File size: 9,633 Bytes
1c22b92 4f9be59 1c22b92 4f9be59 1c22b92 4f9be59 1c22b92 4f9be59 1c22b92 4f9be59 1c22b92 4f9be59 1c22b92 4f9be59 1c22b92 | 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 | """SET 1: PolyPythia seed-merge. EleutherAI/pythia-{size}-seed{1..9}, C(9,2)=36 pairs.
Same data, same arch, same tokenizer; only the init/data-order seed varies -> the merge
obstruction is PURELY COORDINATE. Rungs M0 naive / M1 unit-aligned / task-arith / TIES.
Metric: Delta-floor in nats/token on a held-out corpus (FLORES-200 eng_Latn devtest)."""
import os, sys, json, time, itertools, argparse, gc
sys.path.insert(0, "/root/compose-audit")
from common import *
from transformers import AutoModelForCausalLM, AutoTokenizer
ap = argparse.ArgumentParser()
ap.add_argument("--size", default="14m")
ap.add_argument("--base", default=None, help="repo stem for tokenizer/pseudo-base; defaults to pythia-<size>")
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("--barrier_n", type=int, default=7)
ap.add_argument("--acts_rows", type=int, default=2048)
ap.add_argument("--tag", default="set1")
A = ap.parse_args()
BASE_STEM = A.base or f"pythia-{A.size}"
SEEDS = [int(s) for s in A.seeds.split(",")]
OUT = f"/root/compose-audit/results/{A.tag}_{A.size}.jsonl"
DEV = "cuda"
def log(*a):
print(f"[{time.strftime('%H:%M:%S')}]", *a, flush=True)
# ------------------------------------------------------------------ local: GPTNeoX fused-QKV head perm
def neox_head_match(sd_a, sd_b, d, n_heads, nlayer):
hd = d // n_heads
perms = {}
for L in range(nlayer):
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 or qk not in sd_b:
continue
A_ = np.asarray(sd_a[qk], float).reshape(n_heads, 3 * hd, d)
B_ = np.asarray(sd_b[qk], float).reshape(n_heads, 3 * hd, d)
gain = np.einsum("ixy,jxy->ij", A_, B_)
Ad = np.asarray(sd_a[de], float).reshape(d, n_heads, hd)
Bd = np.asarray(sd_b[de], float).reshape(d, n_heads, hd)
gain = gain + np.einsum("xiy,xjy->ij", Ad, Bd)
perms[L] = AL._assignment(gain)
return perms
def neox_apply_head(sd, perms, d, n_heads):
hd = d // n_heads
out = 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(n_heads, 3 * hd, d)[h].reshape(3 * d, d)
if qb in sd:
out[qb] = np.asarray(sd[qb], float).reshape(n_heads, 3 * hd)[h].reshape(3 * d)
out[de] = np.asarray(sd[de], float).reshape(d, n_heads, hd)[:, h].reshape(d, d)
return out
def align_full(sd_a, sd_b, d, acts_a, acts_b, n_heads, nlayer, method):
"""AL.align_weights_full (residual basis + free MLP axis) + a GPTNeoX fused-QKV head factor,
each accepted only if it does not increase the scale-free block-normalised distance."""
sd, info = AL.align_weights_full(sd_a, sd_b, d, acts_a=acts_a, acts_b=acts_b,
n_heads=None, method=method, strict=True, accept_each=True)
hp = neox_head_match(sd_a, sd, d, n_heads, nlayer)
if hp:
cand = neox_apply_head(sd, hp, d, n_heads)
if AL.block_normalised_distance(sd_a, cand) <= AL.block_normalised_distance(sd_a, sd):
sd, info["heads"] = cand, len(hp)
else:
info["rejected"].append("heads")
return sd, info
# ------------------------------------------------------------------ setup
tok = AutoTokenizer.from_pretrained(f"EleutherAI/{BASE_STEM}")
lines = flores_lines("eng_Latn")
blocks = make_blocks(tok, lines, block=512, max_blocks=A.blocks)
log(f"size={A.size} blocks={tuple(blocks.shape)} tokens={blocks.numel()}")
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"d={D} heads={NH} layers={NL} params={sum(p.numel() for p in shell.parameters())/1e6:.1f}M")
def ev(sd):
sd_load(shell, sd, DEV)
return nll_nats(shell, blocks, DEV, bs=A.bs)
SDS, NLL, ACTS = {}, {}, {}
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)
NLL[s] = nll_nats(m, blocks, DEV, bs=A.bs)
ACTS[s] = capture_acts(m, blocks, DEV, n_rows=A.acts_rows, bs=A.bs)
del m; torch.cuda.empty_cache()
log(f" seed{s} nll={NLL[s]:.4f}")
BASE = None
try:
mb = AutoModelForCausalLM.from_pretrained(f"EleutherAI/{BASE_STEM}", dtype=torch.float32).to(DEV).eval()
BASE = sd_np(mb); BASE_NLL = nll_nats(mb, blocks, DEV, bs=A.bs)
del mb; torch.cuda.empty_cache()
log(f" base({BASE_STEM}, NOT a shared ancestor of the seeds) nll={BASE_NLL:.4f}")
except Exception as e:
log("base load failed:", e)
KEYS = shared_keys(SDS[SEEDS[0]], SDS[SEEDS[1]])
done = set()
if os.path.exists(OUT):
for line in open(OUT):
try: done.add(tuple(json.loads(line)["pair"]))
except Exception: pass
log(f"resuming: {len(done)} pairs already done")
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]
r = {"set": "set1_polypythia", "size": A.size, "pair": [a, b],
"parent_nll": {"a": NLL[a], "b": NLL[b]}, "floor": min(NLL[a], NLL[b]),
"corpus": "flores200_devtest_eng_Latn", "metric": "nats_per_token"}
# ---------------- alignments (fitted BEFORE any merge) ----------------
sb_perm, info_p = align_full(sa, sb, D, ACTS[a], ACTS[b], NH, NL, "permutation")
sb_orth, info_o = align_full(sa, sb, D, ACTS[a], ACTS[b], NH, NL, "orthogonal")
r["align_info"] = {"perm": info_p, "orth": info_o}
# ---------------- predictors (pre-merge) ----------------
fa, fb = flat(sa, KEYS), flat(sb, KEYS)
p = {"weight_cosine": float(fa @ fb / (np.linalg.norm(fa) * np.linalg.norm(fb)))}
p["weight_cosine_bn"] = float(np.mean([
float(np.asarray(sa[k], float).ravel() @ np.asarray(sb[k], float).ravel() /
(np.linalg.norm(sa[k]) * np.linalg.norm(sb[k]) + 1e-12)) for k in KEYS]))
qwd_p = MT.quotient_weight_distance(sa, sb, sb_perm, KEYS)
qwd_o = MT.quotient_weight_distance(sa, sb, sb_orth, KEYS)
p.update({"d_raw": qwd_p["d_raw"], "qmd_perm": qwd_p["qmd"],
"coord_share_perm": qwd_p["coord_fraction"], "norm_ratio_perm": qwd_p["norm_ratio"],
"qmd_orth": qwd_o["qmd"], "coord_share_orth": qwd_o["coord_fraction"]})
for tag, extra in (("perm", qwd_p), ("orth", qwd_o)):
for k in ("d_raw_bn", "qmd_bn", "coordinate_gap_bn", "coord_fraction_bn"):
if k in extra:
p[f"{k}_{tag}"] = extra[k]
bnd_raw = AL.block_normalised_distance(sa, sb, KEYS)
bnd_p = AL.block_normalised_distance(sa, sb_perm, KEYS)
bnd_o = AL.block_normalised_distance(sa, sb_orth, KEYS)
p.update({"bnd_raw": bnd_raw, "bnd_perm": bnd_p, "bnd_orth": bnd_o,
"coord_share_bnd_perm": float((bnd_raw - bnd_p) / bnd_raw),
"coord_share_bnd_orth": float((bnd_raw - bnd_o) / bnd_raw)})
ck, ck_by = mean_cka(ACTS[a], ACTS[b])
p["cka_mean"] = ck; p["cka_last"] = ck_by[max(ck_by)]
mid = NL // 2
for g in ("perm", "procrustes", "ot"):
try:
qr = MT.quotient_residual(ACTS[a][mid], ACTS[b][mid], group=g)
p[f"qmd_act_{g}"] = qr["distance"]; p[f"aligned_cka_{g}"] = qr["aligned_cka"]
except Exception as e:
p[f"qmd_act_{g}"] = float("nan")
if BASE is not None:
ta = flat({k: sa[k] - BASE[k] for k in KEYS}, KEYS)
tb = flat({k: sb[k] - BASE[k] for k in KEYS}, KEYS)
p["task_vector_cosine"] = float(ta @ tb / (np.linalg.norm(ta) * np.linalg.norm(tb)))
r["predictors"] = p
# ---------------- merge rungs ----------------
rungs = {}
rungs["M0_naive_avg"] = MG.average([sa, sb])
rungs["M1_perm_avg"] = MG.average([sa, sb_perm])
rungs["M1_orth_avg"] = MG.average([sa, sb_orth])
if BASE is not None:
rungs["M2_task_arith"] = MG.task_arithmetic(BASE, [sa, sb])
try:
rungs["M3_ties"] = MG.ties(BASE, [sa, sb], density=0.2)
except Exception as e:
log("ties failed", e)
res = {}
for name, sd in rungs.items():
n = ev(sd)
res[name] = {"nll": n, "delta_floor": n - r["floor"],
"delta_vs_naive": None}
for name in res:
res[name]["delta_vs_naive"] = res[name]["nll"] - res["M0_naive_avg"]["nll"]
r["rungs"] = res
# ---------------- barriers ----------------
try:
bn = EV.merge_barrier(sa, sb, ev, n=A.barrier_n)
r["barrier_naive"] = {"barrier": bn["barrier"], "losses": list(map(float, bn["losses"]))}
bp = EV.merge_barrier(sa, sb_perm, ev, n=A.barrier_n)
r["barrier_perm"] = {"barrier": bp["barrier"], "losses": list(map(float, bp["losses"]))}
except Exception as e:
log("barrier failed", e)
r["secs"] = time.time() - t0
fh.write(json.dumps(r) + "\n"); fh.flush()
log(f"pair {a},{b} floor={r['floor']:.3f} M0={res['M0_naive_avg']['delta_floor']:+.3f} "
f"M1perm={res['M1_perm_avg']['delta_floor']:+.3f} M1orth={res['M1_orth_avg']['delta_floor']:+.3f} "
f"({r['secs']:.0f}s)")
del rungs, sb_perm, sb_orth; gc.collect()
fh.close()
log("DONE", A.size)
|