396 lines
14 KiB
Python
396 lines
14 KiB
Python
# Copyright (c) 2025 Zhipu AI Inc (authors: CogAudio Group Members)
|
|
#
|
|
# 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.
|
|
"""
|
|
Core modules for the DiT (Diffusion Transformer) model.
|
|
Includes position embeddings, normalization, attention mechanisms, and transformer blocks.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
from typing import Optional
|
|
import math
|
|
import torch
|
|
from torch import nn
|
|
import torch.nn.functional as F
|
|
from einops import rearrange
|
|
from x_transformers.x_transformers import apply_rotary_pos_emb
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Position Embeddings
|
|
# -----------------------------------------------------------------------------
|
|
|
|
class SinusPositionEmbedding(nn.Module):
|
|
def __init__(self, dim):
|
|
super().__init__()
|
|
self.dim = dim
|
|
|
|
def forward(self, x, scale=1000):
|
|
device = x.device
|
|
half_dim = self.dim // 2
|
|
emb = math.log(10000) / (half_dim - 1)
|
|
emb = torch.exp(torch.arange(half_dim, device=device).float() * -emb)
|
|
emb = scale * x.unsqueeze(1) * emb.unsqueeze(0)
|
|
emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
|
|
return emb
|
|
|
|
|
|
class ConvPositionEmbedding(nn.Module):
|
|
def __init__(self, dim, kernel_size=31, groups=16):
|
|
super().__init__()
|
|
assert kernel_size % 2 != 0
|
|
self.conv1d = nn.Sequential(
|
|
nn.Conv1d(dim, dim, kernel_size, groups=groups, padding=kernel_size // 2),
|
|
nn.Mish(),
|
|
nn.Conv1d(dim, dim, kernel_size, groups=groups, padding=kernel_size // 2),
|
|
nn.Mish(),
|
|
)
|
|
|
|
def forward(self, x: torch.Tensor, mask: Optional[torch.Tensor] = None):
|
|
"""
|
|
Args:
|
|
x: Input tensor of shape (batch, seq_len, dim)
|
|
mask: Boolean mask of shape (batch, seq_len)
|
|
"""
|
|
if mask is not None:
|
|
mask = mask[..., None]
|
|
x = x.masked_fill(~mask, 0.)
|
|
|
|
x = rearrange(x, 'b n d -> b d n')
|
|
x = self.conv1d(x)
|
|
out = rearrange(x, 'b d n -> b n d')
|
|
|
|
if mask is not None:
|
|
out = out.masked_fill(~mask, 0.)
|
|
|
|
return out
|
|
|
|
|
|
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0, theta_rescale_factor=1.):
|
|
"""
|
|
Precompute Rotary Positional Embeddings (RoPE).
|
|
References:
|
|
https://github.com/lucidrains/rotary-embedding-torch
|
|
"""
|
|
theta *= theta_rescale_factor ** (dim / (dim - 2))
|
|
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
|
t = torch.arange(end, device=freqs.device)
|
|
freqs = torch.outer(t, freqs).float()
|
|
freqs_cos = torch.cos(freqs)
|
|
freqs_sin = torch.sin(freqs)
|
|
return torch.cat([freqs_cos, freqs_sin], dim=-1)
|
|
|
|
|
|
def get_pos_embed_indices(start, length, max_pos, scale=1.):
|
|
"""
|
|
Calculates indices for positional embeddings, handling potential scalar scaling.
|
|
"""
|
|
scale = scale * torch.ones_like(start, dtype=torch.float32)
|
|
pos = start.unsqueeze(1) + (
|
|
torch.arange(length, device=start.device, dtype=torch.float32).unsqueeze(0) *
|
|
scale.unsqueeze(1)).long()
|
|
# Clamp to avoid index out of bounds
|
|
pos = torch.where(pos < max_pos, pos, max_pos - 1)
|
|
return pos
|
|
|
|
|
|
class TimestepEmbedding(nn.Module):
|
|
def __init__(self, dim, freq_embed_dim=256):
|
|
super().__init__()
|
|
self.time_embed = SinusPositionEmbedding(freq_embed_dim)
|
|
self.time_mlp = nn.Sequential(
|
|
nn.Linear(freq_embed_dim, dim),
|
|
nn.SiLU(),
|
|
nn.Linear(dim, dim)
|
|
)
|
|
|
|
def forward(self, timestep: torch.Tensor):
|
|
"""
|
|
Args:
|
|
timestep: Tensor of shape (batch,)
|
|
Returns:
|
|
Time embedding of shape (batch, dim)
|
|
"""
|
|
time_hidden = self.time_embed(timestep)
|
|
time = self.time_mlp(time_hidden)
|
|
return time
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Normalization & Layers
|
|
# -----------------------------------------------------------------------------
|
|
|
|
class GRN(nn.Module):
|
|
""" Global Response Normalization (GRN) layer """
|
|
def __init__(self, dim):
|
|
super().__init__()
|
|
self.gamma = nn.Parameter(torch.zeros(1, 1, dim))
|
|
self.beta = nn.Parameter(torch.zeros(1, 1, dim))
|
|
|
|
def forward(self, x):
|
|
Gx = torch.norm(x, p=2, dim=1, keepdim=True)
|
|
Nx = Gx / (Gx.mean(dim=-1, keepdim=True) + 1e-6)
|
|
return self.gamma * (x * Nx) + self.beta + x
|
|
|
|
|
|
class ConvNeXtV2Block(nn.Module):
|
|
"""
|
|
ConvNeXt-V2 Block.
|
|
Reference: https://github.com/facebookresearch/ConvNeXt-V2
|
|
"""
|
|
def __init__(self, dim: int, intermediate_dim: int, dilation: int = 1):
|
|
super().__init__()
|
|
padding = (dilation * (7 - 1)) // 2
|
|
self.dwconv = nn.Conv1d(dim, dim, kernel_size=7, padding=padding, groups=dim, dilation=dilation)
|
|
self.norm = nn.LayerNorm(dim, eps=1e-6)
|
|
self.pwconv1 = nn.Linear(dim, intermediate_dim)
|
|
self.act = nn.GELU()
|
|
self.grn = GRN(intermediate_dim)
|
|
self.pwconv2 = nn.Linear(intermediate_dim, dim)
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
residual = x
|
|
x = x.transpose(1, 2) # (b, n, d) -> (b, d, n)
|
|
x = self.dwconv(x)
|
|
x = x.transpose(1, 2) # (b, d, n) -> (b, n, d)
|
|
x = self.norm(x)
|
|
x = self.pwconv1(x)
|
|
x = self.act(x)
|
|
x = self.grn(x)
|
|
x = self.pwconv2(x)
|
|
return residual + x
|
|
|
|
|
|
class AdaLayerNormZero(nn.Module):
|
|
"""
|
|
Adaptive LayerNorm with modulation (scale/shift) from an embedding (e.g., time/condition).
|
|
Outputs modulation parameters for MSA (Multi-Head Self Attention) and MLP (FeedForward).
|
|
"""
|
|
def __init__(self, dim, additional_dim=0):
|
|
super().__init__()
|
|
self.silu = nn.SiLU()
|
|
self.linear = nn.Linear(dim + additional_dim, dim * 6)
|
|
self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
|
|
|
|
def forward(self, x, emb=None):
|
|
emb = self.linear(self.silu(emb))
|
|
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = torch.chunk(emb, 6, dim=1)
|
|
x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None]
|
|
return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
|
|
|
|
class AdaLayerNormZero_Final(nn.Module):
|
|
"""
|
|
Adaptive LayerNorm for the final layer.
|
|
Only outputs modulation for the current normalization, no downstream parameters.
|
|
"""
|
|
def __init__(self, dim, additional_dim=0):
|
|
super().__init__()
|
|
self.silu = nn.SiLU()
|
|
self.linear = nn.Linear(dim + additional_dim, dim * 2)
|
|
self.norm = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
|
|
|
|
def forward(self, x, emb):
|
|
emb = self.linear(self.silu(emb))
|
|
scale, shift = torch.chunk(emb, 2, dim=1)
|
|
x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
|
|
return x
|
|
|
|
class FeedForward(nn.Module):
|
|
def __init__(self, dim, dim_out=None, mult=4, dropout=0., approximate: str = 'none'):
|
|
super().__init__()
|
|
inner_dim = int(dim * mult)
|
|
dim_out = dim_out if dim_out is not None else dim
|
|
|
|
activation = nn.GELU(approximate=approximate)
|
|
|
|
project_in = nn.Sequential(
|
|
nn.Linear(dim, inner_dim),
|
|
activation
|
|
)
|
|
|
|
self.ff = nn.Sequential(
|
|
project_in,
|
|
nn.Dropout(dropout),
|
|
nn.Linear(inner_dim, dim_out)
|
|
)
|
|
|
|
def forward(self, x):
|
|
return self.ff(x)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Attention Mechanisms
|
|
# -----------------------------------------------------------------------------
|
|
|
|
class AttnProcessorCausalV2:
|
|
def __init__(self):
|
|
pass
|
|
|
|
def __call__(
|
|
self,
|
|
attn: AttentionV2,
|
|
x: torch.Tensor,
|
|
padding_mask: Optional[torch.Tensor] = None,
|
|
rope=None,
|
|
attn_mask=None,
|
|
) -> torch.Tensor:
|
|
|
|
batch_size = x.shape[0]
|
|
|
|
# Projections
|
|
query = attn.to_q(x)
|
|
key = attn.to_k(x)
|
|
value = attn.to_v(x)
|
|
|
|
# Apply Rotary Position Embedding (RoPE)
|
|
if rope is not None:
|
|
freqs, xpos_scale = rope
|
|
q_xpos_scale, k_xpos_scale = (xpos_scale, xpos_scale ** -1.) if xpos_scale is not None else (1., 1.)
|
|
query = apply_rotary_pos_emb(query, freqs, q_xpos_scale)
|
|
key = apply_rotary_pos_emb(key, freqs, k_xpos_scale)
|
|
|
|
# Reshape for multi-head attention
|
|
inner_dim = key.shape[-1]
|
|
head_dim = inner_dim // attn.heads
|
|
|
|
# (batch, seq, heads, dim) -> (batch, heads, seq, dim)
|
|
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)
|
|
|
|
# Scaled Dot Product Attention
|
|
# is_causal=False because masking is handled explicitly via attn_mask
|
|
x = F.scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=0.0, is_causal=False)
|
|
|
|
# Reshape back: (batch, heads, seq, dim) -> (batch, seq, inner_dim)
|
|
x = x.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
|
x = x.to(query.dtype)
|
|
|
|
# Output projection and dropout
|
|
x = attn.to_out[0](x)
|
|
x = attn.to_out[1](x)
|
|
|
|
# Apply padding mask if provided
|
|
if padding_mask is not None:
|
|
padding_mask = rearrange(padding_mask, 'b 1 n -> b n 1').bool()
|
|
x = x.masked_fill(~padding_mask, 0.)
|
|
|
|
return x
|
|
|
|
|
|
class AttentionV2(nn.Module):
|
|
def __init__(
|
|
self,
|
|
processor: AttnProcessorCausalV2,
|
|
dim: int,
|
|
heads: int = 8,
|
|
dim_head: int = 64,
|
|
dropout: float = 0.0,
|
|
context_dim: Optional[int] = None,
|
|
context_pre_only=None,
|
|
):
|
|
super().__init__()
|
|
|
|
if not hasattr(F, "scaled_dot_product_attention"):
|
|
raise ImportError("Attention requires PyTorch 2.0+.")
|
|
|
|
self.processor = processor
|
|
self.dim = dim
|
|
self.heads = heads
|
|
self.inner_dim = dim_head * heads
|
|
self.dropout = dropout
|
|
self.context_dim = context_dim
|
|
self.context_pre_only = context_pre_only
|
|
|
|
self.to_q = nn.Linear(dim, self.inner_dim)
|
|
self.to_k = nn.Linear(dim, self.inner_dim)
|
|
self.to_v = nn.Linear(dim, self.inner_dim)
|
|
|
|
# Context Attention (if applicable)
|
|
if self.context_dim is not None:
|
|
self.to_k_c = nn.Linear(context_dim, self.inner_dim)
|
|
self.to_v_c = nn.Linear(context_dim, self.inner_dim)
|
|
if self.context_pre_only is not None:
|
|
self.to_q_c = nn.Linear(context_dim, self.inner_dim)
|
|
|
|
self.to_out = nn.ModuleList([
|
|
nn.Linear(self.inner_dim, dim),
|
|
nn.Dropout(dropout)
|
|
])
|
|
|
|
if self.context_pre_only is not None and not self.context_pre_only:
|
|
self.to_out_c = nn.Linear(self.inner_dim, dim)
|
|
|
|
def forward(
|
|
self,
|
|
x: torch.Tensor,
|
|
c: torch.Tensor = None,
|
|
padding_mask: Optional[torch.Tensor] = None,
|
|
rope=None,
|
|
c_rope=None,
|
|
attn_mask=None,
|
|
) -> torch.Tensor:
|
|
if c is not None:
|
|
# Note: The current flow logic mainly uses self-attention without 'c' here,
|
|
# but the interface supports it.
|
|
return self.processor(self, x, c=c, padding_mask=padding_mask, rope=rope, c_rope=c_rope,
|
|
attn_mask=attn_mask)
|
|
else:
|
|
return self.processor(self, x, padding_mask=padding_mask, rope=rope, attn_mask=attn_mask)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Transformer Blocks
|
|
# -----------------------------------------------------------------------------
|
|
|
|
class DiTBlockCausalV2(nn.Module):
|
|
def __init__(self, dim, heads, dim_head, ff_mult=4, dropout=0.1, additional_condition_dim=0):
|
|
super().__init__()
|
|
|
|
self.attn_norm = AdaLayerNormZero(dim, additional_condition_dim)
|
|
self.attn = AttentionV2(
|
|
processor=AttnProcessorCausalV2(),
|
|
dim=dim,
|
|
heads=heads,
|
|
dim_head=dim_head,
|
|
dropout=dropout,
|
|
)
|
|
|
|
self.ff_norm = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
|
|
self.ff = FeedForward(dim=dim, mult=ff_mult, dropout=dropout, approximate="tanh")
|
|
|
|
def forward(self, x, t, padding_mask=None, rope=None, attn_mask=None):
|
|
"""
|
|
Args:
|
|
x: Noised input tensor
|
|
t: Time embedding (or concatenated time+condition embedding)
|
|
padding_mask: Mask for padding
|
|
rope: Rotary positional embeddings
|
|
attn_mask: Attention mask (e.g., for causality or blocking)
|
|
"""
|
|
# 1. Pre-norm & Modulation for Attention
|
|
norm, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.attn_norm(x, emb=t)
|
|
|
|
# 2. Attention
|
|
attn_output = self.attn(x=norm, padding_mask=padding_mask, rope=rope, attn_mask=attn_mask)
|
|
|
|
# 3. Residual Connection (modulated)
|
|
x = x + gate_msa.unsqueeze(1) * attn_output
|
|
|
|
# 4. Pre-norm & Modulation for FeedForward
|
|
norm = self.ff_norm(x) * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
|
ff_output = self.ff(norm)
|
|
|
|
# 5. Residual Connection (modulated)
|
|
x = x + gate_mlp.unsqueeze(1) * ff_output
|
|
|
|
return x |