1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
|
import torch, pickle
from pathlib import Path
import lt_ep_train as L
L.DD = Path('data/tinystories_bpe'); L.vocab = pickle.load(open(L.DD/'meta.pkl','rb'))['vocab_size']
blk = L.EQBlock(512,16,256,256, s=1.0, c=1.0, attn_mode='thick'); blk.qknorm=True
ck = torch.load('runs/bptt_spec_snapshot.pt', map_location='cpu')
with torch.no_grad():
for p,w in zip(blk.allp, ck['allp']): p.copy_(w.to('cpu').float())
print(f"# BPTT ckpt step={ck.get('step')} best_val={ck.get('best')}")
print(f"# specnorm cap = 0.9")
for name in ['WQ','WK','WV','WO','fc','pj']:
W = getattr(blk, name).float()
s = torch.linalg.svdvals(W)[0].item()
flag = " <-- EXCEEDS cap" if s > 0.9 else ""
print(f"{name}{tuple(W.shape)}: sigma_max={s:.3f}{flag}")
|
