XVerse / src /flux /block.py
alexnasa's picture
Upload 10 files
f0dfe23 verified
raw
history blame
39.3 kB
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from typing import List, Union, Optional, Tuple, Dict, Any, Callable
from diffusers.models.attention_processor import Attention, F
from .lora_controller import enable_lora
from einops import rearrange
import math
from diffusers.models.embeddings import apply_rotary_emb
def scaled_dot_product_attention(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None) -> torch.Tensor:
# Efficient implementation equivalent to the following:
L, S = query.size(-2), key.size(-2)
B = query.size(0)
scale_factor = 1 / math.sqrt(query.size(-1)) if scale is None else scale
attn_bias = torch.zeros(B, 1, L, S, dtype=query.dtype, device=query.device)
if is_causal:
assert attn_mask is None
assert False
temp_mask = torch.ones(L, S, dtype=torch.bool).tril(diagonal=0)
attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
attn_bias.to(query.dtype)
if attn_mask is not None:
if attn_mask.dtype == torch.bool:
attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
else:
attn_bias += attn_mask
attn_weight = query @ key.transpose(-2, -1) * scale_factor
attn_weight += attn_bias.to(attn_weight.device)
attn_weight = torch.softmax(attn_weight, dim=-1)
return torch.dropout(attn_weight, dropout_p, train=True) @ value, attn_weight
def attn_forward(
attn: Attention,
hidden_states: torch.FloatTensor,
encoder_hidden_states: torch.FloatTensor = None,
condition_latents: torch.FloatTensor = None,
text_cond_mask: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
image_rotary_emb: Optional[torch.Tensor] = None,
cond_rotary_emb: Optional[torch.Tensor] = None,
model_config: Optional[Dict[str, Any]] = {},
store_attn_map: bool = False,
latent_height: Optional[int] = None,
timestep: Optional[torch.Tensor] = None,
last_attn_map: Optional[torch.Tensor] = None,
condition_sblora_weight: Optional[float] = None,
latent_sblora_weight: Optional[float] = None,
) -> torch.FloatTensor:
batch_size, _, _ = (
hidden_states.shape
if encoder_hidden_states is None
else encoder_hidden_states.shape
)
is_sblock = encoder_hidden_states is None
is_dblock = not is_sblock
with enable_lora(
(attn.to_q, attn.to_k, attn.to_v),
(is_dblock and model_config["latent_lora"]) or (is_sblock and model_config["sblock_lora"]), latent_sblora_weight=latent_sblora_weight
):
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:
query = attn.norm_q(query)
if attn.norm_k is not None:
key = attn.norm_k(key)
# the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
if encoder_hidden_states is not None:
# `context` projections.
with enable_lora((attn.add_q_proj, attn.add_k_proj, attn.add_v_proj), model_config["text_lora"]):
encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
if attn.norm_added_q is not None:
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
if attn.norm_added_k is not None:
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
# attention
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
if image_rotary_emb is not None:
query = apply_rotary_emb(query, image_rotary_emb)
key = apply_rotary_emb(key, image_rotary_emb)
if condition_latents is not None:
assert condition_latents.shape[0] == batch_size
cond_length = condition_latents.shape[1]
cond_lora_activate = (is_dblock and model_config["use_condition_dblock_lora"]) or (is_sblock and model_config["use_condition_sblock_lora"])
with enable_lora(
(attn.to_q, attn.to_k, attn.to_v),
dit_activated=not cond_lora_activate, cond_activated=cond_lora_activate, latent_sblora_weight=condition_sblora_weight #TODO implementation for condition lora not share
):
cond_query = attn.to_q(condition_latents)
cond_key = attn.to_k(condition_latents)
cond_value = attn.to_v(condition_latents)
cond_query = cond_query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
cond_key = cond_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
cond_value = cond_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
if attn.norm_q is not None:
cond_query = attn.norm_q(cond_query)
if attn.norm_k is not None:
cond_key = attn.norm_k(cond_key)
if cond_rotary_emb is not None:
cond_query = apply_rotary_emb(cond_query, cond_rotary_emb)
cond_key = apply_rotary_emb(cond_key, cond_rotary_emb)
if model_config.get("text_cond_attn", False):
if encoder_hidden_states is not None:
assert text_cond_mask is not None
img_length = hidden_states.shape[1]
seq_length = encoder_hidden_states_query_proj.shape[2]
assert len(text_cond_mask.shape) == 2 or len(text_cond_mask.shape) == 3
if len(text_cond_mask.shape) == 2:
text_cond_mask = text_cond_mask.unsqueeze(-1)
N = text_cond_mask.shape[-1] # num_condition
else:
raise NotImplementedError()
query = torch.cat([query, cond_query], dim=2) # (B, 24, S+HW+NC)
key = torch.cat([key, cond_key], dim=2)
value = torch.cat([value, cond_value], dim=2)
assert query.shape[2] == key.shape[2]
assert query.shape[2] == cond_length + img_length + seq_length
attention_mask = torch.ones(batch_size, 1, query.shape[2], key.shape[2], device=query.device, dtype=torch.bool)
attention_mask[..., -cond_length:, :-cond_length] = False
attention_mask[..., :-cond_length, -cond_length:] = False
if encoder_hidden_states is not None:
tokens_per_cond = cond_length // N
for i in range(batch_size):
for j in range(N):
start = seq_length + img_length + tokens_per_cond * j
attention_mask[i, 0, :seq_length, start:start+tokens_per_cond] = text_cond_mask[i, :, j].unsqueeze(-1)
elif model_config.get("union_cond_attn", False):
query = torch.cat([query, cond_query], dim=2) # (B, 24, S+HW+NC)
key = torch.cat([key, cond_key], dim=2)
value = torch.cat([value, cond_value], dim=2)
attention_mask = torch.ones(batch_size, 1, query.shape[2], key.shape[2], device=query.device, dtype=torch.bool)
cond_length = condition_latents.shape[1]
assert len(text_cond_mask.shape) == 2 or len(text_cond_mask.shape) == 3
if len(text_cond_mask.shape) == 2:
text_cond_mask = text_cond_mask.unsqueeze(-1)
N = text_cond_mask.shape[-1] # num_condition
tokens_per_cond = cond_length // N
seq_length = 0
if encoder_hidden_states is not None:
seq_length = encoder_hidden_states_query_proj.shape[2]
img_length = hidden_states.shape[1]
else:
seq_length = 128 # TODO, pass it here
img_length = hidden_states.shape[1] - seq_length
if not model_config.get("cond_cond_cross_attn", True):
# no cross attention between different conds
cond_start = seq_length + img_length
attention_mask[:, :, cond_start:, cond_start:] = False
for j in range(N):
start = cond_start + tokens_per_cond * j
end = cond_start + tokens_per_cond * (j + 1)
attention_mask[..., start:end, start:end] = True
# double block
if encoder_hidden_states is not None:
# no cross attention
attention_mask[..., :-cond_length, -cond_length:] = False
if model_config.get("use_attention_double", False) and last_attn_map is not None:
attention_mask = torch.zeros(batch_size, 1, query.shape[2], key.shape[2], device=query.device, dtype=torch.bfloat16)
last_attn_map = last_attn_map.to(query.device)
attention_mask[..., seq_length:-cond_length, :seq_length] = torch.log(last_attn_map/last_attn_map.mean()*model_config["use_atten_lambda"]).view(-1, seq_length)
# single block
else:
# print(last_attn_map)
if model_config.get("use_attention_single", False) and last_attn_map is not None:
attention_mask = torch.zeros(batch_size, 1, query.shape[2], key.shape[2], device=query.device, dtype=torch.bfloat16)
attention_mask[..., :seq_length, -cond_length:] = float('-inf')
# 确保 use_atten_lambda 是列表
use_atten_lambdas = model_config["use_atten_lambda"] if len(model_config["use_atten_lambda"])!=1 else model_config["use_atten_lambda"] * (N+1)
attention_mask[..., -cond_length:, seq_length:-cond_length] = math.log(use_atten_lambdas[0])
last_attn_map = last_attn_map.to(query.device)
cond2latents = []
for i in range(batch_size):
AM = last_attn_map[i] # (H, W, S)
for j in range(N):
start = seq_length + img_length + tokens_per_cond * j
mask = text_cond_mask[i, :, j] # (S,)
weighted_AM = AM * mask.unsqueeze(0).unsqueeze(0) # 扩展 mask 维度以匹配 AM
cond2latent = weighted_AM.mean(-1)
if model_config.get("attention_norm", "mean") == "max":
cond2latent = cond2latent / cond2latent.max() # 归一化
else:
cond2latent = cond2latent / cond2latent.mean() # 归一化
cond2latent = cond2latent.view(-1,) # (WH,)
# 使用对应 condition 的 lambda 值
current_lambda = use_atten_lambdas[j+1]
# 将 cond2latent 复制到 attention_mask[i, 0, :seq_length, start:start+tokens_per_cond]
attention_mask[i, 0, seq_length:-cond_length, start:start+tokens_per_cond] = torch.log(current_lambda * cond2latent.unsqueeze(-1))
# 将 text_cond_mask[i, :, j].unsqueeze(-1) 为 true 的位置设置为当前 lambda 值
cond = mask.unsqueeze(-1).expand(-1, tokens_per_cond)
sub_mask = attention_mask[i, 0, :seq_length, start:start+tokens_per_cond]
attention_mask[i, 0, :seq_length, start:start+tokens_per_cond] = torch.where(cond, math.log(current_lambda), sub_mask)
cond2latents.append(
cond2latent.reshape(latent_height, -1).detach().cpu()
)
if store_attn_map:
if not hasattr(attn, "cond2latents"):
attn.cond2latents = []
attn.cond_timesteps = []
attn.cond2latents.append(torch.stack(cond2latents, dim=0)) # (N, H, W)
attn.cond_timesteps.append(timestep.cpu())
pass
else:
raise NotImplementedError()
if hasattr(attn, "c_factor"):
assert False
attention_mask = torch.zeros(
query.shape[2], key.shape[2], device=query.device, dtype=query.dtype
)
bias = torch.log(attn.c_factor[0])
attention_mask[-cond_length:, :-cond_length] = bias
attention_mask[:-cond_length, -cond_length:] = bias
####################################################################################################
if store_attn_map and encoder_hidden_states is not None:
seq_length = encoder_hidden_states_query_proj.shape[2]
img_length = hidden_states.shape[1]
hidden_states, attention_probs = scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
# (B, 24, S+HW, S+HW) -> (B, 24, HW, S)
t2i_attention_probs = attention_probs[:, :, seq_length:seq_length+img_length, :seq_length]
# (B, 24, S+HW, S+HW) -> (B, 24, S, HW) -> (B, 24, HW, S)
i2t_attention_probs = attention_probs[:, :, :seq_length, seq_length:seq_length+img_length].transpose(-1, -2)
if not hasattr(attn, "attn_maps"):
attn.attn_maps = []
attn.timestep = []
attn.attn_maps.append(
(
rearrange(t2i_attention_probs, 'B attn_head (H W) attn_dim -> B attn_head H W attn_dim', H=latent_height),
rearrange(i2t_attention_probs, 'B attn_head (H W) attn_dim -> B attn_head H W attn_dim', H=latent_height),
)
)
attn.timestep.append(timestep.cpu())
has_nan = torch.isnan(hidden_states).any().item()
if has_nan:
print("[attn_forward] detect nan hidden_states in store_attn_map")
else:
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
has_nan = torch.isnan(hidden_states).any().item()
if has_nan:
print("[attn_forward] detect nan hidden_states")
####################################################################################################
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim).to(query.dtype)
if encoder_hidden_states is not None:
if condition_latents is not None:
encoder_hidden_states, hidden_states, condition_latents = (
hidden_states[:, : encoder_hidden_states.shape[1]],
hidden_states[
:, encoder_hidden_states.shape[1] : -condition_latents.shape[1]
],
hidden_states[:, -condition_latents.shape[1] :],
)
if model_config.get("latent_cond_by_text_attn", False):
# hidden_states += add_latent # (B, HW, D)
hidden_states = new_hidden_states # (B, HW, D)
else:
encoder_hidden_states, hidden_states = (
hidden_states[:, : encoder_hidden_states.shape[1]],
hidden_states[:, encoder_hidden_states.shape[1] :],
)
with enable_lora((attn.to_out[0],), model_config["latent_lora"]):
hidden_states = attn.to_out[0](hidden_states) # linear proj
hidden_states = attn.to_out[1](hidden_states) # dropout
with enable_lora((attn.to_add_out,), model_config["text_lora"]):
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
if condition_latents is not None:
cond_lora_activate = model_config["use_condition_dblock_lora"]
with enable_lora(
(attn.to_out[0],),
dit_activated=not cond_lora_activate, cond_activated=cond_lora_activate,
):
condition_latents = attn.to_out[0](condition_latents)
condition_latents = attn.to_out[1](condition_latents)
return (
(hidden_states, encoder_hidden_states, condition_latents)
if condition_latents is not None
else (hidden_states, encoder_hidden_states)
)
elif condition_latents is not None:
hidden_states, condition_latents = (
hidden_states[:, : -condition_latents.shape[1]],
hidden_states[:, -condition_latents.shape[1] :],
)
return hidden_states, condition_latents
else:
return hidden_states
def set_delta_by_start_end(
start_ends,
src_delta_emb, src_delta_emb_pblock,
delta_emb, delta_emb_pblock, delta_emb_mask,
):
for (i, j, src_s, src_e, tar_s, tar_e) in start_ends:
if src_delta_emb is not None:
delta_emb[i, tar_s:tar_e] = src_delta_emb[j, src_s:src_e]
if src_delta_emb_pblock is not None:
delta_emb_pblock[i, tar_s:tar_e] = src_delta_emb_pblock[j, src_s:src_e]
delta_emb_mask[i, tar_s:tar_e] = True
return delta_emb, delta_emb_pblock, delta_emb_mask
def norm1_context_forward(
self,
x: torch.Tensor,
condition_latents: Optional[torch.Tensor] = None,
timestep: Optional[torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
hidden_dtype: Optional[torch.dtype] = None,
emb: Optional[torch.Tensor] = None,
delta_emb: Optional[torch.Tensor] = None,
delta_emb_cblock: Optional[torch.Tensor] = None,
delta_emb_mask: Optional[torch.Tensor] = None,
delta_start_ends = None,
mod_adapter = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
batch_size, seq_length = x.shape[:2]
if mod_adapter is not None:
assert False
if delta_emb is None:
emb = self.linear(self.silu(emb)) # (B, 3072) -> (B, 18432)
emb = emb.unsqueeze(1) # (B, 1, 18432)
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=-1) # (B, 1, 3072)
x = self.norm(x) * (1 + scale_msa) + shift_msa # (B, 1, 3072)
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
else:
# (B, 3072) > (B, 18432) -> (B, S, 18432)
emb_orig = self.linear(self.silu(emb)).unsqueeze(1).expand((-1, seq_length, -1))
# (B, 3072) -> (B, 1, 3072) -> (B, S, 3072) -> (B, S, 18432)
if delta_emb_cblock is None:
emb_new = self.linear(self.silu(emb.unsqueeze(1) + delta_emb))
else:
emb_new = self.linear(self.silu(emb.unsqueeze(1) + delta_emb + delta_emb_cblock))
emb = torch.where(delta_emb_mask.unsqueeze(-1), emb_new, emb_orig) # (B, S, 18432)
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=-1) # (B, S, 3072)
x = self.norm(x) * (1 + scale_msa) + shift_msa # (B, S, 3072)
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
def norm1_forward(
self,
x: torch.Tensor,
timestep: Optional[torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
hidden_dtype: Optional[torch.dtype] = None,
emb: Optional[torch.Tensor] = None,
delta_emb: Optional[torch.Tensor] = None,
delta_emb_cblock: Optional[torch.Tensor] = None,
delta_emb_mask: Optional[torch.Tensor] = None,
t2i_attn_map: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
if delta_emb is None:
emb = self.linear(self.silu(emb)) # (B, 3072) -> (B, 18432)
emb = emb.unsqueeze(1) # (B, 1, 18432)
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=-1) # (B, 1, 3072)
x = self.norm(x) * (1 + scale_msa) + shift_msa # (B, 1, 3072)
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
else:
raise NotImplementedError()
batch_size, HW = x.shape[:2]
seq_length = t2i_attn_map.shape[-1]
# (B, 3072) > (B, 18432) -> (B, S, 18432)
emb_orig = self.linear(self.silu(emb)).unsqueeze(1).expand((-1, seq_length, -1))
# (B, 3072) -> (B, 1, 3072) -> (B, S, 3072) -> (B, S, 18432)
if delta_emb_cblock is None:
emb_new = self.linear(self.silu(emb.unsqueeze(1) + delta_emb))
else:
emb_new = self.linear(self.silu(emb.unsqueeze(1) + delta_emb + delta_emb_cblock))
# attn_weight (B, HW, S)
emb = torch.where(delta_emb_mask.unsqueeze(-1), emb_new, emb_orig) # (B, S, 18432)
emb = t2i_attn_map @ emb # (B, HW, 18432)
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=-1) # (B, HW, 3072)
x = self.norm(x) * (1 + scale_msa) + shift_msa # (B, HW, 3072)
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
def block_forward(
self,
hidden_states: torch.FloatTensor,
encoder_hidden_states: torch.FloatTensor,
condition_latents: torch.FloatTensor,
temb: torch.FloatTensor,
cond_temb: torch.FloatTensor,
text_cond_mask: Optional[torch.FloatTensor] = None,
delta_emb: Optional[torch.FloatTensor] = None,
delta_emb_cblock: Optional[torch.FloatTensor] = None,
delta_emb_mask: Optional[torch.Tensor] = None,
delta_start_ends = None,
cond_rotary_emb=None,
image_rotary_emb=None,
model_config: Optional[Dict[str, Any]] = {},
store_attn_map: bool = False,
use_text_mod: bool = True,
use_img_mod: bool = False,
mod_adapter = None,
latent_height: Optional[int] = None,
timestep: Optional[torch.Tensor] = None,
last_attn_map: Optional[torch.Tensor] = None,
):
batch_size = hidden_states.shape[0]
use_cond = condition_latents is not None
train_partial_latent_lora = model_config.get("train_partial_latent_lora", False)
train_partial_text_lora = model_config.get("train_partial_text_lora", False)
if train_partial_latent_lora:
train_partial_latent_lora_layers = model_config.get("train_partial_latent_lora_layers", "")
activate_norm1 = activate_ff = True
if "norm1" not in train_partial_latent_lora_layers:
activate_norm1 = False
if "ff" not in train_partial_latent_lora_layers:
activate_ff = False
if train_partial_text_lora:
train_partial_text_lora_layers = model_config.get("train_partial_text_lora_layers", "")
activate_norm1_context = activate_ff_context = True
if "norm1" not in train_partial_text_lora_layers:
activate_norm1_context = False
if "ff" not in train_partial_text_lora_layers:
activate_ff_context = False
if use_cond:
cond_lora_activate = model_config["use_condition_dblock_lora"]
with enable_lora(
(self.norm1.linear,),
dit_activated=activate_norm1 if train_partial_latent_lora else not cond_lora_activate, cond_activated=cond_lora_activate,
):
norm_condition_latents, cond_gate_msa, cond_shift_mlp, cond_scale_mlp, cond_gate_mlp = (
norm1_forward(
self.norm1,
condition_latents,
emb=cond_temb,
)
)
delta_emb_img = delta_emb_img_cblock = None
if use_img_mod and use_text_mod:
if delta_emb is not None:
delta_emb_img, delta_emb = delta_emb.chunk(2, dim=-1)
if delta_emb_cblock is not None:
delta_emb_img_cblock, delta_emb_cblock = delta_emb_cblock.chunk(2, dim=-1)
with enable_lora((self.norm1.linear,), activate_norm1 if train_partial_latent_lora else model_config["latent_lora"]):
if use_img_mod and encoder_hidden_states is not None:
with torch.no_grad():
attn = self.attn
norm_img = self.norm1(hidden_states, emb=temb)[0]
norm_text = self.norm1_context(encoder_hidden_states, emb=temb)[0]
img_query = attn.to_q(norm_img)
img_key = attn.to_k(norm_img)
text_query = attn.add_q_proj(norm_text)
text_key = attn.add_k_proj(norm_text)
inner_dim = img_key.shape[-1]
head_dim = inner_dim // attn.heads
img_query = img_query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # (B, N, HW, D)
img_key = img_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # (B, N, HW, D)
text_query = text_query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # (B, N, S, D)
text_key = text_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) # (B, N, S, D)
if attn.norm_q is not None:
img_query = attn.norm_q(img_query)
if attn.norm_added_q is not None:
text_query = attn.norm_added_q(text_query)
if attn.norm_k is not None:
img_key = attn.norm_k(img_key)
if attn.norm_added_k is not None:
text_key = attn.norm_added_k(text_key)
query = torch.cat([text_query, img_query], dim=2) # (B, N, S+HW, D)
key = torch.cat([text_key, img_key], dim=2) # (B, N, S+HW, D)
if image_rotary_emb is not None:
query = apply_rotary_emb(query, image_rotary_emb)
key = apply_rotary_emb(key, image_rotary_emb)
seq_length = text_query.shape[2]
scale_factor = 1 / math.sqrt(query.size(-1))
t2i_attn_map = query @ key.transpose(-2, -1) * scale_factor # (B, N, S+HW, S+HW)
t2i_attn_map = t2i_attn_map.mean(1)[:, seq_length:, :seq_length] # (B, S+HW, S+HW) -> (B, HW, S)
t2i_attn_map = torch.softmax(t2i_attn_map, dim=-1) # (B, HW, S)
else:
t2i_attn_map = None
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
norm1_forward(
self.norm1,
hidden_states,
emb=temb,
delta_emb=delta_emb_img,
delta_emb_cblock=delta_emb_img_cblock,
delta_emb_mask=delta_emb_mask,
t2i_attn_map=t2i_attn_map,
)
)
# Modulation for double block
with enable_lora((self.norm1_context.linear,), activate_norm1_context if train_partial_text_lora else model_config["text_lora"]):
norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = (
norm1_context_forward(
self.norm1_context,
encoder_hidden_states,
emb=temb,
delta_emb=delta_emb if use_text_mod else None,
delta_emb_cblock=delta_emb_cblock if use_text_mod else None,
delta_emb_mask=delta_emb_mask if use_text_mod else None,
delta_start_ends=delta_start_ends if use_text_mod else None,
mod_adapter=mod_adapter,
condition_latents=condition_latents,
)
)
# Attention.
result = attn_forward(
self.attn,
model_config=model_config,
hidden_states=norm_hidden_states,
encoder_hidden_states=norm_encoder_hidden_states,
condition_latents=norm_condition_latents if use_cond else None,
text_cond_mask=text_cond_mask if use_cond else None,
image_rotary_emb=image_rotary_emb,
cond_rotary_emb=cond_rotary_emb if use_cond else None,
store_attn_map=store_attn_map,
latent_height=latent_height,
timestep=timestep,
last_attn_map=last_attn_map,
)
attn_output, context_attn_output = result[:2]
cond_attn_output = result[2] if use_cond else None
# Process attention outputs for the `hidden_states`.
# 1. hidden_states
attn_output = gate_msa * attn_output # NOTE: changed by img mod
hidden_states = hidden_states + attn_output
# 2. encoder_hidden_states
context_attn_output = c_gate_msa * context_attn_output # NOTE: changed by delta_temb
encoder_hidden_states = encoder_hidden_states + context_attn_output
# 3. condition_latents
if use_cond:
cond_attn_output = cond_gate_msa * cond_attn_output # NOTE: changed by img mod
condition_latents = condition_latents + cond_attn_output
if model_config.get("add_cond_attn", False):
hidden_states += cond_attn_output
# LayerNorm + MLP.
# 1. hidden_states
norm_hidden_states = self.norm2(hidden_states)
norm_hidden_states = (
norm_hidden_states * (1 + scale_mlp) + shift_mlp # NOTE: changed by img mod
)
# 2. encoder_hidden_states
norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
norm_encoder_hidden_states = (
norm_encoder_hidden_states * (1 + c_scale_mlp) + c_shift_mlp # NOTE: changed by delta_temb
)
# 3. condition_latents
if use_cond:
norm_condition_latents = self.norm2(condition_latents)
norm_condition_latents = (
norm_condition_latents * (1 + cond_scale_mlp) + cond_shift_mlp # NOTE: changed by img mod
)
# Feed-forward.
with enable_lora((self.ff.net[2],), activate_ff if train_partial_latent_lora else model_config["latent_lora"]):
# 1. hidden_states
ff_output = self.ff(norm_hidden_states)
ff_output = gate_mlp * ff_output # NOTE: changed by img mod
# 2. encoder_hidden_states
with enable_lora((self.ff_context.net[2],), activate_ff_context if train_partial_text_lora else model_config["text_lora"]):
context_ff_output = self.ff_context(norm_encoder_hidden_states)
context_ff_output = c_gate_mlp * context_ff_output # NOTE: changed by delta_temb
# 3. condition_latents
if use_cond:
cond_lora_activate = model_config["use_condition_dblock_lora"]
with enable_lora(
(self.ff.net[2],),
dit_activated=activate_ff if train_partial_latent_lora else not cond_lora_activate, cond_activated=cond_lora_activate,
):
cond_ff_output = self.ff(norm_condition_latents)
cond_ff_output = cond_gate_mlp * cond_ff_output # NOTE: changed by img mod
# Process feed-forward outputs.
hidden_states = hidden_states + ff_output
encoder_hidden_states = encoder_hidden_states + context_ff_output
if use_cond:
condition_latents = condition_latents + cond_ff_output
# Clip to avoid overflow.
if encoder_hidden_states.dtype == torch.float16:
encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
return encoder_hidden_states, hidden_states, condition_latents if use_cond else None
def single_norm_forward(
self,
x: torch.Tensor,
timestep: Optional[torch.Tensor] = None,
class_labels: Optional[torch.LongTensor] = None,
hidden_dtype: Optional[torch.dtype] = None,
emb: Optional[torch.Tensor] = None,
delta_emb: Optional[torch.Tensor] = None,
delta_emb_cblock: Optional[torch.Tensor] = None,
delta_emb_mask: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
if delta_emb is None:
emb = self.linear(self.silu(emb)) # (B, 3072) -> (B, 9216)
emb = emb.unsqueeze(1) # (B, 1, 9216)
shift_msa, scale_msa, gate_msa = emb.chunk(3, dim=-1) # (B, 1, 3072)
x = self.norm(x) * (1 + scale_msa) + shift_msa # (B, S, 3072) * (B, 1, 3072)
return x, gate_msa
else:
img_text_seq_length = x.shape[1] # S+
text_seq_length = delta_emb_mask.shape[1] # S
# (B, 3072) -> (B, 9216) -> (B, S+, 9216)
emb_orig = self.linear(self.silu(emb)).unsqueeze(1).expand((-1, img_text_seq_length, -1))
# (B, 3072) -> (B, 1, 3072) -> (B, S, 3072) -> (B, S, 9216)
if delta_emb_cblock is None:
emb_new = self.linear(self.silu(emb.unsqueeze(1) + delta_emb))
else:
emb_new = self.linear(self.silu(emb.unsqueeze(1) + delta_emb + delta_emb_cblock))
emb_text = torch.where(delta_emb_mask.unsqueeze(-1), emb_new, emb_orig[:, :text_seq_length]) # (B, S, 9216)
emb_img = emb_orig[:, text_seq_length:] # (B, s, 9216)
emb = torch.cat([emb_text, emb_img], dim=1) # (B, S+, 9216)
shift_msa, scale_msa, gate_msa = emb.chunk(3, dim=-1) # (B, S+, 3072)
x = self.norm(x) * (1 + scale_msa) + shift_msa # (B, S+, 3072)
return x, gate_msa
def single_block_forward(
self,
hidden_states: torch.FloatTensor,
temb: torch.FloatTensor,
image_rotary_emb=None,
condition_latents: torch.FloatTensor = None,
text_cond_mask: torch.FloatTensor = None,
cond_temb: torch.FloatTensor = None,
delta_emb: Optional[torch.FloatTensor] = None,
delta_emb_cblock: Optional[torch.FloatTensor] = None,
delta_emb_mask: Optional[torch.Tensor] = None,
use_text_mod: bool = True,
use_img_mod: bool = False,
cond_rotary_emb=None,
latent_height: Optional[int] = None,
timestep: Optional[torch.Tensor] = None,
store_attn_map: bool = False,
model_config: Optional[Dict[str, Any]] = {},
last_attn_map: Optional[torch.Tensor] = None,
latent_sblora_weight=None,
condition_sblora_weight=None,
):
using_cond = condition_latents is not None
residual = hidden_states
train_partial_lora = model_config.get("train_partial_lora", False)
if train_partial_lora:
train_partial_lora_layers = model_config.get("train_partial_lora_layers", "")
activate_norm = activate_projmlp = activate_projout = True
if "norm" not in train_partial_lora_layers:
activate_norm = False
if "projmlp" not in train_partial_lora_layers:
activate_projmlp = False
if "projout" not in train_partial_lora_layers:
activate_projout = False
with enable_lora((self.norm.linear,), activate_norm if train_partial_lora else model_config["sblock_lora"], latent_sblora_weight=latent_sblora_weight):
# Modulation for single block
norm_hidden_states, gate = single_norm_forward(
self.norm,
hidden_states,
emb=temb,
delta_emb=delta_emb if use_text_mod else None,
delta_emb_cblock=delta_emb_cblock if use_text_mod else None,
delta_emb_mask=delta_emb_mask if use_text_mod else None,
)
with enable_lora((self.proj_mlp,), activate_projmlp if train_partial_lora else model_config["sblock_lora"], latent_sblora_weight=latent_sblora_weight):
mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
if using_cond:
cond_lora_activate = model_config["use_condition_sblock_lora"]
with enable_lora(
(self.norm.linear,),
dit_activated=activate_norm if train_partial_lora else not cond_lora_activate, cond_activated=cond_lora_activate, latent_sblora_weight=condition_sblora_weight
):
residual_cond = condition_latents
norm_condition_latents, cond_gate = self.norm(condition_latents, emb=cond_temb)
with enable_lora(
(self.proj_mlp,),
dit_activated=activate_projmlp if train_partial_lora else not cond_lora_activate, cond_activated=cond_lora_activate, latent_sblora_weight=condition_sblora_weight
):
mlp_cond_hidden_states = self.act_mlp(self.proj_mlp(norm_condition_latents))
attn_output = attn_forward(
self.attn,
model_config=model_config,
hidden_states=norm_hidden_states,
image_rotary_emb=image_rotary_emb,
last_attn_map=last_attn_map,
latent_height=latent_height,
store_attn_map=store_attn_map,
timestep=timestep,
latent_sblora_weight=latent_sblora_weight,
condition_sblora_weight=condition_sblora_weight,
**(
{
"condition_latents": norm_condition_latents,
"cond_rotary_emb": cond_rotary_emb if using_cond else None,
"text_cond_mask": text_cond_mask if using_cond else None,
}
if using_cond
else {}
),
)
if using_cond:
attn_output, cond_attn_output = attn_output
with enable_lora((self.proj_out,), activate_projout if train_partial_lora else model_config["sblock_lora"], latent_sblora_weight=latent_sblora_weight):
hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
# gate = (B, 1, 3072) or (B, S+, 3072)
hidden_states = gate * self.proj_out(hidden_states)
hidden_states = residual + hidden_states
if using_cond:
cond_lora_activate = model_config["use_condition_sblock_lora"]
with enable_lora(
(self.proj_out,),
dit_activated=activate_projout if train_partial_lora else not cond_lora_activate, cond_activated=cond_lora_activate, latent_sblora_weight=condition_sblora_weight
):
condition_latents = torch.cat([cond_attn_output, mlp_cond_hidden_states], dim=2)
cond_gate = cond_gate.unsqueeze(1)
condition_latents = cond_gate * self.proj_out(condition_latents)
condition_latents = residual_cond + condition_latents
if hidden_states.dtype == torch.float16:
hidden_states = hidden_states.clip(-65504, 65504)
return hidden_states if not using_cond else (hidden_states, condition_latents)