Spaces:
Running
on
Zero
Running
on
Zero
| import torch | |
| import torch.nn as nn | |
| import math | |
| import torch.nn.functional as F | |
| from typing import Callable, List, Optional, Tuple, Union | |
| class FluxBlendedAttnProcessor2_0(nn.Module): | |
| """Attention processor used typically in processing the SD3-like self-attention projections.""" | |
| def __init__(self, hidden_dim, ba_scale=1.0, num_ref=1): | |
| super().__init__() | |
| if not hasattr(F, "scaled_dot_product_attention"): | |
| raise ImportError("FluxBlendedAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") | |
| self.blended_attention_k_proj = nn.Linear(hidden_dim, hidden_dim, bias=False) | |
| self.blended_attention_v_proj = nn.Linear(hidden_dim, hidden_dim, bias=False) | |
| self.ba_scale = ba_scale | |
| self.num_ref = num_ref | |
| def __call__( | |
| self, | |
| attn, #: Attention, | |
| hidden_states: torch.FloatTensor, | |
| encoder_hidden_states: torch.FloatTensor = None, | |
| attention_mask: Optional[torch.FloatTensor] = None, | |
| image_rotary_emb: Optional[torch.Tensor] = None, | |
| is_negative_prompt: bool = False | |
| ) -> torch.FloatTensor: | |
| assert encoder_hidden_states is None, "It should be given as None because we are applying it-blender only to the single streams." | |
| batch_size, _, _ = hidden_states.shape | |
| # `sample` projections. | |
| query = attn.to_q(hidden_states) | |
| key = attn.to_k(hidden_states) | |
| value = attn.to_v(hidden_states) | |
| inner_dim = key.shape[-1] | |
| head_dim = inner_dim // attn.heads | |
| query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) | |
| key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) | |
| value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) | |
| if attn.norm_q is not None: | |
| normalized_query = attn.norm_q(query) | |
| if attn.norm_k is not None: | |
| key = attn.norm_k(key) | |
| if image_rotary_emb is not None: | |
| from diffusers.models.embeddings import apply_rotary_emb | |
| query = apply_rotary_emb(normalized_query, image_rotary_emb) | |
| key = apply_rotary_emb(key, image_rotary_emb) | |
| hidden_states = F.scaled_dot_product_attention( | |
| query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False | |
| ) | |
| hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) | |
| hidden_states = hidden_states.to(query.dtype) | |
| # [noisy, clean] | |
| chunk = batch_size//(1+self.num_ref) | |
| ba_query = normalized_query[:chunk] # noisy query | |
| ba_key = self.blended_attention_k_proj(hidden_states[chunk:]) # clean key | |
| ba_value = self.blended_attention_v_proj(hidden_states[chunk:]) # clean value | |
| ba_key = ba_key.view(chunk, -1, attn.heads, head_dim).transpose(1, 2) # the -1 is gonna be multiplied by self.num_ref | |
| ba_value = ba_value.view(chunk, -1, attn.heads, head_dim).transpose(1, 2) | |
| ba_hidden_states = F.scaled_dot_product_attention( | |
| ba_query, ba_key, ba_value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False#, scale=(1 / math.sqrt(ba_query.size(-1)))*self.temperature if self.num_ref > 1 else 1 / math.sqrt(ba_query.size(-1)) | |
| ) | |
| ba_hidden_states = ba_hidden_states.transpose(1, 2).reshape(chunk, -1, attn.heads * head_dim) | |
| ba_hidden_states = ba_hidden_states.to(query.dtype) | |
| zero_tensor_list = [torch.zeros_like(ba_hidden_states)]*self.num_ref | |
| ba_hidden_states = torch.cat([ba_hidden_states]+zero_tensor_list, dim=0) | |
| hidden_states = hidden_states + self.ba_scale * ba_hidden_states | |
| return hidden_states | |