File size: 10,583 Bytes
c13bd26 cd66851 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple, Optional, Union
import math
class YARNScaling:
@staticmethod
def compute_yarn_parameters(
original_max_len: int,
target_max_len: int=8192,
dim: int=128,
base: int = 10000,
beta_fast: int = 32,
beta_slow: int = 1,
alpha: float = 1.0,
device: Optional[torch.device] = None
) -> Tuple[torch.Tensor, float]:
scale = float(target_max_len) / original_max_len
mscale = YARNScaling.compute_mscale(scale, alpha)
# 确保 dim 为 float 以进行除法运算
# RoPE 频率是成对的 (0, 2, ..., d-2)
freqs_idx = torch.arange(0, dim, 2, dtype=torch.float32, device=device)
# 基础频率 (Original RoPE)
freq_extra = 1.0 / (base ** (freqs_idx / dim))
# 如果不需要缩放,直接返回基础频率
if scale <= 1.0:
return freq_extra, 1.0
# 插值频率 (Interpolated for extension)
freq_inter = 1.0 / (scale * base ** (freqs_idx / dim))
def get_limit(beta):
return dim * math.log(original_max_len / (2 * math.pi * beta)) / (2 * math.log(base))
low = max(math.floor(get_limit(beta_fast)), 0)
high = min(math.ceil(get_limit(beta_slow)), dim // 2 - 1)
indices = torch.arange(0, dim // 2, dtype=torch.float32, device=device)
inv_freq = freq_extra.clone()
mask_low_freq = indices > high
inv_freq[mask_low_freq] = freq_inter[mask_low_freq]
mid_mask = (indices >= low) & (indices <= high)
if mid_mask.any():
# 避免除以 0
denom = max(high - low, 1)
t = (indices[mid_mask] - low) / denom
inv_freq[mid_mask] = freq_extra[mid_mask] * (1 - t) + freq_inter[mid_mask] * t
return inv_freq, float(mscale)
@staticmethod
def compute_mscale(scale: float, alpha: float = 1.0) -> float:
"""计算注意力缩放因子 (Temperature scaling)"""
if scale <= 1.0:
return 1.0
return 0.1 * math.log(scale) + 1.0
class YARNRotaryEmbedding(nn.Module):
def __init__(
self,
dim: int = 64,
max_seq_len: int = 8192,
original_max_len: int = 4096,
base: int = 10000,
scaling_factor: float = 1.0,
beta_fast: int = 32,
beta_slow: int = 1,
alpha: float = 1.0,
rope_percentage: float = 1.0,
device: Optional[torch.device] = None
):
super().__init__()
self.dim = dim
self.max_seq_len = max_seq_len
self.original_max_len = original_max_len
self.base = base
self.alpha = alpha
# 计算实际应用 RoPE 的维度
self.rope_dim = int(dim * rope_percentage)
# 确保是偶数
if self.rope_dim % 2 != 0:
self.rope_dim -= 1
# 初始化频率 (Persistent state)
self._init_yarn_frequencies(device)
# 缓存 cos/sin
self.register_buffer("cos_cached", None, persistent=False)
self.register_buffer("sin_cached", None, persistent=False)
def _init_yarn_frequencies(self, device: Optional[torch.device] = None):
inv_freq, mscale = YARNScaling.compute_yarn_parameters(
self.original_max_len,
self.max_seq_len,
self.rope_dim,
self.base,
beta_fast=32,
beta_slow=1,
alpha=self.alpha,
device=device
)
self.register_buffer("inv_freq", inv_freq, persistent=True)
self.register_buffer("mscale", torch.tensor(mscale, dtype=torch.float32, device=device), persistent=True)
def _compute_cos_sin_cache(
self,
needed_len: int,
device: torch.device,
dtype: torch.dtype
):
alloc_len = max(needed_len, self.max_seq_len)
if (self.cos_cached is not None and
self.cos_cached.shape[2] >= alloc_len and
self.cos_cached.device == device):
return
t = torch.arange(alloc_len, dtype=torch.float32, device=device)
freqs = torch.outer(t, self.inv_freq.to(device))
emb = torch.cat((freqs, freqs), dim=-1)
cos_cached = (emb.cos() * self.mscale).view(1, 1, alloc_len, self.rope_dim)
sin_cached = (emb.sin() * self.mscale).view(1, 1, alloc_len, self.rope_dim)
self.cos_cached = cos_cached.to(dtype)
self.sin_cached = sin_cached.to(dtype)
@staticmethod
def rotate_half(x: torch.Tensor) -> torch.Tensor:
x1, x2 = x.chunk(2, dim=-1)
return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb(
self,
q: torch.Tensor,
k: torch.Tensor,
position_ids: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
bsz, num_heads, seq_len, head_dim = q.shape
if position_ids is not None:
max_pos = position_ids.max().item() + 1
needed_len = max(max_pos, seq_len)
else:
needed_len = seq_len
if (self.cos_cached is None or
self.cos_cached.shape[2] < needed_len or
self.cos_cached.device != q.device):
self._compute_cos_sin_cache(needed_len, q.device, q.dtype)
if position_ids is not None:
cos = self.cos_cached[0, 0][position_ids].unsqueeze(1)
sin = self.sin_cached[0, 0][position_ids].unsqueeze(1)
else:
cos = self.cos_cached[:, :, :seq_len, :]
sin = self.sin_cached[:, :, :seq_len, :]
if self.rope_dim < head_dim:
q_rot = q[..., :self.rope_dim]
q_pass = q[..., self.rope_dim:]
k_rot = k[..., :self.rope_dim]
k_pass = k[..., self.rope_dim:]
else:
q_rot = q
k_rot = k
q_pass = None
k_pass = None
q_rot_float = q_rot.float()
k_rot_float = k_rot.float()
cos_float = cos.float()
sin_float = sin.float()
q_embed = (q_rot_float * cos_float) + (self.rotate_half(q_rot_float) * sin_float)
k_embed = (k_rot_float * cos_float) + (self.rotate_half(k_rot_float) * sin_float)
q_embed = q_embed.type_as(q)
k_embed = k_embed.type_as(k)
if q_pass is not None:
q_embed = torch.cat([q_embed, q_pass], dim=-1)
k_embed = torch.cat([k_embed, k_pass], dim=-1)
return q_embed, k_embed
def forward(
self,
q: torch.Tensor,
k: torch.Tensor,
position_ids: Optional[torch.Tensor] = None
) -> Tuple[torch.Tensor, torch.Tensor]:
return self.apply_rotary_pos_emb(q, k, position_ids)
def extra_repr(self) -> str:
return (f"dim={self.dim}, rope_dim={self.rope_dim}, "
f"max_seq_len={self.max_seq_len}, original_max_len={self.original_max_len}, "
f"base={self.base}")
class RMSNorm(nn.Module):
def __init__(
self,
dim: int,
eps: float = 1e-6,
elementwise_affine: bool = True
):
super().__init__()
self.eps = eps
self.elementwise_affine = elementwise_affine
if self.elementwise_affine:
self.weight = nn.Parameter(torch.ones(dim))
else:
self.register_parameter('weight', None)
def _norm(self, x: torch.Tensor) -> torch.Tensor:
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
def forward(self, x: torch.Tensor) -> torch.Tensor:
output = self._norm(x.float())
output = output.type_as(x)
if self.elementwise_affine and self.weight is not None:
output = output * self.weight
return output
class QKNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.query_norm = RMSNorm(dim, eps=eps)
self.key_norm = RMSNorm(dim, eps=eps)
def forward(
self,
q: torch.Tensor,
k: torch.Tensor
) -> Tuple[torch.Tensor, torch.Tensor]:
q = self.query_norm(q)
k = self.key_norm(k)
return q, k
class SwiGLU(nn.Module):
def __init__(
self,
dim: int,
hidden_dim: Optional[int] = None,
multiple_of: int = 256,
ffn_dim_multiplier: Optional[float] = None,
dropout: float = 0.0,
bias: bool = False
):
super().__init__()
if hidden_dim is None:
if ffn_dim_multiplier is not None:
hidden_dim = int(dim * ffn_dim_multiplier)
else:
# 默认: 2/3 * 4 * dim = 8/3 * dim
hidden_dim = int(2 * dim * 4 / 3)
# 确保 hidden_dim 是 multiple_of 的倍数 (通常为了 GPU 核心优化)
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
self.hidden_dim = hidden_dim
# W1: Gate, W3: Up, W2: Down (Standard LLaMA naming conventions)
self.w1 = nn.Linear(dim, hidden_dim, bias=bias)
self.w2 = nn.Linear(hidden_dim, dim, bias=bias)
self.w3 = nn.Linear(dim, hidden_dim, bias=bias)
self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
def forward(self, x: torch.Tensor) -> torch.Tensor:
# SwiGLU(x) = (SiLU(W1·x) ⊙ W3·x) · W2
return self.dropout(self.w2(F.silu(self.w1(x)) * self.w3(x)))
class ParallelAttentionFFN(nn.Module):
def __init__(
self,
dim: int,
attn_module: nn.Module,
ffn_module: nn.Module,
norm_eps: float = 1e-6
):
super().__init__()
self.attn_norm = RMSNorm(dim, eps=norm_eps)
self.ffn_norm = RMSNorm(dim, eps=norm_eps)
self.attn = attn_module
self.ffn = ffn_module
def forward(
self,
x: torch.Tensor,
**attn_kwargs
) -> torch.Tensor:
# 并行计算:从同一个 x (normalize 后) 分叉
attn_input = self.attn_norm(x)
ffn_input = self.ffn_norm(x)
# 计算注意力
attn_out = self.attn(attn_input, **attn_kwargs)
# 计算 FFN (确保不传递 attn 特定的 kwargs)
ffn_out = self.ffn(ffn_input)
# 一次性残差连接
return x + attn_out + ffn_out |