File size: 9,693 Bytes
93cd35a | 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 | """Layer primitives for the ELF transformer."""
import math
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat
# Init defaults:
# - Linear weights: xavier_uniform; biases: 0
# - TimestepEmbedder MLPs and learned tokens: normal(0.02)
# - final_layer.linear: 0 (zero init)
def DEFAULT_KERNEL_INIT(weight: torch.Tensor) -> None:
nn.init.xavier_uniform_(weight)
def DEFAULT_BIAS_INIT(bias: torch.Tensor) -> None:
nn.init.zeros_(bias)
def ZERO_INIT(t: torch.Tensor) -> None:
nn.init.zeros_(t)
def NORMAL_INIT_002(t: torch.Tensor) -> None:
nn.init.normal_(t, mean=0.0, std=0.02)
def _make_linear(in_features: int, out_features: int, bias: bool = True,
kernel_init=DEFAULT_KERNEL_INIT, bias_init=DEFAULT_BIAS_INIT) -> nn.Linear:
"""nn.Linear with explicit initializers."""
layer = nn.Linear(in_features, out_features, bias=bias)
kernel_init(layer.weight)
if bias and bias_init is not None:
bias_init(layer.bias)
return layer
def rotate_half(x: torch.Tensor) -> torch.Tensor:
"""Rotate half the hidden dims of the input."""
x = rearrange(x, '... (d r) -> ... d r', r=2)
x1, x2 = x.unbind(dim=-1)
x = torch.stack((-x2, x1), dim=-1)
return rearrange(x, '... d r -> ... (d r)')
class TextRotaryEmbeddingFast(nn.Module):
"""1D Rotary Position Embedding for text/sequence models."""
def __init__(self, dim: int, pt_seq_len: int = 512,
ft_seq_len: Optional[int] = None, theta: float = 10000.0,
num_empty_token: int = 0):
super().__init__()
self.dim = dim
self.pt_seq_len = pt_seq_len
self.ft_seq_len = ft_seq_len if ft_seq_len is not None else pt_seq_len
self.theta = theta
self.num_empty_token = num_empty_token
freqs_cos, freqs_sin = self._compute_freqs()
self.register_buffer("freqs_cos", freqs_cos, persistent=False)
self.register_buffer("freqs_sin", freqs_sin, persistent=False)
def _compute_freqs(self) -> tuple:
dim = self.dim
ft_seq_len = self.ft_seq_len
pt_seq_len = self.pt_seq_len
freqs = 1.0 / (self.theta ** (
torch.arange(0, dim, 2, dtype=torch.float32)[: dim // 2] / dim
))
pos = torch.arange(ft_seq_len, dtype=torch.float32) / ft_seq_len * pt_seq_len
freqs_main = torch.einsum('..., f -> ... f', pos, freqs)
freqs_main = repeat(freqs_main, '... n -> ... (n r)', r=2)
D = freqs_main.shape[-1]
cos_parts, sin_parts = [], []
# 1. Empty tokens (no rotation): cos=1, sin=0
if self.num_empty_token > 0:
cos_parts.append(torch.ones((self.num_empty_token, D), dtype=freqs.dtype))
sin_parts.append(torch.zeros((self.num_empty_token, D), dtype=freqs.dtype))
# 2. Main tokens (RoPE positions 0 to pt_seq_len-1)
cos_parts.append(torch.cos(freqs_main))
sin_parts.append(torch.sin(freqs_main))
freqs_cos = torch.cat(cos_parts, dim=0) if len(cos_parts) > 1 else cos_parts[0]
freqs_sin = torch.cat(sin_parts, dim=0) if len(sin_parts) > 1 else sin_parts[0]
return freqs_cos, freqs_sin
def forward(self, t: torch.Tensor) -> torch.Tensor:
freqs_cos = self.freqs_cos.to(t.dtype)
freqs_sin = self.freqs_sin.to(t.dtype)
return t * freqs_cos + rotate_half(t) * freqs_sin
class RMSNorm(nn.Module):
"""RMS Normalization layer."""
def __init__(self, hidden_size: int, eps: float = 1e-6):
super().__init__()
self.hidden_size = hidden_size
self.eps = eps
self.weight = nn.Parameter(torch.ones(hidden_size))
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
input_dtype = hidden_states.dtype
variance = hidden_states.float().pow(2).mean(dim=-1, keepdim=True)
inv_std = torch.rsqrt(variance + self.eps).to(input_dtype)
return self.weight.to(input_dtype) * (hidden_states * inv_std)
class BottleneckTextProj(nn.Module):
"""Text projection with bottleneck."""
def __init__(self, text_encoder_dim: int, hidden_size: int, bottleneck_dim: int):
super().__init__()
self.proj1 = _make_linear(text_encoder_dim, bottleneck_dim, bias=False)
self.proj2 = _make_linear(bottleneck_dim, hidden_size, bias=True)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.proj2(self.proj1(x))
class TimestepEmbedder(nn.Module):
"""Embeds scalar timesteps into vector representations."""
def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
super().__init__()
self.hidden_size = hidden_size
self.frequency_embedding_size = frequency_embedding_size
self.mlp_0 = _make_linear(
frequency_embedding_size, hidden_size, bias=True,
kernel_init=NORMAL_INIT_002, bias_init=DEFAULT_BIAS_INIT,
)
self.mlp_2 = _make_linear(
hidden_size, hidden_size, bias=True,
kernel_init=NORMAL_INIT_002, bias_init=DEFAULT_BIAS_INIT,
)
@staticmethod
def timestep_embedding(t: torch.Tensor, dim: int, max_period: int = 10000) -> torch.Tensor:
"""Sinusoidal timestep embeddings: (N,) ints -> (N, dim) floats."""
half = dim // 2
freqs = torch.exp(
-math.log(max_period)
* torch.arange(0, half, dtype=torch.float32, device=t.device)
/ half
)
args = t[:, None].to(torch.float32) * 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: torch.Tensor) -> torch.Tensor:
t_emb = self.mlp_0(self.timestep_embedding(t, self.frequency_embedding_size))
return self.mlp_2(F.silu(t_emb))
def scaled_dot_product_attention(
query: torch.Tensor, key: torch.Tensor, value: torch.Tensor,
attn_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Scaled dot-product attention.
query/key/value: (B, num_heads, L|S, head_dim).
attn_mask: optional int mask (B, S) or (B, L, S); 1=valid, 0=masked.
Returns: (B, num_heads, L, head_dim).
"""
bool_mask: Optional[torch.Tensor] = None
if attn_mask is not None:
if attn_mask.dim() == 2:
bool_mask = attn_mask[:, None, None, :]
elif attn_mask.dim() == 3:
bool_mask = attn_mask[:, None, :, :]
else:
bool_mask = attn_mask
bool_mask = bool_mask.bool()
return F.scaled_dot_product_attention(query, key, value, attn_mask=bool_mask)
class Attention(nn.Module):
"""Multi-head self-attention."""
def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = True,
qk_norm: bool = True, attn_drop: float = 0.0, proj_drop: float = 0.0):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.qk_norm = qk_norm
self.attn_drop = attn_drop
self.proj_drop = proj_drop
head_dim = dim // num_heads
self.qkv = _make_linear(dim, dim * 3, bias=qkv_bias)
self.q_norm = RMSNorm(head_dim) if qk_norm else nn.Identity()
self.k_norm = RMSNorm(head_dim) if qk_norm else nn.Identity()
self.proj = _make_linear(dim, dim, bias=True)
def forward(self, x: torch.Tensor, rope_fn: Optional[nn.Module],
attention_mask: Optional[torch.Tensor] = None,
deterministic: bool = True) -> torch.Tensor:
"""x: (B, N, C). attention_mask: optional int mask (B, N), 1=valid, 0=padded."""
B, N, C = x.shape
head_dim = self.dim // self.num_heads
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, head_dim).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
if self.qk_norm:
q = self.q_norm(q)
k = self.k_norm(k)
if rope_fn is not None:
q = rope_fn(q)
k = rope_fn(k)
x = scaled_dot_product_attention(q, k, v, attn_mask=attention_mask)
x = x.permute(0, 2, 1, 3).reshape(B, N, C)
x = self.proj(x)
if self.proj_drop > 0.0:
x = F.dropout(x, p=self.proj_drop, training=not deterministic)
return x
class SwiGLUFFN(nn.Module):
"""SwiGLU Feed-Forward Network."""
def __init__(self, dim: int, hidden_dim: int, drop: float = 0.0, bias: bool = True):
super().__init__()
hidden_dim_eff = int(hidden_dim * 2 / 3)
self.drop = drop
self.w12 = _make_linear(dim, 2 * hidden_dim_eff, bias=bias)
self.w3 = _make_linear(hidden_dim_eff, dim, bias=bias)
def forward(self, x: torch.Tensor, deterministic: bool = True) -> torch.Tensor:
x12 = self.w12(x)
x1, x2 = x12.chunk(2, dim=-1)
hidden = F.silu(x1) * x2
if self.drop > 0.0:
hidden = F.dropout(hidden, p=self.drop, training=not deterministic)
return self.w3(hidden)
class FinalLayer(nn.Module):
"""The final layer of ELF."""
def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
super().__init__()
self.norm_final = RMSNorm(hidden_size)
# Zero-init linear (kernel & bias both zero).
self.linear = _make_linear(
hidden_size, patch_size * patch_size * out_channels, bias=True,
kernel_init=ZERO_INIT, bias_init=ZERO_INIT,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.linear(self.norm_final(x))
|