File size: 9,636 Bytes
47a719e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Did we try hard enough to make merging work?

The obvious objection to SET 1's negative result is that a plain average of aligned weights is a weak
merge: averaging two networks halves the variance of every pre-activation, and REPAIR (Jordan et al.,
ICLR 2023) shows that restoring those first- and second-order statistics recovers most of the
remaining barrier on vision nets. This adds that rung, training-free: after the permutation-aligned
average, walk the layers in order and affine-correct each Linear's pre-activations so their per-unit
mean and std match the average of the two parents' own statistics on the same corpus.

Rungs: M0 naive · M1 permutation-aligned average · M4 = M1 + REPAIR · M5 = M0 + REPAIR.
Metrics: Δfloor in nats/token (FLORES-200 eng devtest) AND BLiMP accuracy, on the same merges."""
import os, sys, json, time, glob, itertools, argparse, gc, csv
sys.path.insert(0, "/root/compose-audit")
from common import *
from transformers import AutoModelForCausalLM, AutoTokenizer
import pyarrow.parquet as pq

ap = argparse.ArgumentParser()
ap.add_argument("--size", default="14m")
ap.add_argument("--seeds", default="1,2,3,4,5,6,7,8,9")
ap.add_argument("--blocks", type=int, default=48)
ap.add_argument("--bs", type=int, default=16)
ap.add_argument("--n_per_paradigm", type=int, default=200)
ap.add_argument("--acts_rows", type=int, default=2048)
A = ap.parse_args()
SEEDS = [int(s) for s in A.seeds.split(",")]
OUT = f"/root/compose-audit/results/repair_{A.size}.jsonl"
DEV = "cuda"
BLIMP = glob.glob("/root/hf_cache_brainalign/hub/datasets--nyu-mll--blimp/snapshots/*/")[0]


def log(*a):
    print(f"[{time.strftime('%H:%M:%S')}]", *a, flush=True)


def neox_head_match(sd_a, sd_b, d, nh, nl):
    hd, perms = d // nh, {}
    for L in range(nl):
        qk = f"gpt_neox.layers.{L}.attention.query_key_value.weight"
        de = f"gpt_neox.layers.{L}.attention.dense.weight"
        if qk not in sd_a: continue
        Aq = np.asarray(sd_a[qk], float).reshape(nh, 3 * hd, d)
        Bq = np.asarray(sd_b[qk], float).reshape(nh, 3 * hd, d)
        g = np.einsum("ixy,jxy->ij", Aq, Bq)
        Ad = np.asarray(sd_a[de], float).reshape(d, nh, hd)
        Bd = np.asarray(sd_b[de], float).reshape(d, nh, hd)
        perms[L] = AL._assignment(g + np.einsum("xiy,xjy->ij", Ad, Bd))
    return perms


def neox_apply_head(sd, perms, d, nh):
    hd, out = d // nh, dict(sd)
    for L, h in perms.items():
        qk = f"gpt_neox.layers.{L}.attention.query_key_value.weight"
        qb = f"gpt_neox.layers.{L}.attention.query_key_value.bias"
        de = f"gpt_neox.layers.{L}.attention.dense.weight"
        out[qk] = np.asarray(sd[qk], float).reshape(nh, 3 * hd, d)[h].reshape(3 * d, d)
        if qb in sd: out[qb] = np.asarray(sd[qb], float).reshape(nh, 3 * hd)[h].reshape(3 * d)
        out[de] = np.asarray(sd[de], float).reshape(d, nh, hd)[:, h].reshape(d, d)
    return out


def align_perm(sd_a, sd_b, d, aa, ab, nh, nl):
    sd, _ = AL.align_weights_full(sd_a, sd_b, d, acts_a=aa, acts_b=ab, n_heads=None,
                                  method="permutation", strict=True, accept_each=True)
    hp = neox_head_match(sd_a, sd, d, nh, nl)
    if hp:
        cand = neox_apply_head(sd, hp, d, nh)
        if AL.block_normalised_distance(sd_a, cand) <= AL.block_normalised_distance(sd_a, sd):
            sd = cand
    return sd


# ------------------------------------------------------------------ REPAIR
TARGETS = ("mlp.dense_h_to_4h", "attention.query_key_value")


@torch.no_grad()
def preact_stats(model, blocks, dev, bs, nl):
    """{(layer, target): (mean, std)} of each Linear's OUTPUT (= pre-activation), per unit."""
    acc = {}
    hs = []

    def mk(key):
        def hook(mod, inp, out):
            o = out.detach().float().reshape(-1, out.shape[-1])
            s = acc.setdefault(key, [0.0, None, None])
            s[0] += o.shape[0]
            s[1] = o.sum(0) if s[1] is None else s[1] + o.sum(0)
            s[2] = (o * o).sum(0) if s[2] is None else s[2] + (o * o).sum(0)
        return hook

    for L in range(nl):
        blk = model.gpt_neox.layers[L]
        hs.append(blk.mlp.dense_h_to_4h.register_forward_hook(mk((L, "mlp.dense_h_to_4h"))))
        hs.append(blk.attention.query_key_value.register_forward_hook(mk((L, "attention.query_key_value"))))
    for i in range(0, blocks.shape[0], bs):
        model(blocks[i:i + bs].to(dev))
    for h in hs: h.remove()
    out = {}
    for k, (n, s1, s2) in acc.items():
        m = s1 / n
        v = (s2 / n - m * m).clamp_min(1e-12)
        out[k] = (m.cpu().numpy().astype(np.float64), v.sqrt().cpu().numpy().astype(np.float64))
    return out


def repair(sd_merged, stats_a, stats_b, shell, blocks, dev, bs, nl):
    """Walk layers in order; after fixing layers < L the inputs to layer L are already corrected, so
    layer L's own statistics are re-measured before it is corrected. Affine correction on the
    Linear's weight/bias, so the model stays exactly a model of the same architecture."""
    sd = {k: np.array(v, dtype=np.float64, copy=True) for k, v in sd_merged.items()}
    for L in range(nl):
        sd_load(shell, sd, dev)
        cur = preact_stats(shell, blocks, dev, bs, nl)
        for t in TARGETS:
            mu_t = 0.5 * (stats_a[(L, t)][0] + stats_b[(L, t)][0])
            sd_t = 0.5 * (stats_a[(L, t)][1] + stats_b[(L, t)][1])
            mu_m, sd_m = cur[(L, t)]
            g = sd_t / np.maximum(sd_m, 1e-8)
            wk, bk = f"gpt_neox.layers.{L}.{t}.weight", f"gpt_neox.layers.{L}.{t}.bias"
            sd[wk] = sd[wk] * g[:, None]
            sd[bk] = (sd[bk] - mu_m) * g + mu_t
    return sd


# ------------------------------------------------------------------ BLiMP
def load_blimp(n_per):
    out = []
    for d in sorted(glob.glob(BLIMP + "*/")):
        f = glob.glob(d + "*.parquet")
        if not f: continue
        t = pq.read_table(f[0]).to_pydict()
        out.append((os.path.basename(d.rstrip("/")), t["sentence_good"][:n_per], t["sentence_bad"][:n_per]))
    return out


tok = AutoTokenizer.from_pretrained(f"EleutherAI/pythia-{A.size}")
if tok.pad_token is None: tok.pad_token = tok.eos_token


def enc(sents, maxlen=48):
    e = tok(sents, return_tensors="pt", padding=True, truncation=True, max_length=maxlen)
    return e["input_ids"], e["attention_mask"]


ENC = [(n, enc(g), enc(b)) for n, g, b in load_blimp(A.n_per_paradigm)]
lines = flores_lines("eng_Latn")
blocks = make_blocks(tok, lines, block=512, max_blocks=A.blocks)
shell = AutoModelForCausalLM.from_pretrained(f"EleutherAI/pythia-{A.size}-seed{SEEDS[0]}",
                                             dtype=torch.float32).to(DEV).eval()
cfg = shell.config
D, NH, NL = cfg.hidden_size, cfg.num_attention_heads, cfg.num_hidden_layers
log(f"size={A.size} d={D} heads={NH} layers={NL} blimp_paradigms={len(ENC)}")


@torch.no_grad()
def bscore(ids, am):
    o = []
    for i in range(0, ids.shape[0], 128):
        x, m = ids[i:i + 128].to(DEV), am[i:i + 128].to(DEV)
        lp = torch.log_softmax(shell(x, attention_mask=m).logits.float()[:, :-1], -1)
        o.append((lp.gather(-1, x[:, 1:].unsqueeze(-1)).squeeze(-1) * m[:, 1:].float()).sum(1).cpu())
    return torch.cat(o).numpy()


def evaluate(sd):
    sd_load(shell, sd, DEV)
    nll = nll_nats(shell, blocks, DEV, bs=A.bs)
    cor = tot = 0
    for name, (gi, gm), (bi, bm) in ENC:
        sg, sb = bscore(gi, gm), bscore(bi, bm)
        cor += int((sg > sb).sum()); tot += len(sg)
    return nll, cor / tot


SDS, ACTS, PAR, STATS = {}, {}, {}, {}
for s in SEEDS:
    m = AutoModelForCausalLM.from_pretrained(f"EleutherAI/pythia-{A.size}-seed{s}", dtype=torch.float32).to(DEV).eval()
    SDS[s] = sd_np(m); ACTS[s] = capture_acts(m, blocks, DEV, n_rows=A.acts_rows, bs=A.bs)
    del m; torch.cuda.empty_cache()
    PAR[s] = evaluate(SDS[s])
    log(f"  seed{s} nll={PAR[s][0]:.4f} blimp={PAR[s][1]:.4f}")

done = set()
if os.path.exists(OUT):
    for l in open(OUT):
        try: done.add(tuple(json.loads(l)["pair"]))
        except Exception: pass
fh = open(OUT, "a")
for a, b in itertools.combinations(SEEDS, 2):
    if (a, b) in done: continue
    t0 = time.time()
    sa, sb = SDS[a], SDS[b]
    sbp = align_perm(sa, sb, D, ACTS[a], ACTS[b], NH, NL)
    sd_load(shell, sa, DEV); st_a = preact_stats(shell, blocks, DEV, A.bs, NL)
    sd_load(shell, sbp, DEV); st_b = preact_stats(shell, blocks, DEV, A.bs, NL)
    sd_load(shell, sb, DEV); st_braw = preact_stats(shell, blocks, DEV, A.bs, NL)
    rungs = {"M0_naive_avg": MG.average([sa, sb]), "M1_perm_avg": MG.average([sa, sbp])}
    rungs["M4_perm_repair"] = repair(rungs["M1_perm_avg"], st_a, st_b, shell, blocks, DEV, A.bs, NL)
    rungs["M5_naive_repair"] = repair(rungs["M0_naive_avg"], st_a, st_braw, shell, blocks, DEV, A.bs, NL)
    floor = min(PAR[a][0], PAR[b][0]); ceil = max(PAR[a][1], PAR[b][1])
    res = {}
    for k, sd in rungs.items():
        nll, acc = evaluate(sd)
        res[k] = {"nll": nll, "delta_floor": nll - floor, "blimp_acc": acc,
                  "blimp_delta_vs_ceiling": acc - ceil}
    r = {"set": "set1_repair", "size": A.size, "pair": [a, b], "floor": floor, "blimp_ceiling": ceil,
         "parent_nll": {"a": PAR[a][0], "b": PAR[b][0]},
         "parent_blimp": {"a": PAR[a][1], "b": PAR[b][1]}, "rungs": res, "secs": time.time() - t0}
    fh.write(json.dumps(r) + "\n"); fh.flush()
    log(f"pair {a},{b} floor={floor:.2f}/ceil={ceil:.3f} | " +
        " | ".join(f"{k}: {v['delta_floor']:+.2f}n {v['blimp_acc']:.3f}" for k, v in res.items()) +
        f" ({r['secs']:.0f}s)")
    del rungs, sbp; gc.collect()
fh.close()
log("DONE repair", A.size)