File size: 18,021 Bytes
139362f 4d03589 139362f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 |
##############################################################################################################################################
#||||- - - |8.19.2025| - - - || LIQUID STATE SPACE || - - - |1990two| - - -|||| #
##############################################################################################################################################
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import math
from typing import List, Dict, Tuple, Optional
from scipy import linalg
from scipy.signal import cont2discrete
SAFE_MIN = -1e6
SAFE_MAX = 1e6
EPS = 1e-8
#||||- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 𓅸 - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -||||#
def make_safe(tensor, min_val=SAFE_MIN, max_val=SAFE_MAX):
zero = torch.tensor(0.0, device=tensor.device, dtype=tensor.dtype)
maxv = torch.tensor(max_val, device=tensor.device, dtype=tensor.dtype)
tensor = torch.where(torch.isnan(tensor), zero, tensor)
tensor = torch.where(torch.isinf(tensor), maxv, tensor)
return torch.clamp(tensor, min_val, max_val)
def discrete_to_continuous_time(A_discrete, dt=1.0):
try:
A_continuous = linalg.logm(A_discrete.detach().cpu().numpy()) / dt
return torch.tensor(A_continuous, dtype=torch.float32, device=A_discrete.device)
except:
return torch.eye(A_discrete.shape[0], device=A_discrete.device) * 0.01
def continuous_to_discrete_time(A_continuous, B_continuous, dt=1.0):
try:
A_np = A_continuous.detach().cpu().numpy()
B_np = B_continuous.detach().cpu().numpy()
if A_np.ndim == 3:
A_list, B_list = [], []
for i in range(A_np.shape[0]):
Ad, Bd, _, _, _ = cont2discrete((A_np[i], B_np, np.eye(A_np.shape[-1]), 0), dt)
A_list.append(Ad)
B_list.append(Bd)
A_discrete = torch.tensor(np.stack(A_list), dtype=torch.float32, device=A_continuous.device)
B_discrete = torch.tensor(np.stack(B_list), dtype=torch.float32, device=B_continuous.device)
else:
A_discrete, B_discrete, _, _, _ = cont2discrete((A_np, B_np, np.eye(A_np.shape[0]), 0), dt)
A_discrete = torch.tensor(A_discrete, dtype=torch.float32, device=A_continuous.device)
B_discrete = torch.tensor(B_discrete, dtype=torch.float32, device=B_continuous.device)
return A_discrete, B_discrete
except Exception:
n = A_continuous.shape[-1]
eye = torch.eye(n, device=A_continuous.device, dtype=A_continuous.dtype)
if A_continuous.dim() == 3:
eye = eye.unsqueeze(0).expand(A_continuous.size(0), -1, -1)
B_disc = B_continuous.to(dtype=A_continuous.dtype, device=A_continuous.device) \
.unsqueeze(0).expand(A_continuous.size(0), -1, -1)
else:
B_disc = B_continuous.to(dtype=A_continuous.dtype, device=A_continuous.device)
A_discrete = eye + A_continuous * dt
B_discrete = B_disc * dt
return A_discrete, B_discrete
###########################################################################################################################################
#############################################- - - LIQUID TIME CONSTANT CONTROLLER - - -###############################################
class LiquidTimeConstantController(nn.Module):
def __init__(self, state_dim, input_dim, init_tau=1.0):
super().__init__()
self.state_dim = state_dim
self.input_dim = input_dim
self.log_tau = nn.Parameter(torch.ones(state_dim) * math.log(init_tau))
self.tau_adaptation = nn.Sequential(
nn.Linear(state_dim + input_dim, state_dim * 2),
nn.LayerNorm(state_dim * 2),
nn.Tanh(),
nn.Linear(state_dim * 2, state_dim),
nn.Tanh() # Output in [-1, 1] for modulation
)
self.adaptation_rate = nn.Parameter(torch.tensor(0.1))
def get_time_constants(self, state, input_signal):
base_tau = torch.exp(self.log_tau)
base_tau = torch.clamp(base_tau, 0.01, 10.0)
combined_input = torch.cat([state, input_signal], dim=-1)
tau_modulation = self.tau_adaptation(combined_input)
adaptation_rate = torch.clamp(self.adaptation_rate, 0.001, 1.0)
modulated_tau = base_tau * (1.0 + adaptation_rate * tau_modulation)
return torch.clamp(modulated_tau, 0.01, 10.0)
def get_effective_dt(self, tau, target_dt=0.1):
min_tau_val = torch.min(tau).item()
effective_dt = max(0.001, min(float(target_dt), min_tau_val * 0.1))
return effective_dt
###########################################################################################################################################
################################################- - - LIQUID SSM CORE - - -############################################################
class LiquidSSMCore(nn.Module):
def __init__(self, state_dim, input_dim, output_dim, dt=0.1, init_method='hippo'):
super().__init__()
self.state_dim = state_dim
self.input_dim = input_dim
self.output_dim = output_dim
self.dt = dt
if init_method == 'hippo':
self.A_continuous = nn.Parameter(self._init_hippo_matrix(state_dim))
else:
self.A_continuous = nn.Parameter(torch.randn(state_dim, state_dim) * 0.1)
self.B_continuous = nn.Parameter(torch.randn(state_dim, input_dim) * 0.1)
self.C = nn.Parameter(torch.randn(output_dim, state_dim) * 0.1)
self.D = nn.Parameter(torch.zeros(output_dim, input_dim))
self.time_controller = LiquidTimeConstantController(state_dim, input_dim, init_tau=1.0)
self.output_scale = nn.Parameter(torch.ones(output_dim))
self.output_bias = nn.Parameter(torch.zeros(output_dim))
self.state_normalizer = nn.LayerNorm(state_dim)
self.register_buffer('continuous_state', torch.zeros(1, state_dim))
def _init_hippo_matrix(self, N):
A = torch.zeros(N, N)
for i in range(N):
for j in range(N):
if i > j:
A[i, j] = math.sqrt(2 * i + 1) * math.sqrt(2 * j + 1)
elif i == j:
A[i, j] = -(2 * i + 1)
A = A * 0.1
with torch.no_grad():
eig = torch.linalg.eigvals(A).real.abs().max()
if eig > 0:
A = A / eig * 0.9
return A
def reset_state(self, batch_size=1):
device = self.A_continuous.device
self.continuous_state = torch.zeros(batch_size, self.state_dim, device=device)
def liquid_state_evolution(self, input_signal, num_steps=10):
batch_size = input_signal.shape[0]
if self.continuous_state.shape[0] != batch_size:
self.reset_state(batch_size)
tau = self.time_controller.get_time_constants(self.continuous_state, input_signal)
effective_dt = self.time_controller.get_effective_dt(tau, self.dt)
tau_matrix = torch.diag_embed(1.0 / tau)
liquid_A = self.A_continuous - tau_matrix
liquid_A = make_safe(liquid_A, min_val=-10.0, max_val=10.0)
A_discrete, B_discrete = continuous_to_discrete_time(
liquid_A, self.B_continuous, effective_dt
)
step_dt = float(effective_dt) / num_steps
A_discrete, B_discrete = continuous_to_discrete_time(
liquid_A, self.B_continuous, step_dt
)
current_state = self.continuous_state
if A_discrete.dim() == 3:
A_T = A_discrete.transpose(1, 2)
B_T = B_discrete.transpose(1, 2)
input_update = torch.bmm(input_signal.unsqueeze(1), B_T).squeeze(1)
for _ in range(num_steps):
state_update = torch.bmm(current_state.unsqueeze(1), A_T).squeeze(1)
current_state = state_update + input_update
current_state = make_safe(current_state)
else:
A_T = A_discrete.T
B_T = B_discrete.T
input_update = input_signal @ B_T
for _ in range(num_steps):
current_state = current_state @ A_T + input_update
current_state = make_safe(current_state)
current_state = make_safe(current_state)
self.continuous_state = current_state
return current_state, tau, effective_dt
def compute_output(self, state, input_signal):
normalized_state = self.state_normalizer(state)
state_output = torch.matmul(normalized_state, self.C.T)
direct_output = torch.matmul(input_signal, self.D.T)
raw_output = state_output + direct_output
output = self.output_scale * raw_output + self.output_bias
return make_safe(output)
def forward(self, input_signal, return_diagnostics=False):
evolved_state, tau, effective_dt = self.liquid_state_evolution(input_signal)
output = self.compute_output(evolved_state, input_signal)
result = {
'output': output,
'state': evolved_state
}
if return_diagnostics:
result.update({
'time_constants': tau,
'effective_dt': effective_dt,
'state_norm': torch.norm(evolved_state, dim=-1),
'adaptation_rate': self.time_controller.adaptation_rate
})
return result
###########################################################################################################################################
############################################- - - LIQUID SSM SEQUENCE LAYER - - -######################################################
class LiquidSSMSequenceLayer(nn.Module):
def __init__(self, input_dim, state_dim, output_dim, seq_len=None):
super().__init__()
self.input_dim = input_dim
self.state_dim = state_dim
self.output_dim = output_dim
self.seq_len = seq_len
self.liquid_ssm = LiquidSSMCore(state_dim, state_dim, output_dim)
self.input_projection = nn.Sequential(
nn.Linear(input_dim, state_dim),
nn.LayerNorm(state_dim),
nn.GELU()
)
self.output_projection = nn.Sequential(
nn.Linear(output_dim, output_dim * 2),
nn.LayerNorm(output_dim * 2),
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(output_dim * 2, output_dim)
)
self.residual_weight = nn.Parameter(torch.tensor(0.1))
self.sequence_adapter = nn.Sequential(
nn.Linear(state_dim, state_dim),
nn.Tanh(),
nn.Linear(state_dim, 1),
nn.Sigmoid()
)
def forward(self, sequence, return_diagnostics=False):
batch_size, seq_len, input_dim = sequence.shape
self.liquid_ssm.reset_state(batch_size)
outputs = []
diagnostics = [] if return_diagnostics else None
for t in range(seq_len):
current_input = sequence[:, t, :]
projected_input = self.input_projection(current_input)
ssm_result = self.liquid_ssm(projected_input, return_diagnostics=return_diagnostics)
adaptation_factor = self.sequence_adapter(ssm_result['state'])
adapted_output = ssm_result['output'] * adaptation_factor
final_output = self.output_projection(adapted_output)
if final_output.shape == current_input.shape:
residual_strength = torch.clamp(self.residual_weight, 0.0, 1.0)
final_output = final_output + residual_strength * current_input
outputs.append(final_output)
if return_diagnostics:
diagnostics.append({
'timestep': t,
'adaptation_factor': adaptation_factor.mean().item(),
**ssm_result
})
output_sequence = torch.stack(outputs, dim=1)
result = {'output': output_sequence}
if return_diagnostics:
result['diagnostics'] = diagnostics
return result
###########################################################################################################################################
###########################################- - - LIQUID SSM LANGUAGE MODEL - - -#######################################################
class LiquidSSMLanguageModel(nn.Module):
def __init__(self, vocab_size, d_model=512, state_dim=256, num_layers=6, max_seq_len=2048):
super().__init__()
self.vocab_size = vocab_size
self.d_model = d_model
self.state_dim = state_dim
self.num_layers = num_layers
self.max_seq_len = max_seq_len
self.token_embedding = nn.Embedding(vocab_size, d_model)
self.position_embedding = nn.Embedding(max_seq_len, d_model)
self.liquid_layers = nn.ModuleList([
LiquidSSMSequenceLayer(d_model, state_dim, d_model)
for _ in range(num_layers)
])
self.layer_norms = nn.ModuleList([
nn.LayerNorm(d_model) for _ in range(num_layers)
])
self.output_norm = nn.LayerNorm(d_model)
self.lm_head = nn.Linear(d_model, vocab_size)
self.global_adaptation = nn.Sequential(
nn.Linear(d_model, d_model // 4),
nn.GELU(),
nn.Linear(d_model // 4, 1),
nn.Sigmoid()
)
self._init_weights()
def _init_weights(self):
for module in self.modules():
if isinstance(module, nn.Linear):
nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(self, input_ids, labels=None, return_diagnostics=False):
batch_size, seq_len = input_ids.shape
device = input_ids.device
if seq_len > self.max_seq_len:
input_ids = input_ids[:, :self.max_seq_len]
seq_len = self.max_seq_len
if labels is not None:
labels = labels[:, :self.max_seq_len]
input_ids = torch.clamp(input_ids, 0, self.vocab_size - 1)
token_emb = self.token_embedding(input_ids)
pos_ids = torch.arange(seq_len, device=device).unsqueeze(0).expand(batch_size, -1)
pos_emb = self.position_embedding(pos_ids)
x = token_emb + pos_emb
x = make_safe(x)
layer_diagnostics = [] if return_diagnostics else None
for layer_idx, (liquid_layer, layer_norm) in enumerate(zip(self.liquid_layers, self.layer_norms)):
residual = x
x = layer_norm(x)
layer_result = liquid_layer(x, return_diagnostics=return_diagnostics)
x = layer_result['output']
adaptation = self.global_adaptation(x.mean(dim=1, keepdim=True))
x = x * adaptation
x = residual + x
x = make_safe(x)
if return_diagnostics:
layer_diagnostics.append({
'layer': layer_idx,
'adaptation': adaptation.mean().item(),
**layer_result
})
x = self.output_norm(x)
logits = self.lm_head(x)
logits = make_safe(logits, min_val=-50, max_val=50)
loss = None
if labels is not None:
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, self.vocab_size),
shift_labels.view(-1),
ignore_index=-100
)
result = {
'logits': logits,
'loss': loss
}
if return_diagnostics:
result['layer_diagnostics'] = layer_diagnostics
return result
@torch.no_grad()
def generate(self, input_ids, max_length=100, temperature=1.0, top_p=0.95, return_diagnostics=False):
self.eval()
generated = input_ids.clone()
all_diagnostics = [] if return_diagnostics else None
for step in range(max_length - input_ids.shape[1]):
if generated.shape[1] > self.max_seq_len:
break
outputs = self(generated, return_diagnostics=return_diagnostics)
logits = outputs['logits']
if return_diagnostics:
all_diagnostics.append(outputs.get('layer_diagnostics', []))
next_token_logits = logits[:, -1, :] / max(temperature, EPS)
next_token_logits = make_safe(next_token_logits, min_val=-50, max_val=50)
sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
sorted_indices_to_remove = cumulative_probs > top_p
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = False
for b in range(next_token_logits.size(0)):
indices_to_remove = sorted_indices[b][sorted_indices_to_remove[b]]
next_token_logits[b, indices_to_remove] = -float('inf')
probs = F.softmax(next_token_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
next_token = torch.clamp(next_token, 0, self.vocab_size - 1)
generated = torch.cat([generated, next_token], dim=1)
if next_token.item() == 2: # EOS token
break
result = {'generated_ids': generated}
if return_diagnostics:
result['diagnostics'] = all_diagnostics
return result
|