File size: 18,722 Bytes
d8c733f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 | import torch
import torch.nn.functional as F
import time
import numpy as np
# matplotlib è usato solo dalle utility di plot/benchmark, non dal codice del modello.
# Lazy + optional import per non rompere l'import di Vathos in ambienti headless.
try:
import matplotlib.pyplot as plt
plt.style.use('ggplot')
except ImportError:
plt = None # le funzioni di plot solleveranno errore esplicito se chiamate
try:
from colorama import Fore
GOOD = Fore.GREEN
BAD = Fore.RED
RES = Fore.RESET
SEC = Fore.LIGHTBLACK_EX
HM = Fore.YELLOW
NUM = Fore.BLUE
except:
GOOD = ""
BAD = ""
RES = ""
SEC = ""
NUM = ""
HM = ""
FLAG_PASS = 3
def _require_matplotlib():
if plt is None:
raise ImportError(
"Funzione di plotting chiamata ma matplotlib non è installato. "
"Installa con `pip install matplotlib` per usarla."
)
def plot(*args, **kwargs):
_require_matplotlib()
for x_i in args:
plt.plot(x_i)
for x_i in kwargs:
plt.plot(kwargs[x_i], label=str(x_i))
plt.legend()
plt.show()
def cumsum(x):
return torch.cumsum(x, dim=1)
def batched_channelwise_conv1d(K, V):
"""
Computes a convolution between two batched tensors
K: [B, L, d] kernel
V: [B, L, d] signal
Output: [B, L, d]
"""
B, L, d = V.shape
V_reshaped = V.view(1, B * L, d)
K_reshaped = K.view(B * L, 1, d).flip(-1)
pad_total = d - 1
pad_left = pad_total // 2
pad_right = pad_total - pad_left
V_padded = F.pad(V_reshaped, (pad_left, pad_right))
output = F.conv1d(V_padded, K_reshaped, groups=B * L)
return output.view(B, L, d)
def holographic_binding(K, V):
"""
Computes Batched Circular Convolution (Binding) using FFT.
This is O(d log d) and numerically stable.
K: [..., d] kernel
V: [..., d] signal
Output: [..., d]
"""
V_f = torch.fft.rfft(V, dim=-1, norm='ortho')
K_f = torch.fft.rfft(K, dim=-1, norm='ortho')
output_f = V_f * K_f
output = torch.fft.irfft(output_f, n=V.shape[-1], dim=-1, norm='ortho')
return output
def power_weigthed_cumsum(x, a=0.999, rescale=True):
"""
a torch implementation of a power weighted cumsum,
which can be applied to batched inputs of shapes [B, L, d] or [B, L, d1, d2], summing over L
"""
if a - 1 == 0:
return cumsum(x)
if len(x.shape) == 3:
alpha_pow = torch.full([x.shape[1]], a, dtype=x.dtype, device=x.device).cumprod(dim=0).unsqueeze(0).unsqueeze(2)
elif len(x.shape) == 4:
alpha_pow = torch.full([x.shape[1]], a, dtype=x.dtype, device=x.device).cumprod(dim=0).unsqueeze(0).unsqueeze(
2).unsqueeze(2)
else:
raise RuntimeError("x shapes mus be of form [B, L, d] or [B, L, d, d]")
return torch.cumsum(x / alpha_pow, dim=1) * alpha_pow * (alpha_pow[:, 0, ...] if rescale else 1)
def stable_power_weighted_cumsum(x: torch.Tensor, a: float = 0.999, rescale: bool = True) -> torch.Tensor:
B, L = x.shape[:2]
dtype = x.dtype
if L > 1024 and x.dtype == torch.float32:
x = x.to(torch.float64)
a_tensor = torch.tensor(a, dtype=torch.float64, device=x.device)
else:
a_tensor = torch.tensor(a, dtype=x.dtype, device=x.device)
t = torch.arange(1, L + 1, dtype=a_tensor.dtype, device=x.device)
alpha_pow = torch.pow(a_tensor, t)
reshape_dims = [1, L] + [1] * (len(x.shape) - 2)
alpha_pow = alpha_pow.view(*reshape_dims)
scaled_input = x / alpha_pow
s_cumulative = torch.cumsum(scaled_input, dim=1)
s = s_cumulative * alpha_pow
if rescale:
s = s * (1.0 - a_tensor)
return s.to(dtype)
def precompute_power_weigthed_cumsum(x, alpha_pow):
return torch.cumsum(x / alpha_pow, dim=1) * alpha_pow
def flag(text, level=1):
if level <= FLAG_PASS:
print(f"{SEC}||{HM}FLAG LV.{level}{SEC}||{HM} {text}{RES}")
def getname(obj):
return obj.__name__ if hasattr(obj, "__name__") else str(type(obj)).split('.')[-1]
def benchmark_ar_symbolic_model(model, n, input_shape):
"""
Benchmarks the autoregressive symbolic model on three tasks: forward pass,
forward + backward, and forward + backward + optimizer step. Uses random
generated data. Reports average times and shows a bar plot.
Additionally, benchmarks the ratio of forward time / (forward + backward) time
with increasing sequence length L, and plots it.
Args:
model: The model to benchmark (e.g., SequenceModel instance).
n: Number of iterations to average over for each benchmark.
input_shape: Tuple (batch_size, seq_len) for initial input shape.
"""
device = next(model.parameters()).device
is_cuda = device.type == 'cuda'
model.train() # Ensure model is in training mode for gradients
optimizer = torch.optim.Adam(model.parameters())
# Helper to sync if CUDA
def sync():
if is_cuda:
torch.cuda.synchronize()
# Part 1: Benchmark forward, forward+backward, forward+backward+opt
times_fwd = []
times_fb = []
times_fbo = []
for _ in range(n):
x = torch.randint(0, model.vocab_size, input_shape, device=device, requires_grad=False)
# Forward only
sync()
start = time.perf_counter()
out = model(x)
sync()
end = time.perf_counter()
times_fwd.append(end - start)
# Forward + backward
optimizer.zero_grad()
sync()
start = time.perf_counter()
out = model(x)
loss = out.sum() # Dummy loss
loss.backward()
sync()
end = time.perf_counter()
times_fb.append(end - start)
# Forward + backward + optimizer step
optimizer.zero_grad()
sync()
start = time.perf_counter()
out = model(x)
loss = out.sum()
loss.backward()
optimizer.step()
sync()
end = time.perf_counter()
times_fbo.append(end - start)
avg_fwd = np.mean(times_fwd)
avg_fb = np.mean(times_fb)
avg_fbo = np.mean(times_fbo)
print(f"Average Forward Time: {avg_fwd:.6f} s")
print(f"Average Forward + Backward Time: {avg_fb:.6f} s")
print(f"Average Forward + Backward + Opt Time: {avg_fbo:.6f} s")
# Bar plot for the three tasks
labels = ['Forward', 'Forward + Backward', 'Forward + Backward + Opt']
times = [avg_fwd, avg_fb, avg_fbo]
plt.figure(figsize=(8, 5))
plt.bar(labels, times, color=['blue', 'orange', 'green'])
plt.ylabel('Average Time (s)')
plt.title(f'Benchmark Averages over {n} Iterations')
plt.show()
# Part 2: Ratio of forward / (forward + backward) with increasing L
batch_size = input_shape[0]
Ls = [16, 32, 64, 128, 256, 512, 1024] # Reasonable sequence lengths, adjust as needed
ratios = []
avg_times_fwd = []
avg_times_fb = []
inner_n = max(10, n // 10) # Average over fewer iterations for speed, but at least 10
for L in Ls:
shape = (batch_size, L)
x = torch.randint(0, model.vocab_size, shape, device=device, requires_grad=False)
# Time forward (average over inner_n)
fwd_times = []
for _ in range(inner_n):
sync()
start = time.perf_counter()
out = model(x)
sync()
end = time.perf_counter()
fwd_times.append(end - start)
time_fwd = np.mean(fwd_times)
avg_times_fwd.append(time_fwd)
# Time forward + backward (average over inner_n)
fb_times = []
for _ in range(inner_n):
optimizer.zero_grad()
sync()
start = time.perf_counter()
out = model(x)
loss = out.sum()
loss.backward()
sync()
end = time.perf_counter()
fb_times.append(end - start)
time_fb = np.mean(fb_times)
avg_times_fb.append(time_fb)
ratio = time_fwd / time_fb if time_fb > 0 else 0
ratios.append(ratio)
# Plot the ratio
plt.figure(figsize=(8, 5))
plt.plot(Ls, ratios, marker='o', color='red')
plt.xlabel('Sequence Length L')
plt.ylabel('Time Forward / Time (Forward + Backward)')
plt.title('Efficiency Ratio vs Sequence Length')
plt.grid(True)
plt.show()
# Optional: Plot absolute times for reference
plt.figure(figsize=(8, 5))
plt.plot(Ls, avg_times_fwd, marker='o', label='Forward', color='blue')
plt.plot(Ls, avg_times_fb, marker='o', label='Forward + Backward', color='orange')
plt.xlabel('Sequence Length L')
plt.ylabel('Average Time (s)')
plt.title('Absolute Times vs Sequence Length')
plt.legend()
plt.grid(True)
plt.show()
def apply_repetition_penalty(logits: torch.Tensor, generated: torch.Tensor,
penalty: float) -> torch.Tensor:
"""
Standard HF-style repetition penalty, vectorized.
logits: [B, V] (last-position logits)
generated: [B, L] (token ids generated so far, including prompt)
Divides logits>0 by penalty, multiplies logits<=0 by penalty, in-place-safe.
"""
if penalty == 1.0:
return logits
score = torch.gather(logits, 1, generated)
score = torch.where(score > 0, score / penalty, score * penalty)
logits = logits.scatter(1, generated, score)
return logits
def sample_next_token(logits: torch.Tensor, temperature: float = 1.0,
top_k: int | None = None, top_p: float = 1.0) -> torch.Tensor:
"""
Sample one token per row from [B, V] logits with temperature, top-k, top-p.
Returns [B, 1] long tensor.
"""
logits = logits / max(temperature, 1e-8)
if top_k is not None and top_k > 0:
k = min(top_k, logits.size(-1))
v, _ = torch.topk(logits, k)
logits = logits.masked_fill(logits < v[:, [-1]], float('-inf'))
if top_p < 1.0:
sorted_logits, sorted_idx = torch.sort(logits, descending=True)
cum = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
remove = cum > top_p
remove[..., 1:] = remove[..., :-1].clone()
remove[..., 0] = 0
mask = remove.scatter(1, sorted_idx, remove)
logits = logits.masked_fill(mask, float('-inf'))
probs = F.softmax(logits, dim=-1)
return torch.multinomial(probs, num_samples=1)
@torch.no_grad()
def simple_ar_generate(model, prompt, max_len=100, temperature=1.0,
top_k=None, top_p=1.0, repetition_penalty=1.0,
token_end=None, verbose=False):
"""
Dumb autoregressive generation for debugging.
Runs a full forward pass on the whole sequence each step, takes logits at the
last position, samples, appends. O(L^2) total — useless for big models, but it
works on *anything* whose forward(ids) returns [B, L, V] logits, with no
assumptions about KV caches, custom generate(), pos encoders, ve, skips, etc.
Good as a ground-truth reference when debugging more sophisticated generators.
"""
flag("simple_ar_generate is O(L^2) and ignores KV caches: use it for debugging, "
"not for real inference. If your model has a fast .generate(), prefer that.", 2)
model.eval()
if prompt.dim() == 1:
prompt = prompt.unsqueeze(0)
device = next(model.parameters()).device
generated = prompt.clone().to(device)
for _ in range(max_len):
logits = model(generated)
next_logits = logits[:, -1, :]
if repetition_penalty != 1.0:
next_logits = apply_repetition_penalty(next_logits, generated, repetition_penalty)
next_token = sample_next_token(next_logits, temperature, top_k, top_p)
generated = torch.cat([generated, next_token], dim=1)
if verbose:
print(next_token.flatten().tolist())
if token_end is not None and (next_token == token_end).all():
break
return generated
def toeplitz_init(tensor: torch.Tensor, alpha: float, causal: bool = True, mul=0.1):
"""
Initializes a square 2D tensor to a custom Toeplitz matrix with decaying kernels
generated from alpha^t for t=1...L, where L is the side length of the matrix.
If causal is True, the matrix is lower triangular (tril), containing values only
where row index i >= column index j.
Args:
tensor (torch.Tensor): The square 2D tensor to initialize in place.
alpha (float): The base for the exponential decay.
causal (bool, optional): If True, makes the matrix lower triangular. Defaults to False.
Returns:
torch.Tensor: The initialized tensor (modified in place).
"""
if tensor.dim() != 2 or tensor.size(0) != tensor.size(1):
raise ValueError("Tensor must be a square 2D tensor.")
L = tensor.size(0)
i, j = torch.meshgrid(
torch.arange(L, device=tensor.device, dtype=tensor.dtype),
torch.arange(L, device=tensor.device, dtype=tensor.dtype),
indexing='ij'
)
dist = torch.abs(i - j)
t = dist + 1
matrix = alpha ** t
if causal:
matrix = torch.where(i >= j, matrix, torch.tensor(0.0, dtype=tensor.dtype, device=tensor.device))
with torch.no_grad():
tensor.copy_(matrix)
return tensor * mul
def zero_layer_mtp_loss( # TODO: Vibecoded for now
logits: torch.Tensor,
targets: torch.Tensor,
mtp_weights: list[float],
ignore_index: int = -100,
softcap_val: float = 30.0
) -> torch.Tensor:
"""
Production-grade 0-layer MTP loss.
Optimizations:
1. Computes LogSumExp exactly once per token.
2. Uses index gathering for target logits to minimize memory bandwidth.
3. Casts to FP32 for numerical stability before reduction.
"""
assert sum(mtp_weights) == 1, "the sum of MTP weights must be 1.0"
if softcap_val > 0.0:
logits = softcap_val * torch.tanh(logits / softcap_val)
lse = torch.logsumexp(logits, dim=-1).float()
total_loss = torch.tensor(0.0, device=logits.device, dtype=torch.float32)
for k, w in enumerate(mtp_weights):
if w <= 0.0:
continue
if k == 0:
valid_logits = logits
valid_targets = targets
valid_lse = lse
else:
valid_logits = logits[:, :-k]
valid_targets = targets[:, k:]
valid_lse = lse[:, :-k]
mask = (valid_targets != ignore_index)
safe_targets = torch.where(mask, valid_targets, torch.zeros_like(valid_targets))
target_logits = valid_logits.gather(
dim=-1,
index=safe_targets.unsqueeze(-1)
).squeeze(-1).float()
ce_loss = valid_lse - target_logits
masked_loss = ce_loss * mask
num_valid_tokens = mask.sum().clamp(min=1)
loss_k = masked_loss.sum() / num_valid_tokens
total_loss += w * loss_k
return total_loss
def plot_tensor_diagnostics(
name: str,
tensor: torch.Tensor,
on_grad: bool = False,
top_k_svd: int | None = None,
bins: int = 100
):
has_grad = on_grad and hasattr(tensor, 'grad') and tensor.grad is not None
rows = 2 if has_grad else 1
fig, axes = plt.subplots(rows, 2, figsize=(12, 5 * rows))
if rows == 1:
axes = np.expand_dims(axes, axis=0) # standardize indexing
def _compute_and_plot(t: torch.Tensor, row: int, title_prefix: str):
t_cpu = t.detach().float().cpu()
if t_cpu.ndim == 1:
mat = t_cpu.unsqueeze(0)
else:
mat = t_cpu.view(t_cpu.size(0), -1)
s_vals = torch.linalg.svdvals(mat).numpy()
if top_k_svd is not None:
s_vals = s_vals[:top_k_svd]
ax_svd = axes[row, 0]
ax_svd.plot(s_vals, marker='.', linestyle='-', color='b', markersize=4)
ax_svd.set_title(f"{title_prefix} Singular Values\nMax: {s_vals[0]:.4f}, Min: {s_vals[-1]:.4f}")
ax_svd.set_xlabel("Index")
ax_svd.set_ylabel("Singular Value $\sigma_i$")
ax_svd.set_yscale('log') # Log scale is crucial to see the long tail
ax_svd.grid(True, which="both", ls="--", alpha=0.5)
ax_hist = axes[row, 1]
vals_flat = t_cpu.numpy().flatten()
ax_hist.hist(vals_flat, bins=bins, color='g', alpha=0.7, log=True)
mean, std = vals_flat.mean(), vals_flat.std()
ax_hist.set_title(f"{title_prefix} Distribution\n$\mu$: {mean:.2e}, $\sigma$: {std:.2e}")
ax_hist.set_xlabel("Value")
ax_hist.set_ylabel("Count (Log Scale)")
ax_hist.grid(True, ls="--", alpha=0.5)
_compute_and_plot(tensor, row=0, title_prefix=f"Weight: {name}")
if has_grad:
_compute_and_plot(tensor.grad, row=1, title_prefix=f"Gradient: {name}")
elif on_grad:
print(f"Warning: on_grad=True for '{name}', but tensor.grad is None.")
plt.tight_layout()
plt.show()
import math
def compute_normalized_grad_norm(model: torch.nn.Module) -> float:
"""
Returns the RMS gradient norm across all trainable parameters.
RMS = sqrt( sum(g_i^2) / N ) where N = total number of scalar gradients.
Because this *averages* rather than sums, the result is independent of
model size: a 1 M-param and a 100 M-param model both live in [0, ~1]
during healthy training, making cross-run comparisons meaningful.
Returns 0.0 if no gradients are present yet.
"""
total_sq = 0.0
total_n = 0
for p in model.parameters():
if p.grad is not None:
total_sq += p.grad.detach().float().pow(2).sum().item()
total_n += p.grad.numel()
if total_n == 0:
return 0.0
return math.sqrt(total_sq / total_n)
def compute_layer_grad_norms(model: torch.nn.Module) -> dict[str, float]:
"""
Returns a dict { layer_name -> RMS grad norm } for every named module
that owns at least one parameter with a gradient.
Same RMS normalisation as compute_normalized_grad_norm so per-layer
values are still comparable across models of different sizes.
Leaf modules only (skips container modules to avoid double-counting).
"""
norms = {}
for name, module in model.named_modules():
# only leaf modules that directly own parameters
own_params = list(module.parameters(recurse=False))
if not own_params:
continue
sq_sum = 0.0
n = 0
for p in own_params:
if p.grad is not None:
sq_sum += p.grad.detach().float().pow(2).sum().item()
n += p.grad.numel()
if n > 0:
norms[name] = math.sqrt(sq_sum / n)
return norms |