File size: 10,480 Bytes
ceac1e4 | 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 | """sigma_core.py — SigMa positional-encoding core, EXTRACTED VERBATIM from
github.com/bxuanz/SigMa flux/transformer_flux.py (ICML 2026 paper 47JZSOkw5C).
Only the diffusers-independent RoPE/SigMa math is copied here so it runs without
the repo's heavy diffusers/transformers imports (which clash with the local
huggingface-hub version). No logic is modified. Line provenance in transformer_flux.py:
get_adaptive_scale L89-111
find_correction_factor L534-535
find_correction_range L538-544
linear_ramp_mask L547-553
find_newbase_ntk L556-560
get_1d_rotary_pos_embed L568-685
FluxPosEmbed L687-776
"""
import math
import torch
import torch.nn as nn
import numpy as np
from typing import List, Union
def get_adaptive_scale(t: float, scale_factor: float) -> float:
"""
Logit-space SigMa scheduler:
mu_d(t) = sigmoid(gamma_d * (logit(t) - logit(t_c,d))).
"""
t_center = 1.0 / scale_factor
gamma_d = math.sqrt(scale_factor)
# logit(t) is defined on (0, 1); Flux can pass t=1 at the first step.
eps = 1e-6
t = min(max(t, eps), 1.0 - eps)
t_center = min(max(t_center, eps), 1.0 - eps)
def logit(value: float) -> float:
return math.log(value / (1.0 - value))
# 注意:Flux 中 t=1 是噪声,t=0 是图。
# 当 t > t_center (早期),x > 0 -> alpha -> 1 (使用 NTK/YaRN)
# 当 t < t_center (晚期),x < 0 -> alpha -> 0 (回归 Base 以获得锐利纹理)
x = gamma_d * (logit(t) - logit(t_center))
alpha = 1 / (1 + math.exp(-x))
return alpha
def find_correction_factor(num_rotations, dim, base, max_position_embeddings):
return (dim * math.log(max_position_embeddings/(num_rotations * 2 * math.pi)))/(2 * math.log(base)) #Inverse dim formula to find number of rotations
def find_correction_range(low_ratio, high_ratio, dim, base, ori_max_pe_len):
"""
Find the correction range for NTK-by-parts interpolation.
"""
low = np.floor(find_correction_factor(low_ratio, dim, base, ori_max_pe_len))
high = np.ceil(find_correction_factor(high_ratio, dim, base, ori_max_pe_len))
return max(low, 0), min(high, dim-1) #Clamp values just in case
def linear_ramp_mask(min, max, dim):
if min == max:
max += 0.001 #Prevent singularity
linear_func = (torch.arange(dim, dtype=torch.float32) - min) / (max - min)
ramp_func = torch.clamp(linear_func, 0, 1)
return ramp_func
def find_newbase_ntk(dim, base, scale):
"""
Calculate the new base for NTK-aware scaling.
"""
return base * (scale ** (dim / (dim - 2)))
def get_1d_rotary_pos_embed(
dim: int,
pos: Union[np.ndarray, int],
theta: float = 10000.0,
use_real=False,
linear_factor=1.0,
ntk_factor=1.0,
repeat_interleave_real=True,
freqs_dtype=torch.float32,
yarn=False,
max_pe_len=None,
ori_max_pe_len=64, # [重要] 听你的,保持 64 不动!这是画质的基石。
sigma=False,
current_timestep=1.0,
gamma_factor=1.0,
):
assert dim % 2 == 0
if isinstance(pos, int):
pos = torch.arange(pos)
if isinstance(pos, np.ndarray):
pos = torch.from_numpy(pos)
device = pos.device
# 这里的 scale 用于计算 RoPE 频率,必须基于 ori_max_pe_len=64
if yarn and max_pe_len is not None and max_pe_len > ori_max_pe_len:
if not isinstance(max_pe_len, torch.Tensor):
max_pe_len = torch.tensor(max_pe_len, dtype=freqs_dtype, device=device)
# [Track 1: 几何缩放]
# 保持 64 基准,scale 约为 64.0 (4096/64)
# 这一步保证了图像质量不下降
scale = torch.clamp_min(max_pe_len / ori_max_pe_len, 1.0)
scale_val = scale.item()
# YaRN 默认参数
beta_0 = 1.25
beta_1 = 0.75
gamma_0 = 16
gamma_1 = 2
freqs_base = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=device) / dim))
# 这里的 freqs_linear 使用 Base-64 的 scale,保证坐标系正确
freqs_linear = 1.0 / torch.einsum(
'..., f -> ... f',
scale,
(theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=device) / dim))
)
new_base = find_newbase_ntk(dim, theta, scale)
if new_base.dim() > 0:
new_base = new_base.view(-1, 1)
freqs_ntk = 1.0 / torch.pow(
new_base,
(torch.arange(0, dim, 2, dtype=freqs_dtype, device=device) / dim)
)
if freqs_ntk.dim() > 1:
freqs_ntk = freqs_ntk.squeeze()
# -----------------------------------------------------------
# [SigMa core logic]
# -----------------------------------------------------------
if sigma:
adaptive_alpha = get_adaptive_scale(current_timestep, scale_val)
beta_0 = beta_0 * adaptive_alpha
beta_1 = beta_1 * adaptive_alpha
low, high = find_correction_range(beta_0, beta_1, dim, theta, ori_max_pe_len)
low = max(0, low)
high = min(dim // 2, high)
freqs_mask = (1 - linear_ramp_mask(low, high, dim // 2).to(device).to(freqs_dtype))
freqs = freqs_linear * (1 - freqs_mask) + freqs_ntk * freqs_mask
if sigma:
gamma_0 = gamma_0 * adaptive_alpha
gamma_1 = gamma_1 * adaptive_alpha
low, high = find_correction_range(gamma_0, gamma_1, dim, theta, ori_max_pe_len)
low = max(0, low)
high = min(dim // 2, high)
freqs_mask = (1 - linear_ramp_mask(low, high, dim // 2).to(device).to(freqs_dtype))
freqs = freqs * (1 - freqs_mask) + freqs_base * freqs_mask
else:
theta_ntk = theta * ntk_factor
freqs = 1.0 / (theta_ntk ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=device) / dim)) / linear_factor
freqs = torch.outer(pos, freqs)
is_npu = freqs.device.type == "npu"
if is_npu:
freqs = freqs.float()
if use_real and repeat_interleave_real:
freqs_cos = freqs.cos().repeat_interleave(2, dim=1, output_size=freqs.shape[1] * 2).float()
freqs_sin = freqs.sin().repeat_interleave(2, dim=1, output_size=freqs.shape[1] * 2).float()
# MScale 逻辑
if yarn and max_pe_len is not None and max_pe_len > ori_max_pe_len:
scale_factor_tensor = scale if isinstance(scale, torch.Tensor) else torch.tensor(scale)
# MScale 这里的公式 0.1 * ln(scale)
target_mscale = 0.1 * torch.log(scale_factor_tensor) + 1.0
if sigma:
adaptive_alpha = get_adaptive_scale(current_timestep, scale_val)
mscale = (target_mscale - 1.0) * adaptive_alpha + 1.0
else:
mscale = target_mscale
mscale = mscale.to(freqs_cos.device)
freqs_cos = freqs_cos * mscale
freqs_sin = freqs_sin * mscale
return freqs_cos, freqs_sin
class FluxPosEmbed(nn.Module):
def __init__(
self,
theta: int,
axes_dim: List[int],
method: str = 'yarn',
sigma: bool = True,
gamma_factor: float = 1,
):
super().__init__()
self.theta = theta
self.axes_dim = axes_dim
self.base_resolution = 1024
self.patch_size = 16
self.base_patches = self.base_resolution // self.patch_size
self.method = method
self.sigma = sigma if method != 'base' else False
self.current_timestep = 1.0
self.gamma_factor = gamma_factor
def set_timestep(self, timestep: float):
"""Set current timestep for SigMa."""
self.current_timestep = timestep
def forward(self, ids: torch.Tensor) -> torch.Tensor:
n_axes = ids.shape[-1]
cos_out = []
sin_out = []
pos = ids.float()
is_mps = ids.device.type == "mps"
is_npu = ids.device.type == "npu"
freqs_dtype = torch.float32 if (is_mps or is_npu) else torch.float64
for i in range(n_axes):
common_kwargs = {
'dim': self.axes_dim[i],
'pos': pos[:, i],
'theta': self.theta,
'repeat_interleave_real': True,
'use_real': True,
'freqs_dtype': freqs_dtype,
}
if i > 0:
max_pos = pos[:, i].max().item()
current_patches = max_pos + 1
if self.method == 'yarn' and current_patches > self.base_patches:
max_pe_len = torch.tensor(current_patches, dtype=freqs_dtype, device=pos.device)
cos, sin = get_1d_rotary_pos_embed(
**common_kwargs,
yarn=True,
max_pe_len=max_pe_len,
ori_max_pe_len=self.base_patches,
sigma=self.sigma,
current_timestep=self.current_timestep,
gamma_factor=self.gamma_factor,
)
elif self.method == 'ntk' and current_patches > self.base_patches:
# 计算基础 NTK 因子
scale_s = current_patches / self.base_patches
base_ntk = scale_s ** (self.axes_dim[i] / (self.axes_dim[i] - 2))
# [SigMa core update: dynamic NTK]
if self.sigma:
# 1. 计算自适应强度 alpha
adaptive_alpha = get_adaptive_scale(self.current_timestep, scale_s)
# 2. 应用强度
# [修改] 移除 2.0,回归 power 1.0 (adaptive_alpha)
ntk_factor = base_ntk ** (adaptive_alpha)
else:
ntk_factor = base_ntk
ntk_factor = max(1.0, ntk_factor)
cos, sin = get_1d_rotary_pos_embed(**common_kwargs, ntk_factor=ntk_factor)
else:
cos, sin = get_1d_rotary_pos_embed(**common_kwargs)
else:
cos, sin = get_1d_rotary_pos_embed(**common_kwargs)
cos_out.append(cos)
sin_out.append(sin)
freqs_cos = torch.cat(cos_out, dim=-1).to(ids.device)
freqs_sin = torch.cat(sin_out, dim=-1).to(ids.device)
return freqs_cos, freqs_sin
|