0, device=None, dtype=None, norm_layer=nn.LayerNorm, ): factory_kwargs = {'device': device, 'dtype': dtype} super().__init__() self.qdim = qdim self.kdim = kdim self.num_heads = num_heads assert self.qdim % num_heads == 0, "self.qdim must be divisible by num_heads" self.head_dim = self.qdim // num_heads assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8" self.scale = self.head_dim ** -0.5 self.q_proj = nn.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs) self.kv_proj = nn.Linear(kdim, 2 * qdim, bias=qkv_bias, **factory_kwargs) # TODO: eps should be 1 / 65530 if using fp16 self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.inner_attn = FlashCrossAttention(attention_dropout=attn_drop) self.out_proj = nn.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x, y, freqs_cis_img=None): """ Parameters ---------- x: torch.Tensor (batch, seqlen1, hidden_dim) (where hidden_dim = num_heads * head_dim) y: torch.Tensor (batch, seqlen2, hidden_dim2) freqs_cis_img: torch.Tensor (batch, hidden_dim // num_heads), RoPE for image """ b, s1, _ = x.shape # [b, s1, D] _, s2, _ = y.shape # [b, s2, 1024] q = self.q_proj(x).view(b, s1, self.num_heads, self.head_dim) # [b, s1, h, d] kv = self.kv_proj(y).view(b, s2, 2, self.num_heads, self.head_dim) # [b, s2, 2, h, d] k, v = kv.unbind(dim=2) # [b, s2, h, d] q = self.q_norm(q).half() # [b, s1, h, d] k = self.k_norm(k).half() # [b, s2, h, d] # Apply RoPE if needed if freqs_cis_img is not None: qq, _ = apply_rotary_emb(q, None, freqs_cis_img) assert qq.shape == q.shape, f'qq: {qq.shape}, q: {q.shape}' q = qq # [b, s1, h, d] kv = torch.stack([k, v], dim=2) # [b, s1, 2, h, d] context = self.inner_attn(q, kv) # [b, s1, h, d] context = context.view(b, s1, -1) # [b, s1, D] out = self.out_proj(context) out = self.proj_drop(out) out_tuple = (out,) return out_tuple class CrossAttention(nn.Module): """ Use QK Normalization. """ def __init__(self, qdim, kdim, num_heads, qkv_bias=True, qk_norm=False, attn_drop=0.0, proj_drop=0.0, device=None, dtype=None, norm_layer=nn.LayerNorm, ): factory_kwargs = {'device': device, 'dtype': dtype} super().__init__() self.qdim = qdim self.kdim = kdim self.num_heads = num_heads assert self.qdim % num_heads == 0, "self.qdim must be divisible by num_heads" self.head_dim = self.qdim // num_heads assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8" self.scale = self.head_dim ** -0.5 self.q_proj = nn.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs) self.kv_proj = nn.Linear(kdim, 2 * qdim, bias=qkv_bias, **factory_kwargs) # TODO: eps should be 1 / 65530 if using fp16 self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.attn_drop = nn.Dropout(attn_drop) self.out_proj = nn.Linear(qdim, qdim, bias=qkv_bias, **factory_kwargs) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x, y, freqs_cis_img=None): """ Parameters ---------- x: torch.Tensor (batch, seqlen1, hidden_dim) (where hidden_dim = num heads * head dim) y: torch.Tensor (batch, seqlen2, hidden_dim2) freqs_cis_img: torch.Tensor (batch, hidden_dim // 2), RoPE for image """ b, s1, c = x.shape # [b, s1, D] _, s2, c = y.shape # [b, s2, 1024] q = self.q_proj(x).view(b, s1, self.num_heads, self.head_dim) # [b, s1, h, d] kv = self.kv_proj(y).view(b, s2, 2, self.num_heads, self.head_dim) # [b, s2, 2, h, d] k, v = kv.unbind(dim=2) # [b, s, h, d] q = self.q_norm(q) k = self.k_norm(k) # Apply RoPE if needed if freqs_cis_img is not None: qq, _ = apply_rotary_emb(q, None, freqs_cis_img) assert qq.shape == q.shape, f'qq: {qq.shape}, q: {q.shape}' q = qq q = q * self.scale q = q.transpose(-2, -3).contiguous() # q -> B, L1, H, C - B, H, L1, C k = k.permute(0, 2, 3, 1).contiguous() # k -> B, L2, H, C - B, H, C, L2 attn = q @ k # attn -> B, H, L1, L2 attn = attn.softmax(dim=-1) # attn -> B, H, L1, L2 attn = self.attn_drop(attn) x = attn @ v.transpose(-2, -3) # v -> B, L2, H, C - B, H, L2, C x-> B, H, L1, C context = x.transpose(1, 2) # context -> B, H, L1, C - B, L1, H, C context = context.contiguous().view(b, s1, -1) out = self.out_proj(context) # context.reshape - B, L1, -1 out = self.proj_drop(out) out_tuple = (out,) return out_tuple class Attention(nn.Module): """ We rename some layer names to align with flash attention """ def __init__(self, dim, num_heads, qkv_bias=True, qk_norm=False, attn_drop=0., proj_drop=0., norm_layer=nn.LayerNorm, ): super().__init__() self.dim = dim self.num_heads = num_heads assert self.dim % num_heads == 0, 'dim should be divisible by num_heads' self.head_dim = self.dim // num_heads # This assertion is aligned with flash attention assert self.head_dim % 8 == 0 and self.head_dim <= 128, "Only support head_dim <= 128 and divisible by 8" self.scale = self.head_dim ** -0.5 # qkv --> Wqkv self.Wqkv = nn.Linear(dim, dim * 3, bias=qkv_bias) # TODO: eps should be 1 / 65530 if using fp16 self.q_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.k_norm = norm_layer(self.head_dim, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity() self.attn_drop = nn.Dropout(attn_drop) self.out_proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) def forward(self, x, freqs_cis_img=None): B, N, C = x.shape qkv = self.Wqkv(x).reshape(B, N, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4) # [3, b, h, s, d] q, k, v = qkv.unbind(0) # [b, h, s, d] q = self.q_norm(q) # [b, h, s, d] k = self.k_norm(k) # [b, h, s, d] # Apply RoPE if needed if freqs_cis_img is not None: qq, kk = apply_rotary_emb(q, k, freqs_cis_img, head_first=True) assert qq.shape == q.shape and kk.shape == k.shape, \ f'qq: {qq.shape}, q: {q.shape}, kk: {kk.shape}, k: {k.shape}' q, k = qq, kk q = q * self.scale attn = q @ k.transpose(-2, -1) # [b, h, s, d] @ [b, h, d, s] attn = attn.softmax(dim=-1) # [b, h, s, s] attn = self.attn_drop(attn) x = attn @ v # [b, h, s, d] x = x.transpose(1, 2).reshape(B, N, C) # [b, s, h, d] x = self.out_proj(x) x = self.proj_drop(x) out_tuple = (x,) return out_tuple