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
|
"""Cascade-EP trainer — EQUILIBRIUM MODE (the true-EP route).
Two-phase (+-beta) relaxation of all layer states to the nudged equilibria via
Gauss-Seidel reverse sweeps (solver choice only; readout is taken AT the relaxed
states with the standard EP formula), weight grad = (1/2beta)[dF/dtheta|+ - dF/dtheta|-].
Inference = plain forward (standard LLM). Twin of casc_bp_train.py (same seed/data)."""
import argparse, math, pickle, time
import numpy as np, torch, torch.nn as nn, torch.nn.functional as F
from pathlib import Path
ap = argparse.ArgumentParser()
ap.add_argument('--tag', default='casc_eq6')
ap.add_argument('--L', type=int, default=6); ap.add_argument('--C', type=int, default=256)
ap.add_argument('--H', type=int, default=8); ap.add_argument('--T', type=int, default=256)
ap.add_argument('--B', type=int, default=24); ap.add_argument('--steps', type=int, default=4000)
ap.add_argument('--lr', type=float, default=3e-4); ap.add_argument('--warmup', type=int, default=200)
ap.add_argument('--beta', type=float, default=0.003); ap.add_argument('--seed', type=int, default=0)
ap.add_argument('--K', type=int, default=3) # fb (message-passing) rounds
ap.add_argument('--geta', type=float, default=1.0) # fb mixing (1.0 = undamped)
ap.add_argument('--save_every', type=int, default=1000); ap.add_argument('--log', type=int, default=100)
ap.add_argument('--wandb', default=''); ap.add_argument('--wandb_run', default='')
ap.add_argument('--kmax', type=int, default=8) # adaptive fb rounds cap
ap.add_argument('--gate_every', type=int, default=200) # in-training cos(EP,BP) telemetry
args = ap.parse_args()
torch.manual_seed(args.seed)
dev = 'cuda' if torch.cuda.is_available() else 'cpu'
DD = Path('/home/yurenh2/ept/ep_run/data/tinystories_bpe')
vocab = pickle.load(open(DD / 'meta.pkl', 'rb'))['vocab_size']
def get_batch(split):
data = np.memmap(DD / ('train.bin' if split == 'train' else 'val.bin'), dtype=np.uint16, mode='r')
ix = torch.randint(len(data) - args.T - 1, (args.B,))
x = torch.stack([torch.from_numpy(data[i:i + args.T].astype(np.int64)) for i in ix])
y = torch.stack([torch.from_numpy(data[i + 1:i + 1 + args.T].astype(np.int64)) for i in ix])
return x.to(dev), y.to(dev)
class Block(nn.Module):
def __init__(self, C, H):
super().__init__()
self.ln1, self.ln2 = nn.LayerNorm(C), nn.LayerNorm(C)
self.attn = nn.MultiheadAttention(C, H, batch_first=True)
self.ff = nn.Sequential(nn.Linear(C, 4 * C), nn.GELU(), nn.Linear(4 * C, C))
def forward(self, z, mask):
h = self.ln1(z); a, _ = self.attn(h, h, h, attn_mask=mask, need_weights=False)
z = z + a; return z + self.ff(self.ln2(z))
tok = nn.Embedding(vocab, args.C).to(dev)
pos = nn.Embedding(args.T, args.C).to(dev)
blocks = nn.ModuleList([Block(args.C, args.H) for _ in range(args.L)]).to(dev)
mask = torch.triu(torch.full((args.T, args.T), float('-inf'), device=dev), 1)
readout = lambda z: z @ tok.weight.t()
all_params = list(tok.parameters()) + list(pos.parameters()) + list(blocks.parameters())
opt = torch.optim.AdamW(all_params, lr=args.lr, weight_decay=1e-4)
sched = torch.optim.lr_scheduler.LambdaLR(opt, lambda s: min(1.0, (s + 1) / max(args.warmup, 1)))
NBT = args.B * args.T
def free_states(x):
with torch.no_grad():
z = tok(x) + pos(torch.arange(args.T, device=dev))[None]
z0 = z.clone(); zs = []
for b in blocks:
z = b(z, mask); zs.append(z)
return z0, zs
@torch.no_grad()
def tok_sigma(iters=8):
"""top singular value of tok.weight (power iteration on the raw matrix)."""
W = tok.weight
v = torch.randn(W.shape[1], device=dev); v /= v.norm()
sig = 1.0
for _ in range(iters):
u = W @ v; u /= max(u.norm(), 1e-12)
v = W.t() @ u; sig = v.norm(); v /= max(sig, 1e-12)
return float(sig)
def relax(z0, zs_free, y, beta):
"""adaptive-K fb rounds: backward feedback refresh + forward rebuild, until the
per-round state delta contracts below 5% of round-1 (or kmax). Returns (zs, rounds, contracted)."""
zs = [z.clone() for z in zs_free]
d = [None] * args.L
d1 = None; delta = 0.0
for k in range(args.kmax):
zc = zs[args.L - 1].detach().requires_grad_(True)
ce = F.cross_entropy(readout(zc).reshape(-1, vocab), y.reshape(-1))
d[args.L - 1] = (-beta * NBT * torch.autograd.grad(ce, zc)[0]).detach()
for l in range(args.L - 2, -1, -1):
zc = zs[l].detach().requires_grad_(True)
fnext = blocks[l + 1](zc, mask)
d[l] = torch.autograd.grad(fnext, zc, grad_outputs=d[l + 1])[0].detach()
delta = 0.0
with torch.no_grad():
prev = z0
for l in range(args.L):
rebuilt = blocks[l](prev, mask) + d[l]
new = (1 - args.geta) * zs[l] + args.geta * rebuilt if args.geta < 1.0 else rebuilt
delta += float((new - zs[l]).norm())
zs[l] = new
prev = zs[l]
if k == 0:
d1 = max(delta, 1e-12)
elif delta < 0.05 * d1 and k + 1 >= args.K:
return zs, k + 1, True
elif delta > 3.0 * d1:
return zs, k + 1, False
return zs, args.kmax, delta < 0.5 * d1
def dFdtheta(zs, x, y, beta):
"""dF/dtheta at fixed relaxed states (z0 rebuilt WITH graph so emb gets its E-path grad)."""
prev = tok(x) + pos(torch.arange(args.T, device=dev))[None]
E = 0.0
for z, b in zip(zs, blocks): E = E + 0.5 * ((z - b(prev, mask)) ** 2).sum(); prev = z
obj = E / NBT + beta * F.cross_entropy(readout(zs[-1]).reshape(-1, vocab), y.reshape(-1))
gs = torch.autograd.grad(obj, all_params, allow_unused=True)
return [g if g is not None else None for g in gs]
SIG0 = None
def ep_step(x, y):
"""single-sided adaptive EP: beta_t = beta0*sig0^2/sig_tok^2 (top-CE stiffness compensation),
adaptive-K fb relax, grad at the relaxed states. Returns (free_ce, beta_t, rounds, ok)."""
global SIG0
sig = tok_sigma()
if SIG0 is None: SIG0 = sig
beta_t = args.beta * (SIG0 * SIG0) / max(sig * sig, 1e-9)
z0, zs_free = free_states(x)
free_ce = F.cross_entropy(readout(zs_free[-1]).reshape(-1, vocab), y.reshape(-1)).item()
zp, rounds, ok = relax(z0, zs_free, y, +beta_t)
with torch.no_grad():
drift = sum(float((a - b).norm()) for a, b in zip(zp, zs_free)) / max(
sum(float(b.norm()) for b in zs_free), 1e-9)
if (not ok) or (not math.isfinite(drift)) or drift > 0.5:
for p in all_params: p.grad = None
return free_ce, beta_t, rounds, False
prev = tok(x) + pos(torch.arange(args.T, device=dev))[None]
E = 0.0
for z, b in zip(zp, blocks): E = E + 0.5 * ((z - b(prev, mask)) ** 2).sum(); prev = z
obj = E / (NBT * beta_t) + F.cross_entropy(readout(zp[-1]).reshape(-1, vocab), y.reshape(-1))
gs = torch.autograd.grad(obj, all_params, allow_unused=True)
for p, g in zip(all_params, gs):
p.grad = g
return free_ce, beta_t, rounds, True
def bp_gate(x, y):
"""true BP grads for telemetry cos (called before opt.step; reads p.grad separately)."""
z = tok(x) + pos(torch.arange(args.T, device=dev))[None]
for b in blocks: z = b(z, mask)
ce = F.cross_entropy(readout(z).reshape(-1, vocab), y.reshape(-1))
return list(torch.autograd.grad(ce, all_params, allow_unused=True))
@torch.no_grad()
def evaluate(nb=6):
tot = 0.0
for _ in range(nb):
x, y = get_batch('val')
z = tok(x) + pos(torch.arange(args.T, device=dev))[None]
for b in blocks: z = b(z, mask)
tot += F.cross_entropy(readout(z).reshape(-1, vocab), y.reshape(-1)).item()
return tot / nb
wb = None
if args.wandb:
try:
import wandb as _w
wb = _w.init(project=args.wandb, name=args.wandb_run or args.tag, id=args.wandb_run or args.tag,
resume='allow', config=vars(args))
except Exception as e:
print(f'[wandb] disabled ({e})', flush=True)
n = sum(p.numel() for p in all_params)
print(f'[{args.tag}] cascade-EP(EQUILIBRIUM/fb) L{args.L} C{args.C} T{args.T} beta={args.beta} '
f'K={args.K} geta={args.geta} | {n/1e6:.2f}M | {dev}', flush=True)
best, t0 = 1e9, time.time()
skips = 0
for step in range(args.steps + 1):
x, y = get_batch('train')
ce, beta_t, rounds, ok = ep_step(x, y)
if not ok: skips += 1
gcos = float('nan')
if step % args.gate_every == 0 and ok:
gbp = bp_gate(x, y)
num = den1 = den2 = 0.0
for p, g in zip(all_params, gbp):
if p.grad is None or g is None: continue
num += float((p.grad * g).sum()); den1 += float((p.grad ** 2).sum()); den2 += float((g ** 2).sum())
gcos = num / max((den1 ** 0.5) * (den2 ** 0.5), 1e-12)
torch.nn.utils.clip_grad_norm_(all_params, 1.0)
opt.step(); sched.step(); opt.zero_grad(set_to_none=True)
if step % args.log == 0:
val = evaluate(); best = min(best, val)
gtag = '' if math.isnan(gcos) else f' cos={gcos:.4f}'
print(f'step {step:5d}/{args.steps} | train {ce:.4f} val {val:.4f} (best {best:.4f}) '
f'| beta={beta_t:.2e} K={rounds} skips={skips}{gtag} | {step/max(time.time()-t0,1e-9):.3f} it/s', flush=True)
if wb is not None:
try: wb.log({'train_ce': ce, 'val_ce': val, 'best': best, 'beta_t': beta_t,
'rounds': rounds, 'skips': skips, 'gate_cos': (None if math.isnan(gcos) else gcos)}, step=step)
except Exception: pass
if step % args.save_every == 0 and step > 0:
torch.save({'tok': tok.state_dict(), 'pos': pos.state_dict(), 'blocks': blocks.state_dict(),
'step': step, 'val': best, 'config': vars(args)}, Path('runs') / f'{args.tag}_s{step}.pt')
print(f'[{args.tag}] DONE best val CE {best:.4f} (BP twin 2.9746; zil-diagnostic 3.3236)', flush=True)
if wb is not None:
try: wb.summary['best_val_ce'] = best; wb.finish()
except Exception: pass
|
