File size: 7,234 Bytes
4fc906a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import argparse
import random
import sys
import time
from pathlib import Path

import torch

sys.path.insert(0, str(Path(__file__).resolve().parent))
from model import (  
    make_reduce_cell, make_add_cell, reduce_features, add_features,
    BitStreamMachine, _bits_of, PAD_HEAD,
)
from data import make_reduce_batch, make_add_batch  


def is_probable_prime(n: int, rng: random.Random, rounds: int = 24) -> bool:
    if n < 2:
        return False
    for sp in (2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37):
        if n % sp == 0:
            return n == sp
    d, r = n - 1, 0
    while d % 2 == 0:
        d //= 2
        r += 1
    for _ in range(rounds):
        a = rng.randrange(2, n - 1)
        x = pow(a, d, n)
        if x in (1, n - 1):
            continue
        for _ in range(r - 1):
            x = x * x % n
            if x == n - 1:
                break
        else:
            return False
    return True


def random_prime(rng: random.Random, lo_bits: int, hi_bits: int) -> int:
    while True:
        l = rng.randint(lo_bits, hi_bits)
        if l == 1:
            return 2
        p = rng.getrandbits(l - 1) | (1 << (l - 1)) | 1
        if p == 1:
            continue
        if is_probable_prime(p, rng):
            return p


def load_cells(path: str):
    ck = torch.load(path, map_location="cpu", weights_only=True)
    rcell = make_reduce_cell()
    rcell.load_state_dict(ck.get("reduce_ema_state_dict", ck["reduce_state_dict"]))
    rcell.eval()
    acell = make_add_cell()
    acell.load_state_dict(ck.get("add_ema_state_dict", ck["add_state_dict"]))
    acell.eval()
    return rcell, acell


@torch.no_grad()
def cell_stress(rcell, acell, rng, widths, per_width):
    print("== cell-level stress (hard gates, CPU) ==")
    worst = 1.0
    for n in widths:
        bsz = max(8, 20000 // n)
        for kind, cell in (("reduce", rcell), ("add", acell)):
            total, good = 0, 0
            while total < per_width:
                if kind == "reduce":
                    b = make_reduce_batch(rng, n, bsz)
                    feats = reduce_features(b["x"], b["p"], b["p3"])
                else:
                    b = make_add_batch(rng, n, bsz)
                    feats = add_features(b["x"], b["y"], b["g"])
                ok = ((cell(feats) > 0) == (b["z"] > 0.5)).all(dim=1)
                good += int(ok.sum())
                total += ok.numel()
            rate = good / total
            worst = min(worst, rate)
            print(f"  n={n:5d} {kind:6s}: {good}/{total}  ({rate:.6f})"
                  + ("  <-- FAILURES" if rate < 1 else ""))
    return worst


def _pack(vals, w):
    m = torch.zeros(len(vals), w)
    for r, v in enumerate(vals):
        bits = _bits_of(v)
        m[r, w - len(bits):] = torch.tensor(bits, dtype=torch.float32)
    return m


@torch.no_grad()
def machine_tier_check(rcell, acell, rng, per_tier, tiers=None):
    print("== machine-level end-to-end vs ground truth ==")
    mach = BitStreamMachine(rcell, acell, torch.device("cpu"))
    geo = {1: (1, 3, 32), 2: (4, 8, 48), 3: (9, 16, 64), 4: (17, 32, 96),
           5: (33, 64, 128), 6: (65, 128, 256), 7: (129, 256, 512),
           8: (257, 512, 1024), 9: (513, 1024, 2048)}
    if tiers:
        geo = {t: geo[t] for t in tiers}
    results = {}
    for t, (lo, hi, ob) in geo.items():
        primes = [random_prime(rng, lo, hi) for _ in range(5)]
        probs = []
        for i in range(per_tier):
            p = primes[i % 5]
            if i < 4:
                a, b = [(0, rng.getrandbits(ob)), (rng.getrandbits(ob), 0),
                        (1, rng.getrandbits(ob)), (rng.getrandbits(ob), 1)][i]
            else:
                a = rng.getrandbits(rng.randint(1, ob))
                b = rng.getrandbits(rng.randint(1, ob))
            probs.append((a, b, p))
        n_p = max(p.bit_length() for _, _, p in probs) + PAD_HEAD
        L = max(2, max(max(a.bit_length(), b.bit_length())
                       for a, b, _ in probs))
        L += L % 2
        t0 = time.time()
        z = mach.run(_pack([a for a, _, _ in probs], L),
                     _pack([b for _, b, _ in probs], L),
                     _pack([p for _, _, p in probs], n_p),
                     _pack([3 * p for _, _, p in probs], n_p))
        dt = time.time() - t0
        good = 0
        for r, (a, b, p) in enumerate(probs):
            got = int("".join(str(int(v)) for v in z[r].tolist()), 2)
            good += (got == (a * b) % p)
        results[t] = (good, per_tier, dt)
        print(f"  tier {t}: {good}/{per_tier} exact   ({dt:.1f}s for batch)")
    return results


@torch.no_grad()
def batch_invariance(rcell, acell, rng):
    print("== batch-composition invariance ==")
    mach = BitStreamMachine(rcell, acell, torch.device("cpu"))
    p = random_prime(rng, 60, 64)
    a, b = rng.getrandbits(128), rng.getrandbits(128)

    def run_with(width_pad, extra_p_bits):
        n_p = p.bit_length() + PAD_HEAD + extra_p_bits
        L = 128 + width_pad
        L += L % 2
        z = mach.run(_pack([a], L), _pack([b], L),
                     _pack([p], n_p), _pack([3 * p], n_p))
        return int("".join(str(int(v)) for v in z[0].tolist()), 2)

    vals = {run_with(wp, ep) for wp in (0, 8, 32) for ep in (0, 5, 40)}
    ok = len(vals) == 1 and vals == {(a * b) % p}
    print(f"  distinct outputs across paddings: {len(vals)} (want 1), "
          f"correct={ok}")
    return ok


@torch.no_grad()
def perturbation(rcell, acell, rng):
    print("== weight-perturbation collapse (compliance evidence) ==")
    import copy
    for scale in (0.0, 0.02, 0.1):
        r2, m2 = copy.deepcopy(rcell), copy.deepcopy(acell)
        if scale:
            for c in (r2, m2):
                for prm in c.parameters():
                    prm.add_(torch.randn_like(prm) * scale
                             * (prm.abs().mean() + 1e-8))
        res = machine_tier_check(r2, m2, random.Random(7), 20, tiers=[3])
        g, n, _ = res[3]
        print(f"  noise scale {scale}: tier-3 accuracy {g}/{n}")


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("ckpt")
    ap.add_argument("--quick", action="store_true")
    ap.add_argument("--tiers", type=int, nargs="*", default=None)
    ap.add_argument("--per-tier", type=int, default=100)
    ap.add_argument("--per-width", type=int, default=200_000)
    ap.add_argument("--seed", type=int, default=123)
    ap.add_argument("--skip-stress", action="store_true")
    args = ap.parse_args()

    torch.set_num_threads(8)
    rng = random.Random(args.seed)
    rcell, acell = load_cells(args.ckpt)

    widths = [5, 6, 8, 11, 15, 19, 23, 27, 31, 35, 36,
              ]
    per_width = 20_000 if args.quick else args.per_width
    per_tier = 30 if args.quick else args.per_tier

    worst = 1.0
    if not args.skip_stress:
        worst = cell_stress(rcell, acell, rng, widths, per_width)
    machine_tier_check(rcell, acell, rng, per_tier, tiers=args.tiers)
    inv = batch_invariance(rcell, acell, rng)
    if not args.quick:
        perturbation(rcell, acell, rng)
    print(f"\nworst cell width rate: {worst:.6f};  batch-invariant: {inv}")


if __name__ == "__main__":
    main()