Spaces:
Running
Running
File size: 19,802 Bytes
535348a 1a6c909 535348a fed77dc 535348a fed77dc 535348a fed77dc 535348a fed77dc 535348a fed77dc 535348a fed77dc 535348a fed77dc 535348a 1a6c909 535348a | 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 | """
diagnose_model.py — "Predicting How Transformers Attend" Diagnostic Tool
=========================================================================
Single-command characterization of any causal LM via power-law attention decay.
Measures:
γ (gamma) — attention decay exponent A(d) ∝ d^{-γ}
T_attn = 1/γ — attention temperature
Phase — A (deconfined / RoPE), B (confined / AbsPE), C (ALiBi), Hagedorn
Z, U, S, F — thermodynamic potentials (partition function, energy, entropy, free energy)
C_V, χ — heat capacity, susceptibility
D_90 — context depth capturing 90% of Z (KV compression estimate)
ΔH_90 — holographic quality loss at D_90
KL_grammar — attention grammar anomaly (deviation from power-law prior)
θ_eff — effective RoPE base (Padé diagnostic)
γ_pred — theoretical prediction C/ln(θ) where C=ln(10000)=9.2103
Usage:
python diagnose_model.py --model EleutherAI/pythia-70m
python diagnose_model.py --model meta-llama/Meta-Llama-3-8B --local /path/to/weights --load_in_4bit
python diagnose_model.py --model Qwen/Qwen2.5-7B --theta 1000000 --N 1000
python diagnose_model.py --model EleutherAI/pythia-70m --fast # quick mode, 3 distances
Output:
Prints diagnostic table to stdout.
Saves JSON to ./diagnose_results/{model_short}.json
"""
import sys
sys.stdout.reconfigure(encoding='utf-8')
import argparse
import json
import math
import random
import time
from pathlib import Path
import numpy as np
# ── Constants ──────────────────────────────────────────────────────────────────
C_THEORY = math.log(10000) # 9.2103 — γ × ln(θ) = C for standard RoPE
DISTANCES_FULL = [10, 20, 30, 50, 100, 200, 500, 1000, 2000]
DISTANCES_FAST = [10, 50, 200, 1000]
N_PROMPTS = 30 # per distance (fast mode default)
N_PROMPTS_FULL = 80
SEEDS = [42, 123, 7]
THETA_KNOWN = {
"EleutherAI/pythia-14m": 10_000,
"EleutherAI/pythia-31m": 10_000,
"EleutherAI/pythia-70m": 10_000,
"EleutherAI/pythia-160m": 10_000,
"EleutherAI/pythia-410m": 10_000,
"EleutherAI/pythia-1b": 10_000,
"EleutherAI/pythia-1.4b": 10_000,
"EleutherAI/pythia-2.8b": 10_000,
"mistralai/Mistral-7B-v0.1": 10_000,
"tiiuae/falcon-7b": 10_000,
"microsoft/phi-2": 10_000,
"meta-llama/Llama-2-7b-hf": 10_000,
"google/gemma-2-9b-it": 10_000,
"EleutherAI/gpt-j-6B": 10_000,
"meta-llama/Meta-Llama-3-8B": 500_000,
"Qwen/Qwen2.5-7B": 1_000_000,
"mistralai/Mistral-Nemo-Instruct-2407": 1_000_000,
"codellama/CodeLlama-13b-Instruct-hf": 1_000_000,
}
OUTPUT_DIR = Path("./diagnose_results")
# ── Thermodynamic functions ────────────────────────────────────────────────────
# Euler-Mascheroni constant — needed for accurate H_N approximation at γ=1.
EULER_GAMMA = 0.5772156649015329
def partition_Z(gamma: float, N: int) -> float:
"""Z(γ, N) = sum_{d=1}^N d^{-γ}.
γ=1: H_N ~ log N + γ_E + 1/(2N) − ... [Euler-Mascheroni asymptotic]
γ≠1: integral approximation + d=1 boundary.
"""
if abs(gamma - 1.0) < 1e-5:
return math.log(N) + EULER_GAMMA # was math.log(N+0.5), missing γ_E
return (N ** (1 - gamma) - 1) / (1 - gamma) + 1
def mean_log_d(gamma: float, N: int) -> float:
Z = partition_Z(gamma, N)
if Z <= 0:
return 0.0
if abs(gamma - 1.0) < 1e-5:
integral = math.log(N) ** 2 / 2
else:
g1 = 1.0 - gamma
integral = N ** g1 * (math.log(N) / g1 - 1 / g1 ** 2) + 1 / g1 ** 2
return integral / Z
def entropy_S(gamma: float, N: int) -> float:
return math.log(partition_Z(gamma, N)) + gamma * mean_log_d(gamma, N)
def free_energy_F(gamma: float, N: int) -> float:
"""Helmholtz free energy: F = -T·log(Z) = -log(Z)/γ (T_attn = 1/γ).
Was: -log(Z) [β·F = log-partition convention; ambiguous when reported as F].
Now: -log(Z)/γ [physical F, consistent with U = -∂(log Z)/∂γ and S = (U − F)/T].
"""
Z = max(partition_Z(gamma, N), 1e-30)
return -math.log(Z) / max(gamma, 1e-9)
def heat_capacity_Cv(gamma: float, N: int, delta: float = 1e-4) -> float:
if gamma <= delta or gamma >= 20:
return float("nan")
dU = (mean_log_d(gamma + delta, N) - mean_log_d(gamma - delta, N)) / (2 * delta)
return -gamma ** 2 * dU
def D_f_closed(gamma: float, f: float, N: int) -> int:
"""KV compression window — DISCRETE truth (exact for the sum).
Smallest D such that ∑_{d=1}^D d^{-γ} / ∑_{d=1}^N d^{-γ} ≥ f.
The paper's "exact continuous formula"
D_f = [(1−f) + f·N^(1−γ)]^{1/(1−γ)} (and the γ=1 limit N^f)
is a CONTINUUM INTEGRAL APPROXIMATION that diverges from the discrete
sum by 5–50% in Phase B (γ>1), where the agent serves users.
Since N is bounded by context window (≤ ~10⁶), direct summation is
O(N) and fast (<10 ms). We use it for accuracy.
"""
if N <= 0:
return 1
if not (0.0 < gamma):
return N # ill-defined; safe upper bound
# Direct discrete cumulative
weights = [d ** (-gamma) for d in range(1, N + 1)]
total = sum(weights)
if total <= 0 or not math.isfinite(total):
# Fall back to continuum closed form (rare numerical edge case)
return _D_f_closed_continuum(gamma, f, N)
target = f * total
cum = 0.0
for d, w in enumerate(weights, start=1):
cum += w
if cum >= target:
return d
return N
def _D_f_closed_continuum(gamma: float, f: float, N: int) -> int:
"""Continuum closed form (paper Theorem 7.1) — asymptotic, kept as fallback."""
if abs(gamma - 1.0) < 1e-9:
return max(1, min(N, int(round(N ** f))))
one_minus_g = 1.0 - gamma
base = (1 - f) + f * (N ** one_minus_g)
if base <= 0:
return 1
try:
d_f = base ** (1.0 / one_minus_g)
except (OverflowError, ValueError):
return N
if not math.isfinite(d_f):
return N
return max(1, min(N, int(round(d_f))))
def delta_H(theta: float, Df: int, N: int) -> float:
sqrt2 = math.sqrt(2)
return math.log((theta + Df / sqrt2) / (theta + N / sqrt2))
def theta_eff_pade(theta: float, T: float) -> float:
return theta + T / math.sqrt(2)
def phase_label(gamma: float) -> str:
if gamma < 0.95:
return "A — deconfined (RoPE/long)"
if gamma > 1.05:
return "B — confined (AbsPE/short)"
return "Hagedorn (crossover γ≈1)"
def kl_divergence(p: np.ndarray, q: np.ndarray) -> float:
p = p / p.sum()
q = q / q.sum()
eps = 1e-12
mask = p > eps
return float(np.sum(p[mask] * np.log(p[mask] / (q[mask] + eps))))
# ── Attention measurement ──────────────────────────────────────────────────────
def set_seed(seed: int):
random.seed(seed)
np.random.seed(seed)
try:
import torch
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
except ImportError:
pass
def measure_attn_distance(model, tokenizer, distance: int, n_prompts: int,
seed: int, device: str, vocab_high: int) -> float:
import torch
set_seed(seed)
rng = random.Random(seed)
seq_len = distance + 50
target_pos = seq_len - distance - 1
last_pos = seq_len - 1
vocab_low = 1000
attn_values = []
model.eval()
with torch.no_grad():
for _ in range(n_prompts):
tokens = [rng.randint(vocab_low, vocab_high) for _ in range(seq_len)]
input_ids = torch.tensor([tokens], dtype=torch.long).to(device)
try:
out = model(input_ids, output_attentions=True, return_dict=True)
except Exception:
continue
if out.attentions is None:
raise RuntimeError(
"output_attentions returned None. "
"Try loading with attn_implementation='eager'."
)
vals = []
for layer_attn in out.attentions:
w = layer_attn[0, :, last_pos, target_pos].float().cpu().numpy()
finite = w[np.isfinite(w)]
if len(finite):
vals.append(float(np.mean(finite)))
if vals:
attn_values.append(float(np.mean(vals)))
return float(np.mean(attn_values)) if attn_values else float("nan")
def fit_power_law(distances: list, means: list) -> dict:
d = np.array(distances, dtype=float)
m = np.array(means, dtype=float)
mask = np.isfinite(m) & (m > 0)
if mask.sum() < 2:
return {"gamma": float("nan"), "log_A": 0.0, "R2": 0.0}
log_d = np.log(d[mask])
log_m = np.log(m[mask])
X = np.stack([np.ones(mask.sum()), -log_d], axis=1)
coeffs, *_ = np.linalg.lstsq(X, log_m, rcond=None)
log_A, gamma = float(coeffs[0]), float(coeffs[1])
pred = log_A - gamma * log_d
ss_res = float(np.sum((log_m - pred) ** 2))
ss_tot = float(np.sum((log_m - np.mean(log_m)) ** 2))
R2 = 1.0 - ss_res / ss_tot if ss_tot > 0 else 0.0
return {"gamma": gamma, "log_A": log_A, "R2": round(R2, 6)}
# ── Attention Grammar anomaly ──────────────────────────────────────────────────
def grammar_kl(attn_by_d: dict, gamma: float, log_A: float) -> float:
dists = sorted(attn_by_d.keys())
p_obs = np.array([attn_by_d[d] for d in dists], dtype=float)
p_obs = np.maximum(p_obs, 1e-30)
p_obs /= p_obs.sum()
A = math.exp(log_A)
p_prior = np.array([A * d ** (-gamma) for d in dists], dtype=float)
p_prior = np.maximum(p_prior, 1e-30)
p_prior /= p_prior.sum()
return kl_divergence(p_obs, p_prior)
# ── Main diagnostic ───────────────────────────────────────────────────────────
def run_diagnostic(args) -> dict:
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_name = args.model
theta_nom = args.theta or THETA_KNOWN.get(model_name, 10_000)
print(f"\n{'='*65}")
print(f"TRANSFORMER THERMODYNAMICS DIAGNOSTIC")
print(f"{'='*65}")
print(f" Model : {model_name}")
print(f" theta_nom : {theta_nom:,}")
print(f" N : {args.N}")
print(f" Mode : {'fast' if args.fast else 'full'}")
print()
# ── Load model ──────────────────────────────────────────────────────
local_path = args.local or model_name
print(f"Loading model from: {local_path} ...")
t0 = time.time()
load_kwargs = dict(
trust_remote_code=True,
attn_implementation="eager",
)
device = "cuda" if (not args.cpu and torch.cuda.is_available()) else "cpu"
if args.load_in_4bit and device == "cuda":
try:
from transformers import BitsAndBytesConfig
load_kwargs["quantization_config"] = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
load_kwargs["device_map"] = "auto"
except ImportError:
print(" [warn] bitsandbytes not available; loading in float32")
elif device == "cpu":
load_kwargs["dtype"] = torch.float32
tokenizer = AutoTokenizer.from_pretrained(local_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(local_path, **load_kwargs)
if device == "cpu":
model = model.to("cpu")
model.eval()
print(f" Loaded in {time.time()-t0:.1f}s device={device}")
vocab_high = min(tokenizer.vocab_size - 1, 49_000)
distances = DISTANCES_FAST if args.fast else DISTANCES_FULL
n_prompts = N_PROMPTS if args.fast else N_PROMPTS_FULL
N = args.N
# ── Measure attention by distance ────────────────────────────────────
print(f"\nMeasuring attention decay at {len(distances)} distances × {n_prompts} prompts ...")
attn_by_d = {}
for dist in distances:
if dist > N:
continue
t1 = time.time()
mean_val = measure_attn_distance(
model, tokenizer, dist, n_prompts, SEEDS[0], device, vocab_high
)
attn_by_d[dist] = mean_val
print(f" d={dist:5d} attn={mean_val:.6f} ({time.time()-t1:.1f}s)")
# ── Fit power law ────────────────────────────────────────────────────
valid_d = [d for d, v in attn_by_d.items() if math.isfinite(v) and v > 0]
valid_v = [attn_by_d[d] for d in valid_d]
fit = fit_power_law(valid_d, valid_v)
gamma = fit["gamma"]
log_A = fit["log_A"]
R2 = fit["R2"]
if not math.isfinite(gamma):
print("\n[ERROR] Power-law fit failed. Too few valid distances.")
return {}
# ── Thermodynamics ───────────────────────────────────────────────────
Z = partition_Z(gamma, N)
U = mean_log_d(gamma, N)
S = entropy_S(gamma, N)
F = free_energy_F(gamma, N)
Cv = heat_capacity_Cv(gamma, N)
chi = 1.0 / abs(gamma - 1.0) if abs(gamma - 1.0) > 1e-4 else 1e6
xi = 1.0 / abs(math.log(gamma)) if abs(math.log(gamma)) > 1e-10 else 1e6
T_attn = 1.0 / gamma
D90 = D_f_closed(gamma, 0.90, N)
dH90 = delta_H(theta_nom, D90, N)
theta_eff = theta_eff_pade(theta_nom, float(N))
# Theoretical γ prediction — γ_Padé(θ, T_eval) (paper §3.3, supersedes
# the earlier shorthand γ ≈ C/lnθ which assumed T = 10000).
if theta_nom > 0:
T_for_pred = max(distances) if distances else N # use largest measured T
z_sqrt2 = T_for_pred * math.sqrt(2)
gamma_pred = (2 * theta_nom - z_sqrt2) / (2 * theta_nom + z_sqrt2)
else:
gamma_pred = None
# Attention grammar KL
kl_ag = grammar_kl(attn_by_d, gamma, log_A)
# Phase
phase = phase_label(gamma)
# ── Report ───────────────────────────────────────────────────────────
print(f"\n{'='*65}")
print(f"RESULTS")
print(f"{'='*65}")
print(f" γ (gamma) = {gamma:.4f} [R²={R2:.4f}]")
if gamma_pred is not None:
delta_g = gamma - gamma_pred
print(f" γ_Padé(θ,T) = {gamma_pred:.4f} Δγ = {delta_g:+.4f}")
print(f" Phase : {phase}")
print(f" T_attn = 1/γ = {T_attn:.4f}")
print()
print(f" Thermodynamics (N={N}):")
print(f" Z (partition) = {Z:.4f}")
print(f" U = E[log d] = {U:.4f}")
print(f" S (entropy) = {S:.4f}")
print(f" F (free ener) = {F:.4f}")
cv_str = f"{Cv:.4f}" if math.isfinite(Cv) else "N/A"
print(f" C_V (heat cap)= {cv_str}")
chi_str = f"{chi:.2f}" if chi < 1e5 else "∞ (near Hagedorn)"
print(f" χ (suscept.) = {chi_str}")
xi_str = f"{xi:.2f}" if xi < 1e5 else "∞"
print(f" ξ (corr. len) = {xi_str}")
print()
print(f" KV Compression (f=0.90):")
print(f" D_90 = {D90} tokens ({D90/N*100:.1f}% of N={N})")
print(f" dH_90 = {dH90:.4f} nats")
print()
print(f" RoPE Diagnostic:")
print(f" theta_nom = {theta_nom:,}")
print(f" theta_eff_Pade = {theta_eff:.1f}")
print()
print(f" Attention Grammar:")
print(f" KL(obs||prior) = {kl_ag:.4f} ", end="")
if kl_ag > 0.05:
print("[HIGH — non-power-law circuits present]")
elif kl_ag > 0.01:
print("[MODERATE — some circuit deviation]")
else:
print("[LOW — pure positional attention]")
print(f"\n γ interpretation:")
if gamma < 0.7:
print(f" Very long-range attention (large θ, LLaMA-3/Qwen2.5 class)")
elif gamma < 0.95:
print(f" Long-range attention (standard RoPE, Phase A)")
elif gamma < 1.05:
print(f" Hagedorn crossover — attention at phase boundary")
elif gamma < 1.3:
print(f" Short-range attention (AbsPE or short context training)")
else:
print(f" Highly local attention (possible SWA or very short context)")
# ── Save ─────────────────────────────────────────────────────────────
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
short = model_name.replace("/", "--")
result = {
"model": model_name,
"theta_nom": theta_nom,
"N": N,
"fast_mode": args.fast,
"fit_power_law": fit,
"gamma": gamma,
"gamma_pred": gamma_pred,
"delta_gamma": (gamma - gamma_pred) if gamma_pred else None,
"phase": phase,
"T_attn": T_attn,
"Z": Z, "U": U, "S": S, "F": F, "Cv": Cv,
"chi": chi, "xi": xi,
"D90": D90,
"D90_frac": D90 / N,
"delta_H_90": dH90,
"theta_eff_pade": theta_eff,
"kl_grammar": kl_ag,
"attn_by_distance": {str(d): v for d, v in attn_by_d.items()},
}
out_path = OUTPUT_DIR / f"{short}.json"
out_path.write_text(json.dumps(result, indent=2, default=float), encoding="utf-8")
print(f"\n Saved: {out_path}")
print(f"{'='*65}\n")
return result
def main():
parser = argparse.ArgumentParser(
description="Predicting How Transformers Attend — diagnostic for any causal LM"
)
parser.add_argument("--model", required=True,
help="HuggingFace model ID (e.g. EleutherAI/pythia-70m)")
parser.add_argument("--local", default=None,
help="Local path to model weights (if not downloading)")
parser.add_argument("--theta", type=int, default=None,
help="RoPE θ (auto-detected for known models)")
parser.add_argument("--N", type=int, default=2000,
help="Context length N for thermodynamic calculations (default 2000)")
parser.add_argument("--fast", action="store_true",
help="Fast mode: fewer distances and prompts (~5 min on CPU)")
parser.add_argument("--load_in_4bit", action="store_true",
help="Load model in 4-bit quantization (requires bitsandbytes)")
parser.add_argument("--cpu", action="store_true",
help="Force CPU even if CUDA available")
args = parser.parse_args()
try:
run_diagnostic(args)
except KeyboardInterrupt:
print("\n[interrupted]")
except Exception as e:
print(f"\n[ERROR] {e}")
raise
if __name__ == "__main__":
main()
|