# Copyright (c) 2025 Resemble AI # Author: Manmay Nakhashi # MIT License import math import torch from torch import nn import torch.nn.functional as F from einops import rearrange class RelativePositionBias(nn.Module): def __init__(self, scale, causal=False, num_buckets=32, max_distance=128, heads=8): super().__init__() self.scale = scale self.causal = causal self.num_buckets = num_buckets self.max_distance = max_distance self.relative_attention_bias = nn.Embedding(num_buckets, heads) @staticmethod def _relative_position_bucket(relative_position, causal=True, num_buckets=32, max_distance=128): ret = 0 n = -relative_position if not causal: num_buckets //= 2 ret += (n < 0).long() * num_buckets n = torch.abs(n) else: n = torch.max(n, torch.zeros_like(n)) max_exact = num_buckets // 2 is_small = n < max_exact val_if_large = max_exact + ( torch.log(n.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact) ).long() val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1)) ret += torch.where(is_small, n, val_if_large) return ret def forward(self, qk_dots): i, j, device = *qk_dots.shape[-2:], qk_dots.device q_pos = torch.arange(i, dtype=torch.long, device=device) k_pos = torch.arange(j, dtype=torch.long, device=device) rel_pos = k_pos[None, :] - q_pos[:, None] rp_bucket = self._relative_position_bucket(rel_pos, causal=self.causal, num_buckets=self.num_buckets, max_distance=self.max_distance) values = self.relative_attention_bias(rp_bucket) bias = rearrange(values, 'i j h -> () h i j') return qk_dots + (bias * self.scale) class AttentionQKV(nn.Module): def __init__(self, n_heads, head_dim, dropout_rate=0.1, scale=None, flash=False): super().__init__() self.n_heads = n_heads self.head_dim = head_dim self.scale = scale if scale is not None else head_dim ** -0.5 self.flash = flash self.dropout_rate = dropout_rate self.dropout = nn.Dropout(dropout_rate) self.flash_config = self.setup_flash_config() if flash else None def setup_flash_config(self): # Setup flash attention configuration flash_config = { 'enable_flash': True, 'enable_math': True, 'enable_mem_efficient': True } return flash_config def forward(self, q, k, v, mask=None): q, k, v = [self.split_heads(tensor) for tensor in [q, k, v]] if self.flash: out = self.flash_attention(q, k, v, mask=mask) else: out = self.scaled_dot_product_attention(q, k, v, mask=mask) return self.combine_heads(out) def scaled_dot_product_attention(self, q, k, v, mask=None): sim = torch.einsum("bhlt,bhls->bhts", q, k) * self.scale if mask is not None: sim = sim.masked_fill(mask == 0, float('-inf')) attn = torch.softmax(sim, dim=-1) attn = self.dropout(attn) return torch.einsum("bhts,bhls->bhlt", attn, v) def flash_attention(self, q, k, v, mask=None): config = self.flash_config if self.flash_config else {} with torch.backends.cuda.sdp_kernel(**config): out = F.scaled_dot_product_attention( q, k, v, attn_mask=mask, dropout_p=self.dropout_rate if self.training else 0. ) return out def split_heads(self, x): bs, length, _ = x.shape x = x.view(bs, length, self.n_heads, self.head_dim) return x.permute(0, 2, 1, 3) def combine_heads(self, x): bs, _, length, _ = x.shape x = x.permute(0, 2, 1, 3).contiguous() return x.view(bs, length, -1) class AttentionBlock2(nn.Module): """ An attention block that allows spatial positions to attend to each other, using AttentionQKV and separate linear transformations for Q, K, and V. """ def __init__( self, channels, num_heads=1, num_head_channels=-1, relative_pos_embeddings=False, flash_attention=True, dropout_rate=0.2, scale=None ): super().__init__() self.channels = channels if num_head_channels == -1: self.num_heads = num_heads else: assert ( channels % num_head_channels == 0 ), f"channels {channels} is not divisible by num_head_channels {num_head_channels}" self.num_heads = channels // num_head_channels self.norm = nn.LayerNorm(channels) # Separate linear layers for Q, K, and V self.to_q = nn.Linear(channels, channels) self.to_k = nn.Linear(channels, channels) self.to_v = nn.Linear(channels, channels) self.attention = AttentionQKV(self.num_heads, channels // self.num_heads, dropout_rate=dropout_rate, flash=flash_attention, scale=scale) self.proj_out = nn.Linear(channels, channels) if relative_pos_embeddings: self.relative_pos_embeddings = RelativePositionBias(scale=(channels // self.num_heads) ** .5, causal=False, heads=num_heads, num_buckets=32, max_distance=64) else: self.relative_pos_embeddings = None def forward(self, x1, x2, mask=None): b1, c1, *spatial1 = x1.shape b2, c2, *spatial2 = x2.shape x1_norm = self.norm(x1) x2_norm = self.norm(x2) q = self.to_q(x1_norm) k = self.to_k(x2_norm) v = self.to_v(x2_norm) h = self.attention(q, k, v, mask=mask) h = self.proj_out(h) return (x1 + h).reshape(b1, c1, *spatial1) class Perceiver(nn.Module): """Inspired by https://arxiv.org/abs/2103.03206""" def __init__(self, pre_attention_query_token=32, pre_attention_query_size=1024, embedding_dim=1024, num_attn_heads=4): """ Initialize the perceiver module. :param pre_attention_query_token: Number of query tokens for pre-attention :param pre_attention_query_size: Size of each query token :param embedding_dim: Dimension of the embedding space :param num_attn_heads: Number of attention heads """ super().__init__() # Initialize the pre-attention query parameter self.pre_attention_query = torch.nn.Parameter( torch.empty(1, pre_attention_query_token, pre_attention_query_size) ) # Calculate the variance for uniform initialization query_variance = math.sqrt(3.0) * math.sqrt(2.0 / (pre_attention_query_token + pre_attention_query_token)) # Initialize the pre-attention query with uniform distribution self.pre_attention_query.data.uniform_(-query_variance, query_variance) # Initialize the attention block self.attn = AttentionBlock2(embedding_dim, num_attn_heads) def forward(self, h): """ Forward pass of the perceiver module. :param h: Input tensor :return: Output after applying attention mechanisms """ # Expand the pre-attention query to match the batch size of the input query_ = self.pre_attention_query.expand(h.shape[0], -1, -1) # Apply the first attention mechanism (cross-attention) pre_att = self.attn(query_, h) # Apply the second attention mechanism (self-attention) attn = self.attn(pre_att, pre_att) return attn