File size: 6,000 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 | #!/usr/bin/env python3
"""
Comprehensive intervention worker for 100k checkpoints (Layer 0).
For each checkpoint: generates N_SEQ random sequences, intervenes at every
sorted-output position with multiple intensities, records per-trial details.
Methodology matches existing perlocation/pernumber experiments:
- unsorted_lb_num=0, unsorted_ub_num=1 (boost one wrong unsorted number)
- ub=60 (wide neighbourhood)
- Same GPTIntervention mechanism from grid-run/model_analysis.py
"""
import argparse
import json
import os
import sys
import time
import numpy as np
import torch
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'grid-run'))
from model_analysis import GPT, GPTConfig, GPTIntervention
N_SEQ = 3000
INTENSITIES = [2.0, 4.0, 6.0, 10.0]
UB = 60
def remap_state_dict(sd):
new = {}
for k, v in sd.items():
nk = k
for i in range(10):
nk = nk.replace(f'transformer.h.{i}.attn.', f'transformer.h.{i}.c_attn.')
nk = nk.replace(f'transformer.h.{i}.mlp.', f'transformer.h.{i}.c_fc.')
new[nk] = v
return new
def load_model(ckpt_path, device):
ckpt = torch.load(ckpt_path, map_location='cpu')
mc = ckpt['model_config']
config = GPTConfig(block_size=mc['block_size'], vocab_size=mc['vocab_size'] - 1,
with_layer_norm=mc.get('use_final_LN', True))
model = GPT(config)
sd = remap_state_dict(ckpt['model_state_dict'])
wpe_max = config.block_size * 4 + 1
if 'transformer.wpe.weight' in sd and sd['transformer.wpe.weight'].shape[0] > wpe_max:
sd['transformer.wpe.weight'] = sd['transformer.wpe.weight'][:wpe_max]
for k in [k for k in sd if k.endswith('.c_attn.bias') and 'c_attn.c_attn' not in k]:
del sd[k]
if 'lm_head.weight' in sd:
del sd['lm_head.weight']
model.load_state_dict(sd, strict=False)
model.to(device)
model.eval()
return model, config
def get_batch(vs, bs, device):
x = torch.randperm(vs)[:bs]
vals, _ = torch.sort(x)
return torch.cat((x, torch.tensor([vs]), vals), dim=0).unsqueeze(0).to(device)
@torch.no_grad()
def run_checkpoint(model, config, device):
bs = config.block_size
vs = config.vocab_size
pos_l, num_l, nxt_l, gap_l = [], [], [], []
int_l, cor_l, pred_l = [], [], []
n_ok = n_fail = 0
for si in range(N_SEQ):
idx = get_batch(vs, bs, device)
# --- baseline predictions (intensity=0) ---
logits, _ = model(idx)
bpreds = torch.argmax(logits, dim=-1)[0]
for p in range(bs - 1):
loc = bs + 1 + p
num_val = idx[0, loc].item()
nxt_val = idx[0, loc + 1].item()
pos_l.append(p)
num_l.append(num_val)
nxt_l.append(nxt_val)
gap_l.append(nxt_val - num_val)
int_l.append(0.0)
pr = bpreds[loc].item()
cor_l.append(int(pr == nxt_val))
pred_l.append(pr)
# --- interventions ---
try:
im = GPTIntervention(model, idx)
except Exception:
continue
for p in range(bs - 1):
loc = bs + 1 + p
num_val = idx[0, loc].item()
nxt_val = idx[0, loc + 1].item()
gap = nxt_val - num_val
for intensity in INTENSITIES:
try:
im.intervent_attention(
attention_layer_num=0, location=loc,
unsorted_lb=UB, unsorted_ub=UB,
unsorted_lb_num=0, unsorted_ub_num=1,
unsorted_intensity_inc=intensity,
sorted_lb=0, sorted_num=0, sorted_intensity_inc=0.0)
pr, ac = im.check_if_still_works()
pos_l.append(p)
num_l.append(num_val)
nxt_l.append(nxt_val)
gap_l.append(gap)
int_l.append(intensity)
cor_l.append(int(pr == ac))
pred_l.append(pr)
im.revert_attention(0)
n_ok += 1
except Exception:
try:
im.revert_attention(0)
except Exception:
pass
n_fail += 1
if (si + 1) % 500 == 0:
print(f" {si+1}/{N_SEQ} ok={n_ok} fail={n_fail}", flush=True)
return dict(
position=np.array(pos_l, dtype=np.int16),
number=np.array(num_l, dtype=np.int16),
next_number=np.array(nxt_l, dtype=np.int16),
gap=np.array(gap_l, dtype=np.int16),
intensity=np.array(int_l, dtype=np.float32),
correct=np.array(cor_l, dtype=np.int8),
predicted=np.array(pred_l, dtype=np.int16),
)
def main():
ap = argparse.ArgumentParser()
ap.add_argument('--tasks-file', required=True)
ap.add_argument('--gpu', type=int, required=True)
args = ap.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
device = 'cuda'
with open(args.tasks_file) as f:
tasks = json.load(f)
print(f"GPU {args.gpu}: {len(tasks)} checkpoints to process", flush=True)
for t in tasks:
if os.path.exists(t['out']):
print(f" Skip {t['name']} (cached)", flush=True)
continue
t0 = time.time()
model, config = load_model(t['ckpt_path'], device)
print(f" Loaded {t['name']} ({time.time()-t0:.1f}s)", flush=True)
t0 = time.time()
res = run_checkpoint(model, config, device)
os.makedirs(os.path.dirname(t['out']), exist_ok=True)
np.savez_compressed(t['out'], **res)
dt = time.time() - t0
n = len(res['position'])
print(json.dumps({
'done': t['name'], 'gpu': args.gpu,
'elapsed': round(dt, 1), 'n_trials': n
}), flush=True)
print(f"GPU {args.gpu}: all done", flush=True)
if __name__ == '__main__':
main()
|