File size: 18,331 Bytes
beda614 | 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 | """
Parallel launcher for 100k-checkpoint analysis.
Assigns checkpoints to 8 GPUs (one persistent worker per GPU).
Assembles plots incrementally as each checkpoint completes.
"""
import json
import os
import subprocess
import sys
import time
import glob
import threading
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
CKPT_DIR = os.path.join(SCRIPT_DIR, 'final_models')
OUTPUT_BASE = os.path.join(SCRIPT_DIR, 'outputs')
NUM_GPUS = 8
UB_VALUES = [5, 10, 15, 20, 30, 50, 60]
def discover_checkpoints():
pt_files = sorted(glob.glob(os.path.join(CKPT_DIR, '*.pt')))
checkpoints = []
for pt in pt_files:
bn = os.path.basename(pt)
if '.summary.' in bn:
continue
parts = bn.replace('.pt', '').split('__')
config_str = parts[0]
ckpt_type = parts[1] if len(parts) > 1 else 'final'
tokens = config_str.split('_')
dseed = iseed = None
for t in tokens:
if t.startswith('dseed'):
dseed = t.replace('dseed', '')
elif t.startswith('iseed'):
iseed = t.replace('iseed', '')
if ckpt_type.startswith('ckpt'):
itr = int(ckpt_type.replace('ckpt', ''))
ckpt_label = f'ckpt{itr}'
else:
itr = 100000
ckpt_label = 'final'
folder_name = f"plots_N256_B16_ds{dseed}_is{iseed}_{ckpt_label}"
checkpoints.append({
'path': pt,
'dseed': dseed,
'iseed': iseed,
'itr': itr,
'ckpt_label': ckpt_label,
'folder_name': folder_name,
})
return checkpoints
def make_tasks_for_checkpoint(ckpt):
"""Generate all analysis tasks for a single checkpoint."""
tmp_dir = os.path.join(OUTPUT_BASE, 'tmp_results', ckpt['folder_name'])
tasks = []
tasks.append({
'ckpt_path': ckpt['path'], 'type': 'baseline', 'layer': 0,
'out': os.path.join(tmp_dir, 'baseline.npz'),
'name': f"{ckpt['folder_name']}_baseline",
'itr': ckpt['itr'],
})
for layer in [0, 1]:
tasks.append({
'ckpt_path': ckpt['path'], 'type': 'ablation', 'layer': layer,
'out': os.path.join(tmp_dir, f'ablation_layer{layer}.npz'),
'name': f"{ckpt['folder_name']}_ablation_L{layer}",
'itr': ckpt['itr'],
})
for layer in [0, 1]:
tasks.append({
'ckpt_path': ckpt['path'], 'type': 'cinclogits', 'layer': layer,
'out': os.path.join(tmp_dir, f'cinclogits_layer{layer}.npz'),
'name': f"{ckpt['folder_name']}_cinclogits_L{layer}",
'itr': ckpt['itr'],
})
for ub in UB_VALUES:
for layer in [0, 1]:
suffix = '' if ub == 5 else f'_ub{ub}'
tasks.append({
'ckpt_path': ckpt['path'], 'type': 'intensity', 'layer': layer,
'ub': ub,
'out': os.path.join(tmp_dir, f'intensity_layer{layer}{suffix}.npz'),
'name': f"{ckpt['folder_name']}_intensity_ub{ub}_L{layer}",
'itr': ckpt['itr'],
})
return tasks
def is_checkpoint_done(ckpt):
"""Check if all tasks for a checkpoint are done."""
tasks = make_tasks_for_checkpoint(ckpt)
return all(os.path.exists(t['out']) for t in tasks)
def assemble_plots_for_checkpoint(ckpt):
"""Assemble all plots for a single checkpoint."""
folder_name = ckpt['folder_name']
tmp_dir = os.path.join(OUTPUT_BASE, 'tmp_results', folder_name)
plot_dir = os.path.join(OUTPUT_BASE, folder_name)
os.makedirs(plot_dir, exist_ok=True)
tag = (f"N=256 block=16 lr=0.01 iters={ckpt['itr']} "
f"dseed={ckpt['dseed']} iseed={ckpt['iseed']}")
for fn, label in [(_assemble_baseline, 'baseline'),
(_assemble_ablation, 'ablation'),
(_assemble_cinclogits, 'cinclogits'),
(_assemble_intensity, 'intensity')]:
try:
fn(tmp_dir, plot_dir, tag)
except Exception as e:
print(f" WARN {label} for {folder_name}: {e}", flush=True)
n_plots = len([f for f in os.listdir(plot_dir) if f.endswith('.png')])
return n_plots
def monitor_workers(procs, checkpoints, ckpt_by_gpu):
"""Monitor worker processes and assemble plots as checkpoints complete."""
assembled = set()
total_tasks = sum(len(make_tasks_for_checkpoint(c)) for c in checkpoints)
total_done = 0
# Assemble already-cached checkpoints
for ckpt in checkpoints:
if is_checkpoint_done(ckpt) and ckpt['folder_name'] not in assembled:
n = assemble_plots_for_checkpoint(ckpt)
assembled.add(ckpt['folder_name'])
print(f" [PLOTS] {ckpt['folder_name']}: {n} plots (cached)", flush=True)
t0 = time.time()
while any(p.poll() is None for p in procs.values()):
time.sleep(5)
for ckpt in checkpoints:
fn = ckpt['folder_name']
if fn not in assembled and is_checkpoint_done(ckpt):
n = assemble_plots_for_checkpoint(ckpt)
assembled.add(fn)
elapsed = time.time() - t0
print(f" [PLOTS] {fn}: {n} plots assembled ({elapsed:.0f}s)",
flush=True)
done_now = sum(1 for c in checkpoints
for t in make_tasks_for_checkpoint(c)
if os.path.exists(t['out']))
if done_now > total_done + 10:
total_done = done_now
elapsed = time.time() - t0
rate = total_done / elapsed if elapsed > 0 else 0
eta = (total_tasks - total_done) / rate if rate > 0 else 0
print(f" [PROGRESS] {total_done}/{total_tasks} tasks, "
f"{len(assembled)}/{len(checkpoints)} ckpts plotted "
f"({elapsed:.0f}s, ETA ~{eta:.0f}s)", flush=True)
# Final assembly for any remaining
for ckpt in checkpoints:
fn = ckpt['folder_name']
if fn not in assembled and is_checkpoint_done(ckpt):
n = assemble_plots_for_checkpoint(ckpt)
assembled.add(fn)
print(f" [PLOTS] {fn}: {n} plots assembled (final)", flush=True)
# Check for failures
for gpu, proc in procs.items():
if proc.returncode != 0:
print(f" [WARN] GPU {gpu} worker exited with code {proc.returncode}",
flush=True)
elapsed = time.time() - t0
return assembled, elapsed
def main():
t_start = time.time()
checkpoints = discover_checkpoints()
print(f"Found {len(checkpoints)} checkpoints")
# Filter to only checkpoints that still need work
todo = [c for c in checkpoints if not is_checkpoint_done(c)]
already_done = len(checkpoints) - len(todo)
print(f"Already complete: {already_done}, to process: {len(todo)}")
# Assemble plots for already-done checkpoints first
for ckpt in checkpoints:
if is_checkpoint_done(ckpt):
plot_dir = os.path.join(OUTPUT_BASE, ckpt['folder_name'])
if not os.path.isdir(plot_dir) or not any(f.endswith('.png') for f in os.listdir(plot_dir)):
n = assemble_plots_for_checkpoint(ckpt)
print(f" [PLOTS] {ckpt['folder_name']}: {n} plots (from cache)", flush=True)
if not todo:
print("All checkpoints already analyzed and plotted!")
return
# Build task lists per GPU: distribute checkpoints round-robin
gpu_tasks = {g: [] for g in range(NUM_GPUS)}
for i, ckpt in enumerate(todo):
gpu = i % NUM_GPUS
tasks = make_tasks_for_checkpoint(ckpt)
# Skip already-cached tasks
tasks = [t for t in tasks if not os.path.exists(t['out'])]
gpu_tasks[gpu].extend(tasks)
# Sort each GPU's tasks so same-checkpoint tasks are contiguous (model loaded once)
for g in gpu_tasks:
gpu_tasks[g].sort(key=lambda t: (t['ckpt_path'], t['type']))
total_tasks = sum(len(v) for v in gpu_tasks.values())
print(f"\nTotal tasks to run: {total_tasks}, distributed across {NUM_GPUS} GPUs")
for g in range(NUM_GPUS):
n = len(gpu_tasks[g])
ckpts = len(set(t['ckpt_path'] for t in gpu_tasks[g])) if n else 0
print(f" GPU {g}: {n} tasks across {ckpts} checkpoints")
# Write task files and launch workers
task_dir = os.path.join(OUTPUT_BASE, 'task_files')
os.makedirs(task_dir, exist_ok=True)
procs = {}
log_dir = os.path.join(OUTPUT_BASE, 'gpu_worker_logs')
os.makedirs(log_dir, exist_ok=True)
for g in range(NUM_GPUS):
if not gpu_tasks[g]:
continue
task_file = os.path.join(task_dir, f'gpu{g}_tasks.json')
with open(task_file, 'w') as f:
json.dump(gpu_tasks[g], f)
log_file = open(os.path.join(log_dir, f'gpu{g}.log'), 'w')
proc = subprocess.Popen(
[sys.executable, os.path.join(SCRIPT_DIR, 'gpu_worker.py'),
'--tasks-file', task_file, '--gpu', str(g)],
stdout=log_file, stderr=subprocess.STDOUT,
cwd=SCRIPT_DIR)
procs[g] = proc
print(f"\nLaunched {len(procs)} GPU workers. Monitoring...\n", flush=True)
assembled, elapsed = monitor_workers(procs, checkpoints, gpu_tasks)
total_elapsed = time.time() - t_start
print(f"\n{'='*60}")
print(f"ALL DONE β {len(assembled)}/{len(checkpoints)} checkpoints plotted")
print(f"Elapsed: {total_elapsed:.0f}s ({total_elapsed/60:.1f}m)")
print(f"Output: {OUTPUT_BASE}")
print(f"{'='*60}")
# ββ Plot assembly functions ββββββββββββββββββββββββββββββββββββββββββββββββββ
def _assemble_baseline(tmp_dir, plot_dir, tag):
f = os.path.join(tmp_dir, 'baseline.npz')
if not os.path.exists(f):
return
d = np.load(f)
full_seq_acc = float(d['full_seq_acc'])
cond_acc = d['cond_acc']
cond_eligible = d['cond_eligible']
fig, ax = plt.subplots(figsize=(4, 4))
bars = ax.bar([0], [full_seq_acc], 0.5, color='#e6850e')
for b in bars:
h = b.get_height()
ax.text(b.get_x() + b.get_width() / 2, h + 0.01,
f'{h:.3f}', ha='center', va='bottom', fontsize=11, fontweight='bold')
ax.set_xticks([0])
ax.set_xticklabels(['Model'], fontsize=12)
ax.set_ylabel('Full-sequence accuracy', fontsize=12)
ax.set_title(f'Baseline accuracy (500 trials)\n{tag}', fontsize=11, fontweight='bold')
ax.grid(True, axis='y', alpha=0.2, linestyle=':')
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.set_ylim(0, min(1.15, full_seq_acc * 1.2 + 0.05))
fig.tight_layout()
fig.savefig(os.path.join(plot_dir, 'baseline_accuracy.png'), dpi=300, bbox_inches='tight')
plt.close()
fig, ax = plt.subplots(figsize=(6, 4))
pos = np.arange(len(cond_acc))
valid = cond_eligible >= 10
if valid.any():
ax.plot(pos[valid], cond_acc[valid], marker='s', markersize=3, linewidth=1.2,
color='#e6850e')
if not valid.all():
cutoff = np.where(~valid)[0][0]
ax.axvline(x=cutoff - 0.5, color='#e6850e', linestyle=':', alpha=0.5)
ax.set_xlabel('Output position', fontsize=10)
ax.set_ylabel('Conditional per-token accuracy', fontsize=10)
ax.set_title(f'Per-token accuracy (given correct prefix) β baseline (500 trials)\n{tag}',
fontsize=11, fontweight='bold')
ax.grid(True, alpha=0.3)
ax.set_ylim(0, 1.05)
fig.tight_layout()
fig.savefig(os.path.join(plot_dir, 'baseline_conditional_accuracy.png'),
dpi=300, bbox_inches='tight')
plt.close()
def _assemble_ablation(tmp_dir, plot_dir, tag):
data = {}
for layer in [0, 1]:
f = os.path.join(tmp_dir, f'ablation_layer{layer}.npz')
if not os.path.exists(f):
return
d = np.load(f)
data[layer] = {
'full_seq_acc': float(d['full_seq_acc']),
'per_pos_acc': d['per_pos_acc'],
'cond_acc': d['cond_acc'],
'cond_eligible': d['cond_eligible'],
}
fig, ax = plt.subplots(figsize=(5, 4.5))
vals = [data[0]['full_seq_acc'], data[1]['full_seq_acc']]
bars = ax.bar([0, 1], vals, 0.5, color=['#1f77b4', '#ff7f0e'])
for b in bars:
h = b.get_height()
ax.text(b.get_x() + b.get_width() / 2, h + 0.01,
f'{h:.3f}', ha='center', va='bottom', fontsize=10, fontweight='bold')
ax.set_xticks([0, 1])
ax.set_xticklabels(['Skip Layer 0', 'Skip Layer 1'], fontsize=12)
ax.set_ylabel('Full-sequence accuracy', fontsize=12)
ax.set_title(f'Accuracy with attention layer removed (500 trials)\n{tag}',
fontsize=11, fontweight='bold')
ax.grid(True, axis='y', alpha=0.2, linestyle=':')
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.set_ylim(0, min(1.15, max(vals) * 1.25 + 0.05))
fig.tight_layout()
fig.savefig(os.path.join(plot_dir, 'ablation_accuracy.png'), dpi=300, bbox_inches='tight')
plt.close()
fig, axes = plt.subplots(1, 2, figsize=(10, 4), sharey=True)
for i, layer in enumerate([0, 1]):
ax = axes[i]
pos = np.arange(len(data[layer]['per_pos_acc']))
ax.plot(pos, data[layer]['per_pos_acc'], marker='o', markersize=3,
linewidth=1.2, color=['#1f77b4', '#ff7f0e'][i])
ax.set_xlabel('Output position', fontsize=10)
ax.set_title(f'Skip Layer {layer}', fontsize=11, fontweight='bold')
ax.grid(True, alpha=0.3)
ax.set_ylim(0, 1.05)
axes[0].set_ylabel('Per-position accuracy', fontsize=10)
fig.suptitle(f'Per-position accuracy with attention removed (500 trials)\n{tag}',
fontsize=11, fontweight='bold')
fig.tight_layout()
fig.savefig(os.path.join(plot_dir, 'ablation_per_position.png'),
dpi=300, bbox_inches='tight')
plt.close()
fig, axes = plt.subplots(1, 2, figsize=(10, 4), sharey=True)
for i, layer in enumerate([0, 1]):
ax = axes[i]
ca = data[layer]['cond_acc']
ce = data[layer]['cond_eligible']
pos = np.arange(len(ca))
valid = ce >= 10
color = ['#1f77b4', '#ff7f0e'][i]
if valid.any():
ax.plot(pos[valid], ca[valid], marker='o', markersize=3,
linewidth=1.2, color=color)
if not valid.all():
cutoff = np.where(~valid)[0][0]
ax.axvline(x=cutoff - 0.5, color=color, linestyle=':', alpha=0.5)
ax.set_xlabel('Output position', fontsize=10)
ax.set_title(f'Skip Layer {layer}', fontsize=11, fontweight='bold')
ax.grid(True, alpha=0.3)
ax.set_ylim(0, 1.05)
axes[0].set_ylabel('Conditional per-token accuracy', fontsize=10)
fig.suptitle(
f'Per-token accuracy (given prefix correct) with attention removed (500 trials)\n{tag}',
fontsize=11, fontweight='bold')
fig.tight_layout()
fig.savefig(os.path.join(plot_dir, 'ablation_conditional_accuracy.png'),
dpi=300, bbox_inches='tight')
plt.close()
def _assemble_cinclogits(tmp_dir, plot_dir, tag):
for layer in [0, 1]:
f = os.path.join(tmp_dir, f'cinclogits_layer{layer}.npz')
if not os.path.exists(f):
continue
d = np.load(f)
cl_ic = d['clogit_icscore']
icl_ic = d['iclogit_icscore']
frac_ic = np.mean(cl_ic + icl_ic)
eps = 1e-10
corr = np.sum(cl_ic) / (np.sum(cl_ic + icl_ic) + eps)
fig, ax = plt.subplots(figsize=(4.5, 4))
bw = 0.5
x = np.array([0, 1])
b1 = ax.bar(x[0], frac_ic, bw, color='#e6850e')
b2 = ax.bar(x[1], corr, bw, color='#1f77b4')
for bars in [b1, b2]:
for b in bars:
h = b.get_height()
ax.text(b.get_x() + b.get_width() / 2, h + 0.008,
f'{h:.3f}', ha='center', va='bottom', fontsize=10, fontweight='bold')
ax.set_xticks(x)
ax.set_xticklabels(['Fraction of\nincorrect scores',
'Logit correction ratio\namong incorrect scores'], fontsize=10)
ax.set_ylabel('Fraction', fontsize=12)
ax.set_title(f'Incorrect scores & logit correction (Layer {layer})\n{tag}',
fontsize=11, fontweight='bold')
ax.grid(True, axis='y', alpha=0.2, linestyle=':')
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ymax = max(frac_ic, corr)
ax.set_ylim(0, ymax * 1.25 if ymax > 0 else 1.0)
fig.tight_layout()
fig.savefig(os.path.join(plot_dir, f'cinclogits_layer{layer}.png'),
dpi=300, bbox_inches='tight')
plt.close()
def _assemble_intensity(tmp_dir, plot_dir, tag):
for ub in UB_VALUES:
for layer in [0, 1]:
suffix = '' if ub == 5 else f'_ub{ub}'
f = os.path.join(tmp_dir, f'intensity_layer{layer}{suffix}.npz')
if not os.path.exists(f):
continue
d = np.load(f)
intensities = d['intensities']
rates = d['success_rates']
plt.figure(figsize=(4.5, 3.2))
plt.plot(intensities, rates, marker='o', linewidth=1.5, markersize=5,
color='#e6850e')
plt.xlabel('Intervention Intensity', fontsize=9)
plt.ylabel('Success Probability', fontsize=9)
title = f'Robustness to Attention Intervention (Layer {layer})'
if ub != 5:
title += f' [ub={ub}]'
title += f'\n{tag}'
plt.title(title, fontsize=9)
plt.grid(True, alpha=0.3)
plt.xticks(intensities, fontsize=8)
plt.yticks(fontsize=8)
plt.ylim(0, 1.05)
plt.tight_layout()
fname = f'intensity_layer{layer}'
if ub != 5:
fname += f'_ub{ub}'
plt.savefig(os.path.join(plot_dir, f'{fname}.png'),
dpi=300, bbox_inches='tight')
plt.close()
if __name__ == '__main__':
main()
|