Spaces:
Sleeping
Sleeping
File size: 20,813 Bytes
3da2703 | 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 | """
Node 7: Weight Signal Extractor — Sub-env 3.
Extracts canonical weight statistics from a trained LoRA ``.safetensors`` file.
**All statistics are computed from canonical SVD components (U, S, Vt, Q)**
produced by ``canonicalize_lora_factors()`` imported from
``src/utils/canonical.py``. Raw A/B matrices are never used for statistics
directly — they are fed through the W2T QR → SVD pipeline first, resolving
column-space factorization ambiguity.
``token_position_to_phoneme`` is loaded from the tokenizer config JSON shipped
alongside the ``.safetensors`` file. It is NOT derived from the weights. If
``tokenizer_config_path`` is ``None``, the field is set to ``None``.
Statistic formulas:
- Layer norms: ``torch.linalg.norm(canonical_effective_update)``
- Layer entropy: inverted normalized entropy of canonical singular values
- Rank utilization: from canonical SVD S directly (no extra computation)
"""
from __future__ import annotations
import json
from pathlib import Path
from typing import Optional
import numpy as np
import torch
from safetensors.torch import load_file
from src.schemas.subenv2 import SyntheticWeightDescriptor
from src.schemas.subenv3 import WeightSignalObservation
from src.utils.canonical import (
CanonicalComponents,
canonicalize_lora_factors,
layer_entropy_from_singular_values,
singular_direction_anomaly_scores,
)
# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------
# Fraction of a layer's max singular value below which an S entry is "near zero".
_SPARSITY_THRESHOLD_RATIO: float = 0.01
# Cumulative energy fraction for counting "dominant" singular directions.
_DOMINANT_ENERGY_FRACTION: float = 0.90
# Standard-deviation multipliers for flagging anomalous Vt-row entropy.
_ENTROPY_ANOMALY_SIGMA: float = 2.0
# Number of histogram bins for the canonical S distribution.
_HISTOGRAM_BINS: int = 20
# ---------------------------------------------------------------------------
# Private helpers — safetensors parsing
# ---------------------------------------------------------------------------
def _find_lora_pairs(state_dict: dict[str, torch.Tensor]) -> dict[str, tuple[torch.Tensor, torch.Tensor]]:
"""Parse a LoRA state-dict into ``{layer_name: (A, B)}`` pairs.
Supports two common key conventions:
- PEFT-style: ``...lora_A...weight`` / ``...lora_B...weight``
- Kohya-style: ``...lora_down.weight`` / ``...lora_up.weight``
``A`` is the "down" (input) projection; ``B`` is the "up" (output)
projection, matching the LoRA convention ``delta_W = B @ A``.
Keys that are not 2-D tensors are silently skipped (e.g. alpha scalars).
"""
a_keys: dict[str, str] = {}
b_keys: dict[str, str] = {}
for key in state_dict:
if state_dict[key].ndim != 2:
continue
lo = key.lower()
if "lora_a" in lo or "lora_down" in lo:
# Derive canonical layer name by stripping the LoRA suffix
base = (
key.replace(".lora_A.weight", "")
.replace(".lora_A.default.weight", "")
.replace(".lora_down.weight", "")
)
a_keys[base] = key
elif "lora_b" in lo or "lora_up" in lo:
base = (
key.replace(".lora_B.weight", "")
.replace(".lora_B.default.weight", "")
.replace(".lora_up.weight", "")
)
b_keys[base] = key
pairs: dict[str, tuple[torch.Tensor, torch.Tensor]] = {}
for base in a_keys:
if base in b_keys:
A = state_dict[a_keys[base]] # shape: (rank, in_features) → transpose → (in_features, rank)
B = state_dict[b_keys[base]] # shape: (out_features, rank)
# PEFT stores A as (rank, in), B as (out, rank).
# canonicalize_lora_factors expects A: (in, rank), B: (out, rank).
if A.shape[0] < A.shape[1]:
A = A.T
pairs[base] = (A, B)
return pairs
# ---------------------------------------------------------------------------
# Private helpers — per-layer statistics (all from canonical components)
# ---------------------------------------------------------------------------
def _layer_norm_from_canonical(cc: CanonicalComponents) -> float:
"""Frobenius norm of the canonical effective update.
Equivalent to ``torch.linalg.norm(U @ diag(S) @ Vt)``, which simplifies
to ``torch.linalg.norm(S)`` since U and Vt are orthonormal.
"""
return float(torch.linalg.norm(cc.S))
def _layer_sparsity_from_canonical(cc: CanonicalComponents) -> float:
"""Fraction of canonical singular values near zero.
A value is «near zero» when it is below
``_SPARSITY_THRESHOLD_RATIO × max(S)``.
"""
s = cc.S
if s.numel() == 0:
return 1.0
threshold = _SPARSITY_THRESHOLD_RATIO * float(s.max())
near_zero = int((s < threshold).sum())
return near_zero / s.numel()
def _layer_rank_utilization_from_canonical(cc: CanonicalComponents) -> float:
"""Effective rank / nominal rank from canonical SVD S.
Effective rank = number of singular values above the sparsity threshold.
Nominal rank = total number of singular values (= LoRA rank).
"""
s = cc.S
if s.numel() == 0:
return 0.0
threshold = _SPARSITY_THRESHOLD_RATIO * float(s.max())
effective_rank = int((s >= threshold).sum())
return effective_rank / s.numel()
def _layer_entropy_from_canonical(cc: CanonicalComponents) -> float:
"""Inverted normalized entropy of canonical singular values.
Returns a value in [0, 1] where:
- 0.0 indicates uniform singular values (healthy utilization),
- 1.0 indicates concentrated energy (rank collapse tendency).
"""
return layer_entropy_from_singular_values(cc.S)
def _per_row_entropies(cc: CanonicalComponents) -> list[float]:
"""Per-direction anomaly scores from normalized singular values.
Values are in [0, 1]; higher values indicate singular directions that
dominate layer energy.
"""
return singular_direction_anomaly_scores(cc.S)
def _u_column_norm_variance(cc: CanonicalComponents) -> float:
"""Variance of U column norms across the canonical left singular vectors."""
u = cc.U # (out_features, rank)
if u.numel() == 0:
return 0.0
col_norms = torch.linalg.norm(u, dim=0) # (rank,)
return float(col_norms.var())
def _dominant_directions(cc: CanonicalComponents, fraction: float = _DOMINANT_ENERGY_FRACTION) -> int:
"""Number of canonical singular values that capture ``fraction`` of total energy."""
s = cc.S
if s.numel() == 0:
return 0
energy = s ** 2
total = float(energy.sum())
if total < 1e-12:
return 0
cumsum = torch.cumsum(energy, dim=0)
k = int((cumsum < fraction * total).sum()) + 1
return min(k, s.numel())
# ---------------------------------------------------------------------------
# Private helpers — cross-layer statistics
# ---------------------------------------------------------------------------
def _layer_update_vector(cc: CanonicalComponents) -> np.ndarray:
"""Flatten the canonical effective update to a 1-D vector for correlation."""
# Effective update = U @ diag(S) @ Vt; we use S as a compact representation.
return cc.S.detach().cpu().float().numpy()
def _correlation_matrix(vectors: list[np.ndarray]) -> list[list[float]]:
"""Pairwise Pearson correlation matrix of layer canonical S vectors.
Vectors are zero-padded or truncated to the shortest common length before
correlation is computed, since different layers may have different ranks.
"""
if not vectors:
return []
min_len = min(len(v) for v in vectors)
mat = np.stack([v[:min_len] for v in vectors], axis=0) # (n_layers, min_len)
try:
corr = np.corrcoef(mat) # (n_layers, n_layers)
except Exception:
n = len(vectors)
corr = np.eye(n)
return [[float(corr[i, j]) for j in range(corr.shape[1])] for i in range(corr.shape[0])]
def _attention_head_specialization(
layer_pairs: dict[str, tuple[CanonicalComponents, int]],
) -> dict[str, float]:
"""Per-layer attention head specialization score.
For each attention-related layer, measures how much the canonical energy
is concentrated in a subset of singular directions relative to the number
of heads. A score near 1.0 indicates one head dominates; near 0.0
indicates uniform distribution across heads.
Args:
layer_pairs: ``{layer_name: (CanonicalComponents, n_heads)}`` where
``n_heads`` is inferred from the layer name (see implementation).
Returns:
``{layer_name: specialization_score}`` for attention layers only.
"""
result: dict[str, float] = {}
for name, (cc, n_heads) in layer_pairs.items():
if n_heads < 2:
continue
s = cc.S.detach().cpu().float().numpy()
if len(s) == 0:
continue
energy = s ** 2
total = energy.sum()
if total < 1e-12:
continue
# Fraction of energy in the top (rank // n_heads) directions
head_size = max(1, len(s) // n_heads)
top_energy = energy[:head_size].sum()
result[name] = float(top_energy / total)
return result
def _weight_magnitude_histogram(all_s_values: list[float], n_bins: int = _HISTOGRAM_BINS) -> list[float]:
"""Binned histogram of all canonical singular values (S) across layers."""
if not all_s_values:
return [0.0] * n_bins
arr = np.array(all_s_values, dtype=np.float32)
counts, _ = np.histogram(arr, bins=n_bins)
total = counts.sum()
return [float(c / total) if total > 0 else 0.0 for c in counts]
def _gradient_noise_estimate(layer_canonical: dict[str, CanonicalComponents]) -> float:
"""Estimate gradient noise from the flatness of the canonical S spectrum.
A noisy training signal causes a more uniform singular value spectrum (high
ratio of minimum to maximum S). Returns a value in [0.0, 1.0] where 1.0
is maximum estimated noise.
"""
ratios: list[float] = []
for cc in layer_canonical.values():
s = cc.S.detach().cpu().float().numpy()
if len(s) < 2 or s[0] < 1e-8:
continue
ratios.append(float(s[-1] / s[0])) # min / max (S is descending from SVD)
return float(np.mean(ratios)) if ratios else 0.0
def _overfitting_signature(
rank_utilizations: dict[str, float],
layer_norms: dict[str, float],
) -> float:
"""Overfitting signature: low rank utilization combined with high norms.
Returns a value in [0.0, 1.0] where 1.0 indicates a strong overfitting
pattern (collapsed canonical directions + inflated update norms).
"""
if not rank_utilizations:
return 0.0
mean_util = float(np.mean(list(rank_utilizations.values())))
mean_norm = float(np.mean(list(layer_norms.values()))) if layer_norms else 0.0
# Low utilization (near 0) and high norm (uncapped) → high overfitting score.
# Norm is normalised with a soft cap at 10.0.
norm_factor = float(np.clip(mean_norm / 10.0, 0.0, 1.0))
return float(np.clip((1.0 - mean_util) * norm_factor, 0.0, 1.0))
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def _infer_n_heads(layer_name: str) -> int:
"""Heuristically infer the number of attention heads from a layer name."""
# Common naming patterns: attn.q_proj, self_attn.k_proj, attention.query
_ATTN_KEYWORDS = ("q_proj", "k_proj", "v_proj", "query", "key", "value", "attn")
lo = layer_name.lower()
if not any(kw in lo for kw in _ATTN_KEYWORDS):
return 0 # not an attention layer — signal to skip head specialization
# Return a reasonable default; could be overridden via config in production.
return 8
def extract_weight_signals(
weight_path: Path,
tokenizer_config_path: Optional[Path] = None,
dataset_health_summary: Optional[SyntheticWeightDescriptor] = None,
suspected_anomalous_phonemes: Optional[list[str]] = None,
) -> WeightSignalObservation:
"""Extract canonical weight statistics from a LoRA ``.safetensors`` file.
``canonicalize_lora_factors()`` is called for every LoRA (A, B) pair
before any statistics are computed. All ``canonical_*`` fields in the
returned observation are derived exclusively from the canonical SVD
components (U, S, Vt, Q), never from raw A/B matrices.
Args:
weight_path: Path to the ``.safetensors`` LoRA weight file.
tokenizer_config_path: Path to the audio tokenizer config JSON that
ships alongside the ``.safetensors`` file. Must contain a
``"token_position_to_phoneme"`` key mapping int positions to
phoneme strings. If ``None``, the field is set to ``None``.
dataset_health_summary: Optional ``SyntheticWeightDescriptor`` from
Sub-env 2 Node 6, forwarded as prior context.
suspected_anomalous_phonemes: Optional list of phonemes flagged by
Sub-env 2, forwarded as prior context.
Returns:
A fully populated :class:`WeightSignalObservation`.
Raises:
FileNotFoundError: If ``weight_path`` does not exist.
ValueError: If no valid LoRA (A, B) pairs are found in the file.
"""
weight_path = Path(weight_path)
if not weight_path.exists():
raise FileNotFoundError(f"Weight file not found: {weight_path}")
weight_file_id = weight_path.name
# ------------------------------------------------------------------
# Load state dict and parse LoRA pairs
# ------------------------------------------------------------------
state_dict: dict[str, torch.Tensor] = load_file(str(weight_path))
lora_pairs = _find_lora_pairs(state_dict)
if not lora_pairs:
raise ValueError(
f"No valid LoRA (A, B) pairs found in '{weight_path}'. "
"Check that the file contains lora_A/lora_B or lora_down/lora_up keys."
)
target_modules = sorted(lora_pairs.keys())
# Infer rank from first pair
first_A, _ = next(iter(lora_pairs.values()))
lora_rank = first_A.shape[1] # A is (in_features, rank)
total_parameters = sum(
A.numel() + B.numel() for A, B in lora_pairs.values()
)
# ------------------------------------------------------------------
# Canonicalize every (A, B) pair — all subsequent stats use cc only
# ------------------------------------------------------------------
layer_canonical: dict[str, CanonicalComponents] = {}
for name, (A, B) in lora_pairs.items():
layer_canonical[name] = canonicalize_lora_factors(A.float(), B.float())
# ------------------------------------------------------------------
# Layer-wise statistics (all from canonical components)
# ------------------------------------------------------------------
layer_norms: dict[str, float] = {}
layer_sparsity: dict[str, float] = {}
layer_rank_utilization: dict[str, float] = {}
canonical_entropy_per_layer: dict[str, float] = {}
# Collect all S values for histogram and per-direction anomaly scores
all_s_values: list[float] = []
all_row_entropies: list[tuple[int, float]] = [] # (position_idx, anomaly_score)
for name, cc in layer_canonical.items():
layer_norms[name] = _layer_norm_from_canonical(cc)
layer_sparsity[name] = _layer_sparsity_from_canonical(cc)
layer_rank_utilization[name] = _layer_rank_utilization_from_canonical(cc)
canonical_entropy_per_layer[name] = _layer_entropy_from_canonical(cc)
all_s_values.extend(cc.S.detach().cpu().float().tolist())
for row_idx, h in enumerate(_per_row_entropies(cc)):
all_row_entropies.append((row_idx, h))
# ------------------------------------------------------------------
# High-entropy token positions (anomalous singular directions across layers)
# ------------------------------------------------------------------
if all_row_entropies:
entropies_only = np.array([e for _, e in all_row_entropies], dtype=np.float32)
mu = float(entropies_only.mean())
sigma = float(entropies_only.std())
threshold = max(mu + _ENTROPY_ANOMALY_SIGMA * sigma, 0.1)
high_entropy_token_positions: list[int] = sorted(
set(
pos
for pos, h in all_row_entropies
if h > threshold
)
)
else:
high_entropy_token_positions = []
# ------------------------------------------------------------------
# Token-to-phoneme mapping (from tokenizer config, NOT from weights)
# ------------------------------------------------------------------
token_position_to_phoneme: Optional[dict[int, str]] = None
if tokenizer_config_path is not None:
tokenizer_config_path = Path(tokenizer_config_path)
with tokenizer_config_path.open("r", encoding="utf-8") as fh:
tok_cfg = json.load(fh)
raw_map: dict = tok_cfg.get("token_position_to_phoneme", {})
token_position_to_phoneme = {int(k): str(v) for k, v in raw_map.items()}
# ------------------------------------------------------------------
# Canonical U-component statistics (aggregated across all layers)
# ------------------------------------------------------------------
u_norm_variances = [_u_column_norm_variance(cc) for cc in layer_canonical.values()]
canonical_output_norm_variance = float(np.mean(u_norm_variances)) if u_norm_variances else 0.0
dominant_per_layer = [_dominant_directions(cc) for cc in layer_canonical.values()]
canonical_dominant_directions = int(np.mean(dominant_per_layer)) if dominant_per_layer else 0
# ------------------------------------------------------------------
# Cross-layer correlation matrix (over canonical S vectors)
# ------------------------------------------------------------------
layer_vectors = [_layer_update_vector(cc) for cc in layer_canonical.values()]
layer_correlation_matrix = _correlation_matrix(layer_vectors)
# ------------------------------------------------------------------
# Attention head specialization
# ------------------------------------------------------------------
attn_pairs: dict[str, tuple[CanonicalComponents, int]] = {
name: (cc, _infer_n_heads(name))
for name, cc in layer_canonical.items()
if _infer_n_heads(name) >= 2
}
attention_head_specialization = _attention_head_specialization(attn_pairs)
# ------------------------------------------------------------------
# Training quality signals
# ------------------------------------------------------------------
weight_magnitude_histogram = _weight_magnitude_histogram(all_s_values)
gradient_noise_estimate = _gradient_noise_estimate(layer_canonical)
overfitting_sig = _overfitting_signature(layer_rank_utilization, layer_norms)
# ------------------------------------------------------------------
# Assemble observation
# ------------------------------------------------------------------
return WeightSignalObservation(
weight_file_id=weight_file_id,
lora_rank=lora_rank,
target_modules=target_modules,
total_parameters=total_parameters,
# Layer-wise (canonical)
layer_norms=layer_norms,
layer_sparsity=layer_sparsity,
layer_rank_utilization=layer_rank_utilization,
# Canonical Vt analysis
canonical_entropy_per_layer=canonical_entropy_per_layer,
high_entropy_token_positions=high_entropy_token_positions,
# Token-to-phoneme (from tokenizer config only)
token_position_to_phoneme=token_position_to_phoneme,
# Canonical U analysis
canonical_output_norm_variance=canonical_output_norm_variance,
canonical_dominant_directions=canonical_dominant_directions,
# Cross-layer patterns
layer_correlation_matrix=layer_correlation_matrix,
attention_head_specialization=attention_head_specialization,
# Training quality
weight_magnitude_histogram=weight_magnitude_histogram,
gradient_noise_estimate=gradient_noise_estimate,
overfitting_signature=overfitting_sig,
# Sub-env 2 context
dataset_health_summary=dataset_health_summary,
suspected_anomalous_phonemes=suspected_anomalous_phonemes,
)
|