File size: 9,764 Bytes
f9740ea | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 | import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import math
from einops import rearrange
# from flash_attn.ops.fused_dense import FusedMLP, FusedDense
from huggingface_hub import PyTorchModelHubMixin
from omegaconf import OmegaConf
from . import rotary
from .fused_add_dropout_scale import (
bias_dropout_add_scale_fused_train,
bias_dropout_add_scale_fused_inference,
get_bias_dropout_add_scale,
modulate_fused,
)
def modulate(x, shift, scale):
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
#################################################################################
# Layers #
#################################################################################
class LayerNorm(nn.Module):
def __init__(self, dim):
super().__init__()
self.weight = nn.Parameter(torch.ones([dim]))
self.dim = dim
def forward(self, x):
with torch.amp.autocast("cuda", enabled=False):
x = F.layer_norm(x.float(), [self.dim])
return x * self.weight[None,None,:]
def residual_linear(x, W, x_skip, residual_scale):
"""x_skip + residual_scale * W @ x"""
dim_out, dim_in = W.shape[0], W.shape[1]
return torch.addmm(
x_skip.view(-1, dim_out),
x.view(-1, dim_in),
W.T,
alpha=residual_scale
).view(*x.shape[:-1], dim_out)
#################################################################################
# Embedding Layers for Timesteps and Class Labels #
#################################################################################
class TimestepEmbedder(nn.Module):
"""
Embeds scalar timesteps into vector representations.
"""
def __init__(self, hidden_size, frequency_embedding_size=256, silu=True):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size, bias=True),
)
self.frequency_embedding_size = frequency_embedding_size
@staticmethod
def timestep_embedding(t, dim, max_period=10000):
"""
Create sinusoidal timestep embeddings.
:param t: a 1-D Tensor of N indices, one per batch element.
These may be fractional.
:param dim: the dimension of the output.
:param max_period: controls the minimum frequency of the embeddings.
:return: an (N, D) Tensor of positional embeddings.
"""
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
half = dim // 2
freqs = torch.exp(
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
).to(device=t.device)
args = t[:, None].float() * freqs[None]
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
return embedding
def forward(self, t):
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
t_emb = self.mlp(t_freq)
return t_emb
class LabelEmbedder(nn.Module):
"""
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
"""
def __init__(self, num_classes, cond_size):
super().__init__()
self.embedding_table = nn.Embedding(num_classes + 1, cond_size)
self.num_classes = num_classes
# TODO think of initializing with 0.02 std deviation like in original DiT paper
def forward(self, labels):
embeddings = self.embedding_table(labels)
return embeddings
#################################################################################
# Core Model #
#################################################################################
class DDiTBlock(nn.Module):
def __init__(self, dim, n_heads, cond_dim, mlp_ratio=4, dropout=0.1):
super().__init__()
self.n_heads = n_heads
self.norm1 = LayerNorm(dim)
self.attn_qkv = nn.Linear(dim, 3 * dim, bias=False)
self.attn_out = nn.Linear(dim, dim, bias=False)
self.dropout1 = nn.Dropout(dropout)
self.norm2 = LayerNorm(dim)
self.mlp = nn.Sequential(
nn.Linear(dim, mlp_ratio * dim, bias=True),
nn.GELU(approximate="tanh"),
nn.Linear(mlp_ratio * dim, dim, bias=True)
)
self.dropout2 = nn.Dropout(dropout)
self.dropout = dropout
self.adaLN_modulation = nn.Linear(cond_dim, 6 * dim, bias=True)
self.adaLN_modulation.weight.data.zero_()
self.adaLN_modulation.bias.data.zero_()
def _get_bias_dropout_scale(self):
return (
bias_dropout_add_scale_fused_train
if self.training
else bias_dropout_add_scale_fused_inference
)
def forward(self, x, rotary_cos_sin, c, seqlens=None):
bias_dropout_scale_fn = self._get_bias_dropout_scale()
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c)[:, None].chunk(6, dim=2)
# attention operation
x_skip = x
x = modulate_fused(self.norm1(x), shift_msa, scale_msa)
# dtype0 = x.dtype
qkv = self.attn_qkv(x)
qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.n_heads)
with torch.amp.autocast("cuda", enabled=False):
cos, sin = rotary_cos_sin
qkv = rotary.apply_rotary_pos_emb(
qkv, cos.to(qkv.dtype), sin.to(qkv.dtype)
)
if seqlens is not None:
raise NotImplementedError(
"Variable-length attention is not used by SEDD's forward pass."
)
q, k, v = qkv.unbind(dim=2)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=0.0, is_causal=False
)
x = rearrange(x, 'b h s d -> b s (h d)')
x = bias_dropout_scale_fn(self.attn_out(x), None, gate_msa, x_skip, self.dropout)
# mlp operation
x = bias_dropout_scale_fn(self.mlp(modulate_fused(self.norm2(x), shift_mlp, scale_mlp)), None, gate_mlp, x, self.dropout)
return x
class EmbeddingLayer(nn.Module):
def __init__(self, dim, vocab_dim):
"""
Mode arg: 0 -> use a learned layer, 1 -> use eigenvectors,
2-> add in eigenvectors, 3 -> use pretrained embedding matrix
"""
super().__init__()
self.embedding = nn.Parameter(torch.empty((vocab_dim, dim)))
torch.nn.init.kaiming_uniform_(self.embedding, a=math.sqrt(5))
def forward(self, x):
return self.embedding[x]
class DDitFinalLayer(nn.Module):
def __init__(self, hidden_size, out_channels, cond_dim):
super().__init__()
self.norm_final = LayerNorm(hidden_size)
self.linear = nn.Linear(hidden_size, out_channels)
self.linear.weight.data.zero_()
self.linear.bias.data.zero_()
self.adaLN_modulation = nn.Linear(cond_dim, 2 * hidden_size, bias=True)
self.adaLN_modulation.weight.data.zero_()
self.adaLN_modulation.bias.data.zero_()
def forward(self, x, c):
shift, scale = self.adaLN_modulation(c)[:, None].chunk(2, dim=2)
x = modulate_fused(self.norm_final(x), shift, scale)
x = self.linear(x)
return x
class SEDD(nn.Module, PyTorchModelHubMixin):
def __init__(self, config):
super().__init__()
# hack to make loading in configs easier
if type(config) == dict:
config = OmegaConf.create(config)
self.config = config
self.absorb = config.graph.type == "absorb"
vocab_size = config.tokens + (1 if self.absorb else 0)
self.vocab_embed = EmbeddingLayer(config.model.hidden_size, vocab_size)
self.sigma_map = TimestepEmbedder(config.model.cond_dim)
self.rotary_emb = rotary.Rotary(config.model.hidden_size // config.model.n_heads)
self.blocks = nn.ModuleList([
DDiTBlock(config.model.hidden_size, config.model.n_heads, config.model.cond_dim, dropout=config.model.dropout) for _ in range(config.model.n_blocks)
])
self.output_layer = DDitFinalLayer(config.model.hidden_size, vocab_size, config.model.cond_dim)
self.scale_by_sigma = config.model.scale_by_sigma
def _get_bias_dropout_scale(self):
return (
bias_dropout_add_scale_fused_train
if self.training
else bias_dropout_add_scale_fused_inference
)
def forward(self, indices, sigma):
x = self.vocab_embed(indices)
c = F.silu(self.sigma_map(sigma))
rotary_cos_sin = self.rotary_emb(x)
with torch.amp.autocast("cuda", dtype=torch.bfloat16):
for i in range(len(self.blocks)):
x = self.blocks[i](x, rotary_cos_sin, c, seqlens=None)
x = self.output_layer(x, c)
if self.scale_by_sigma:
assert self.absorb, "Haven't configured this to work."
esigm1_log = torch.where(sigma < 0.5, torch.expm1(sigma), sigma.exp() - 1).log().to(x.dtype)[:, None, None]
x = x - esigm1_log - np.log(x.shape[-1] - 1)# this will be approximately averaged at 0
x = torch.scatter(x, -1, indices[..., None], torch.zeros_like(x[..., :1]))
return x
|