File size: 6,994 Bytes
89da923 | 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 | """
SAE Model - TopK Sparse Autoencoder
"""
import math
from dataclasses import dataclass
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
@dataclass
class SAEConfig:
"""SAE 配置"""
input_dim: int = 4096
dict_size: int = 32768
k: int = 128
device: str = "cuda"
dtype: str = "bfloat16"
def get_torch_dtype(self) -> torch.dtype:
dtype_map = {
"float32": torch.float32,
"float16": torch.float16,
"bfloat16": torch.bfloat16,
}
return dtype_map.get(self.dtype, torch.bfloat16)
class TopKSAE(nn.Module):
"""
TopK Sparse Autoencoder
"""
def __init__(self, config: SAEConfig):
super().__init__()
self.config = config
self.decoder = nn.Linear(config.dict_size, config.input_dim, bias=False)
self._normalize_decoder()
self.encoder = nn.Linear(config.input_dim, config.dict_size, bias=False)
self.encoder.weight.data = self.decoder.weight.data.T.clone()
self.pre_bias = nn.Parameter(torch.zeros(config.input_dim))
self.to(config.device)
self.to(config.get_torch_dtype())
def _normalize_decoder(self):
"""归一化 decoder 权重（每列归一化）"""
with torch.no_grad():
self.decoder.weight.data = F.normalize(
self.decoder.weight.data, dim=0
)
def encode(self, x: torch.Tensor) -> torch.Tensor:
"""编码：input -> latent activations
Args:
x: [batch, input_dim]
Returns:
latent: [batch, dict_size] 稀疏激活
"""
centered_x = x - self.pre_bias
pre_activation = self.encoder(centered_x)
topk_values, topk_indices = torch.topk(
pre_activation, k=self.config.k, dim=-1
)
latents = torch.zeros_like(pre_activation)
latents.scatter_(-1, topk_indices, F.relu(topk_values))
return latents
def decode(self, latents: torch.Tensor) -> torch.Tensor:
"""解码：latent -> reconstruction
Args:
latent: [batch, dict_size]
Returns:
reconstruction: [batch, input_dim]
"""
return self.decoder(latents) + self.pre_bias
def forward(
self,
x: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
latents = self.encode(x)
x_hat = self.decode(latents)
return x_hat, latents
def compute_loss(
self,
x: torch.Tensor,
) -> Tuple[torch.Tensor, dict]:
"""计算损失
Args:
x: [batch, input_dim]
Returns:
(loss, loss_dict)
"""
x_hat, latent = self.forward(x)
loss = (((x_hat - x) ** 2).mean(dim=-1) / (x**2).mean(dim=-1)).mean()
loss_dict = {
"mean_activation": latent[latent > 0].mean().item() if (latent > 0).any() else 0,
}
return loss, loss_dict
def get_feature_activations(
self,
x: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""获取特征激活及其索引
Args:
x: [batch, input_dim]
Returns:
values: [batch, k] TopK 激活值
indices: [batch, k] TopK 特征索引
"""
centered_x = x - self.pre_bias
pre_activation = self.encoder(centered_x)
topk_values, topk_indices = torch.topk(
pre_activation, k=self.config.k, dim=-1
)
return F.relu(topk_values), topk_indices
def get_decoder_vectors(self, indices: torch.Tensor) -> torch.Tensor:
"""获取指定特征的 decoder 向量
Args:
indices: [batch, k] 特征索引
Returns:
vectors: [batch, k, input_dim] decoder 向量
"""
return self.decoder.weight[:, indices].permute(1, 2, 0)
def steer(
self,
x: torch.Tensor,
feature_idx: int,
strength: float,
) -> torch.Tensor:
"""在原始 activation 上直接加减 decoder 方向来 steer 指定特征。
h_new = h + (strength - 1) * z_i * d_i
其中 z_i 是 feature_idx 的 TopK 编码激活值，d_i 是其 decoder 列向量。
不走完整 encode→decode，避免 TopK 重建误差污染其他特征。
strength=1.0 时变化量为 0（真正的 baseline）。
Args:
x: [batch, input_dim] 输入激活
feature_idx: 要 steering 的特征索引
strength: 目标特征缩放系数（1.0 不变，>1 增强，<1 抑制，0 完全消除）
Returns:
steered_x: [batch, input_dim] 调整后的激活
"""
with torch.no_grad():
latents = self.encode(x) # [batch, dict_size]
z_i = latents[:, feature_idx] # [batch]
d_i = self.decoder.weight[:, feature_idx] # [input_dim]
delta = (strength - 1) * z_i.unsqueeze(-1) * d_i.unsqueeze(0)
return x + delta
def steer_multi(
self,
x: torch.Tensor,
feature_indices: list,
strengths: list,
) -> torch.Tensor:
"""在原始 activation 上直接加减 decoder 方向来同时 steer 多个特征。
h_new = h + sum_i (strength_i - 1) * z_i * d_i
不走完整 encode→decode，避免 TopK 重建误差污染其他特征。
strength=1.0 时对应特征的变化量为 0（真正的 baseline）。
Args:
x: [batch, input_dim] 输入激活
feature_indices: 要 steering 的特征索引列表
strengths: 对应的缩放系数列表
Returns:
steered_x: [batch, input_dim] 调整后的激活
"""
with torch.no_grad():
latents = self.encode(x) # [batch, dict_size]
delta = torch.zeros_like(x)
for feat_idx, strength in zip(feature_indices, strengths):
z_i = latents[:, feat_idx] # [batch]
d_i = self.decoder.weight[:, feat_idx] # [input_dim]
delta += (strength - 1) * z_i.unsqueeze(-1) * d_i.unsqueeze(0)
return x + delta
def save(self, path: str):
"""保存模型"""
torch.save({
"config": self.config,
"state_dict": self.state_dict(),
}, path)
@classmethod
def load(cls, path: str, device: str = "cuda") -> "TopKSAE":
"""加载模型"""
checkpoint = torch.load(path, map_location=device, weights_only=False)
config = checkpoint["config"]
config.device = device
model = cls(config)
model.load_state_dict(checkpoint["state_dict"])
model.to(config.get_torch_dtype())
return model
|