File size: 10,354 Bytes
3bdb743 | 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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from transformers import PreTrainedModel
try:
from .configuration_pillars import PillarsConfig
except ImportError:
from configuration_pillars import PillarsConfig
try:
from x_transformers import Encoder
except ImportError:
raise ImportError("To use PILLARS, you must run: pip install x-transformers")
# --- UTILS ---
class ComplexDropout(nn.Module):
def __init__(self, p=0.5):
super().__init__()
self.p = p
def forward(self, z):
if not self.training or self.p == 0.0: return z
mask = torch.ones_like(z.real)
mask = F.dropout(mask, self.p, self.training, inplace=False)
return z * mask
class RobustPhaseNorm(nn.Module):
def __init__(self, d_model, eps=1e-5):
super().__init__()
self.scale = nn.Parameter(torch.ones(d_model))
self.eps = eps
def forward(self, x):
mag = torch.abs(x)
rms = torch.sqrt(torch.mean(mag**2, dim=-1, keepdim=True) + self.eps)
return (x / rms) * self.scale
class ModReLU(nn.Module):
def __init__(self, features):
super().__init__()
self.b = nn.Parameter(torch.zeros(features))
def forward(self, z):
mag = torch.abs(z)
new_mag = F.relu(mag + self.b)
phase = z / (mag + 1e-6)
return new_mag * phase
class ComplexToRealBridge(nn.Module):
def __init__(self, d_model):
super().__init__()
self.proj = nn.Linear(d_model * 2, d_model)
self.norm = nn.LayerNorm(d_model)
def forward(self, x_complex):
cat = torch.cat([x_complex.real, x_complex.imag], dim=-1)
return self.norm(self.proj(cat))
# --- COMPONENTS ---
class DynamicRoSE(nn.Module):
def __init__(self, num_embeddings, embedding_dim, max_period=10000.0):
super().__init__()
self.embedding_dim = embedding_dim
self.raw_embedding = nn.Embedding(num_embeddings, embedding_dim)
self.adapter = nn.Linear(embedding_dim, embedding_dim * 2)
freqs = torch.exp(torch.arange(0, embedding_dim, dtype=torch.float32) * -(math.log(max_period) / embedding_dim))
self.register_buffer('freqs', freqs)
self.rotation_predictor = nn.Linear(embedding_dim, embedding_dim * 2)
def forward(self, input_ids):
real_base = self.raw_embedding(input_ids)
B, L, D = real_base.shape
complex_params = self.adapter(real_base)
z_t = torch.complex(complex_params[..., :D], complex_params[..., D:])
rot_raw = self.rotation_predictor(real_base)
rot_x, rot_y = rot_raw.chunk(2, dim=-1)
rot_mag = torch.sqrt(rot_x**2 + rot_y**2 + 1e-6)
dynamic_rot = torch.complex(rot_x / rot_mag, rot_y / rot_mag)
pos = torch.arange(L, device=input_ids.device).float()
static_angles = torch.outer(pos, self.freqs)
static_rot = torch.polar(torch.ones_like(static_angles), static_angles)
z_final = z_t * static_rot.unsqueeze(0) * dynamic_rot
return z_final, real_base
class HyenaNeuralFilter(nn.Module):
def __init__(self, d_model, max_len=1024, hidden_dim=64):
super().__init__()
self.d_model = d_model
freqs = torch.exp(torch.arange(0, hidden_dim, 2, dtype=torch.float32) * -(math.log(10000.0) / hidden_dim))
self.register_buffer("freqs", freqs)
self.mlp = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim), nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim), nn.SiLU(),
nn.Linear(hidden_dim, d_model * 2)
)
def forward(self, L, device):
t = torch.linspace(0, 1, steps=L, device=device).unsqueeze(-1)
emb = torch.cat([torch.sin(t * self.freqs), torch.cos(t * self.freqs)], dim=-1)
out = self.mlp(emb).view(L, self.d_model, 2)
return torch.complex(out[..., 0], out[..., 1])
class GatedHarmonicConvolution(nn.Module):
def __init__(self, d_model, max_len=1024, dropout=0.1):
super().__init__()
self.d_model = d_model
self.filter_len = max_len
self.neural_filter = HyenaNeuralFilter(d_model, max_len=max_len)
self.gate_proj = nn.Linear(d_model * 2, d_model * 2)
self.mix_real = nn.Linear(d_model, d_model)
self.mix_imag = nn.Linear(d_model, d_model)
self.out_real = nn.Linear(d_model, d_model)
self.out_imag = nn.Linear(d_model, d_model)
self.activation = ModReLU(d_model)
self.norm = RobustPhaseNorm(d_model)
self.dropout = ComplexDropout(dropout)
def forward(self, x, src_mask=None):
residual = x
x_norm = self.norm(x)
if src_mask is not None:
x_norm = x_norm.masked_fill(src_mask.unsqueeze(-1), 0.0)
B, L, D = x_norm.shape
eff_L = min(L, self.filter_len)
x_freq = torch.fft.fft(x_norm, n=eff_L, dim=1, norm='ortho')
h = self.neural_filter(eff_L, x.device).unsqueeze(0)
x_filtered = x_freq * h
x_time = torch.fft.ifft(x_filtered, n=eff_L, dim=1, norm='ortho')
if L > eff_L: x_time = F.pad(x_time, (0,0,0,L-eff_L))
else: x_time = x_time[:, :L, :]
gates = torch.sigmoid(self.gate_proj(torch.cat([x_norm.real, x_norm.imag], dim=-1)))
g_r, g_i = gates.chunk(2, dim=-1)
x_gated = torch.complex(x_time.real * g_r, x_time.imag * g_i)
mr, mi = self.mix_real, self.mix_imag
x_mixed = torch.complex(mr(x_gated.real) - mi(x_gated.imag), mr(x_gated.imag) + mi(x_gated.real))
x_act = self.activation(x_mixed)
or_, oi = self.out_real, self.out_imag
out = torch.complex(or_(x_act.real) - oi(x_act.imag), or_(x_act.imag) + oi(x_act.real))
return self.dropout(out) + residual
class PRISMEncoder(nn.Module):
def __init__(self, num_layers, d_model, max_len, dropout=0.1):
super().__init__()
self.layers = nn.ModuleList([
GatedHarmonicConvolution(d_model, max_len, dropout)
for _ in range(num_layers)
])
self.final_norm = RobustPhaseNorm(d_model)
def forward(self, x, src_mask=None):
for layer in self.layers:
if self.training: x = torch.utils.checkpoint.checkpoint(layer, x, src_mask, use_reentrant=False)
else: x = layer(x, src_mask)
return self.final_norm(x)
class FNetBlock(nn.Module):
def __init__(self, d_model, d_ff, dropout):
super().__init__()
self.norm_mix = nn.LayerNorm(d_model)
self.norm_ff = nn.LayerNorm(d_model)
self.mix_dropout = nn.Dropout(dropout)
self.ff = nn.Sequential(
nn.Linear(d_model, d_ff), nn.GELU(), nn.Dropout(dropout),
nn.Linear(d_ff, d_model), nn.Dropout(dropout)
)
def forward(self, x):
residual = x
x = self.norm_mix(x)
with torch.cuda.amp.autocast(enabled=False):
x = x.float()
x = torch.fft.fftn(x, dim=(-2, -1), norm='ortho').real
x = x.to(dtype=residual.dtype)
x = self.mix_dropout(x)
x = x + residual
residual = x
x = self.norm_ff(x)
x = self.ff(x)
return x + residual
class FNetEncoder(nn.Module):
def __init__(self, depth, d_model, d_ff, dropout):
super().__init__()
self.layers = nn.ModuleList([
FNetBlock(d_model, d_ff, dropout) for _ in range(depth)
])
self.norm_out = nn.LayerNorm(d_model)
def forward(self, x):
for layer in self.layers:
x = layer(x)
return self.norm_out(x)
# --- MAIN MODEL ---
class PillarsModel(PreTrainedModel):
config_class = PillarsConfig
def __init__(self, config):
super().__init__(config)
self.config = config
# 1. SHARED ROOT
self.rose = DynamicRoSE(config.vocab_size, config.d_model)
# 2. DOWNSAMPLE
self.particle_down = nn.Linear(config.d_model, config.d_branch)
self.wave_down = nn.Linear(config.d_model * 2, config.d_branch * 2)
# 3. RATE STREAM (FNet)
self.fnet_pos = nn.Embedding(config.seq_len, config.d_branch)
self.stream_rate = FNetEncoder(depth=config.depth, d_model=config.d_branch, d_ff=config.d_branch*4, dropout=config.dropout)
# 4. PHASE STREAM (PRISM)
self.stream_phase = PRISMEncoder(num_layers=config.depth, d_model=config.d_branch, max_len=config.seq_len, dropout=config.dropout)
self.phase_bridge = ComplexToRealBridge(config.d_branch)
# 5. FUSION
self.fusion_proj = nn.Linear(config.d_branch * 2, config.d_model)
self.fusion_norm = nn.LayerNorm(config.d_model)
# 6. REFINER
self.refiner = Encoder(
dim=config.d_model,
depth=config.refine_depth,
heads=8,
attn_flash=True,
rotary_pos_emb=True,
attn_dropout=config.dropout,
ff_dropout=config.dropout
)
# 7. HEAD
self.head_bias = nn.Parameter(torch.zeros(config.vocab_size))
def forward(self, input_ids, labels=None):
# A. Shared Root
wave_src, particle_src = self.rose(input_ids)
# B. Downsample
p_small = self.particle_down(particle_src)
w_flat = torch.cat([wave_src.real, wave_src.imag], dim=-1)
w_small_flat = self.wave_down(w_flat)
w_small = torch.complex(w_small_flat[..., :self.config.d_branch], w_small_flat[..., self.config.d_branch:])
# C. Branches
pos_emb = self.fnet_pos(torch.arange(input_ids.shape[1], device=input_ids.device))
rate_out = self.stream_rate(p_small + pos_emb)
phase_out = self.phase_bridge(self.stream_phase(w_small))
# D. Fusion
stacked = torch.cat([rate_out, phase_out], dim=-1)
context = self.fusion_norm(self.fusion_proj(stacked))
# E. Refiner & Output
refined = self.refiner(context)
# Weight tying: Use rose embeddings as output weights
logits = F.linear(refined, self.rose.raw_embedding.weight, self.head_bias)
loss = None
if labels is not None:
loss_fct = nn.CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))
return {"loss": loss, "logits": logits}
return logits
|