File size: 11,286 Bytes
13eabdd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
"""Task-vector weight-space columns: the informative version, for a suite with a shared base.

F8's weight-space family (`weight_cosine`, `subspace_overlap`, `spectral_overcounting`) was designed
for pairs that do NOT share a starting point. Every MergeBench family does share one -- its five
experts are fine-tunes of a single pretrained checkpoint -- so those columns sit at their analytic
ceiling: weight cosine 1.00000-1.00001, subspace overlap 0.9999-1.0000, spectral over-counting
-1.000000 (its exact A==B limit). None of them can carry a correlation.

The object every merge operator actually manipulates is the **task vector** tau = theta_expert -
theta_base, and task vectors are not saturated. This module recomputes the weight-space family on
them:

    tv_cosine                cos(tau_a, tau_b) over every shared floating tensor
    tv_norm_ratio            min(||tau||)/max(||tau||) -- are the two experts equally far from base?
    tv_qmd_raw               ||tau_a - tau_b|| / sqrt(||tau_a|| ||tau_b||); the numerator equals
                             theta_a - theta_b, but the normaliser is the task-vector scale rather
                             than the (near-identical) weight scale, so unlike `qmd_raw` it is not
                             dominated by how big the models are
    tv_subspace_overlap      `metrics.subspace_overlap` on the embedding task vectors
    tv_spectral_overcounting `metrics.spectral_overcounting` on the embedding task vectors -- this
                             is the column F8 wanted: over-counting between two genuine task
                             directions, not between two copies of the same base

Streamed key by key on the GPU straight out of the safetensors, so a 9B family never puts a float32
state dict in host RAM. Deliberately SEPARATE from the measurement sweep and idempotent, so it can
be retro-fitted onto families that already ran and had their expert weights deleted.

    PYTHONPATH=src python -m mergeschool.mergebench.taskvec --families gemma-2-2b
    bash scripts/run_taskvec.sh                       # every family with rows on disk
"""
from __future__ import annotations

import argparse
import gc
import json
import os
import shutil
import time

import numpy as np
import pandas as pd

from mergeschool import paths
from mergeschool.mergebench import suite as SU

OUT = paths.RESULTS / "mergebench"
log = lambda *a: print(f"[tv {time.strftime('%H:%M:%S')}]", *a, flush=True)  # noqa: E731

TV_COLS = ["tv_cosine", "tv_norm_ratio", "tv_qmd_raw", "tv_subspace_overlap",
           "tv_spectral_overcounting", "tv_norm_a", "tv_norm_b"]
ALLOW = ["*.safetensors", "*.json", "tokenizer*", "*.model"]


def _index(d):
    """{tensor key -> file} for one checkpoint directory."""
    from safetensors import safe_open
    idx = {}
    for f in sorted(x for x in os.listdir(d) if x.endswith(".safetensors")):
        with safe_open(os.path.join(d, f), framework="pt") as fh:
            for k in fh.keys():
                idx[k] = os.path.join(d, f)
    return idx


def compute_family(family, device="cuda", local=None, keep=False, doc=None):
    """{pair_id: {tv_* columns}} for one family. Downloads whatever `local` does not supply."""
    from huggingface_hub import snapshot_download
    from safetensors import safe_open
    import torch

    doc = doc or SU.enumerate_suite()
    experts = doc["families"][family]
    parent = SU.FAMILY_PARENT.get(family)
    if not parent:
        raise ValueError(f"no pretrained parent recorded for {family}")

    # DISK GATE. This pass runs alongside the measurement sweep, which itself holds up to two
    # families at once. Worst case measured: the sweep on two 8B families (161 GB) plus this pass on
    # gemma-2-9b and its base (111 GB) leaves ~337 GB -- BELOW the 350 GB floor. So wait for room
    # rather than race the sweep into it. `need` is this family's experts + base with 25% headroom.
    need_gb = (SU.family_bytes(family, doc) * 1.2 / 1e9) + 25.0
    floor = float(os.environ.get("MB_DISK_FLOOR_GB", "400"))
    waited = 0
    while True:
        free = shutil.disk_usage("/").free / 1e9
        if free - need_gb >= floor:
            break
        if waited == 0:
            log(f"  {family}: WAITING for disk -- need ~{need_gb:.0f} GB, free {free:.0f} GB, "
                f"floor {floor:.0f} GB")
        if waited > 10800:
            raise RuntimeError(f"disk never freed for {family}: free {free:.0f} GB, "
                               f"need {need_gb:.0f} GB above a {floor:.0f} GB floor")
        time.sleep(60)
        waited += 60
    if waited:
        log(f"  {family}: disk free after {waited//60} min wait")

    fetched = []
    local = dict(local or {})
    for dom, repo in sorted(experts.items()):
        if dom not in local:
            local[dom] = snapshot_download(repo, allow_patterns=ALLOW, max_workers=4)
            fetched.append(local[dom])
    base_dir = snapshot_download(parent, allow_patterns=ALLOW, max_workers=4)
    log(f"  {family}: base {parent} ready ({shutil.disk_usage('/').free/1e9:.0f} GB free)")

    doms = sorted(experts)
    idx = {d: _index(local[d]) for d in doms}
    bidx = _index(base_dir)
    keys = sorted(set(bidx).intersection(*[set(idx[d]) for d in doms]))

    dot = {(a, b): 0.0 for i, a in enumerate(doms) for b in doms[i + 1:]}
    sq = dict(dot)
    nrm = {d: 0.0 for d in doms}
    emb = {}
    nkeys = 0
    for k in keys:
        with safe_open(bidx[k], framework="pt") as fh:
            tb = fh.get_tensor(k)
        if not tb.is_floating_point():
            continue
        b = tb.to(device=device, dtype=torch.float32).reshape(-1)
        tau, ok = {}, True
        for d in doms:
            with safe_open(idx[d][k], framework="pt") as fh:
                t = fh.get_tensor(k)
            if t.shape != tb.shape:
                ok = False
                break
            tau[d] = t.to(device=device, dtype=torch.float32).reshape(-1) - b
        if not ok:
            del b
            continue
        nkeys += 1
        for d in doms:
            nrm[d] += float(tau[d].pow(2).sum())
        for (a, c) in dot:
            dot[(a, c)] += float(torch.dot(tau[a], tau[c]))
            sq[(a, c)] += float((tau[a] - tau[c]).pow(2).sum())
        # the embedding task vectors, kept for the two spectral columns
        if "embed_tokens" in k or k.endswith("wte.weight"):
            for d in doms:
                emb[d] = tau[d].reshape(tb.shape).cpu().numpy()
        del b, tau
    torch.cuda.empty_cache()
    log(f"  {family}: task vectors over {nkeys} shared tensors"
        + (f", embeddings {tuple(next(iter(emb.values())).shape)}" if emb else ", no embedding key"))

    from mergeschool.mergebench.sweep import spectral_pair
    out = {}
    for i, a in enumerate(doms):
        for c in doms[i + 1:]:
            na, nc = nrm[a] ** 0.5, nrm[c] ** 0.5
            row = {"tv_norm_a": na, "tv_norm_b": nc,
                   "tv_cosine": float(dot[(a, c)] / (na * nc)) if na * nc > 0 else np.nan,
                   "tv_norm_ratio": float(min(na, nc) / max(na, nc)) if max(na, nc) > 0 else np.nan,
                   "tv_qmd_raw": (float(np.sqrt(sq[(a, c)]) / np.sqrt(na * nc))
                                  if na * nc > 0 else np.nan)}
            if a in emb and c in emb:
                try:
                    sp = spectral_pair(emb[a], emb[c], device=device)
                    row["tv_subspace_overlap"] = sp["subspace_overlap"]
                    row["tv_spectral_overcounting"] = sp["spectral_overcounting"]
                except Exception as e:
                    log(f"  {family} {a}-{c}: spectral on task vectors failed "
                        f"({type(e).__name__}: {e})")
            out[f"{family}__{a}__{c}"] = row
    emb.clear()
    gc.collect()
    if not keep:
        for d in fetched:
            shutil.rmtree(os.path.dirname(os.path.dirname(d)), ignore_errors=True)
        shutil.rmtree(os.path.dirname(os.path.dirname(base_dir)), ignore_errors=True)
        log(f"  {family}: released weights ({shutil.disk_usage('/').free/1e9:.0f} GB free)")
    return out


CACHE = OUT / "taskvec_cache.json"


def cache_write(family, tv):
    """Persist a family's task-vector columns.

    Without this the pass is order-dependent in the worst way: it ran ahead of the measurement sweep,
    found no rows to merge into, and silently discarded 27 minutes of downloads and GPU work for six
    families. The values depend only on the checkpoints, never on whether the sweep has caught up,
    so they are written here and merged whenever rows appear.
    """
    doc = {}
    if CACHE.exists():
        try:
            doc = json.loads(CACHE.read_text())
        except Exception:
            doc = {}
    doc.update(tv)
    CACHE.parent.mkdir(parents=True, exist_ok=True)
    CACHE.write_text(json.dumps(doc, indent=1))
    log(f"  {family}: cached {len(tv)} pairs -> {CACHE.name} ({len(doc)} total)")


def merge_cached():
    """Merge every cached task-vector row into whichever shard now holds it. Idempotent."""
    if not CACHE.exists():
        return 0
    try:
        return merge_into_shards(json.loads(CACHE.read_text()))
    except Exception as e:
        log(f"  cache merge failed: {type(e).__name__}: {e}")
        return 0


def merge_into_shards(tv):
    """Write the tv_* columns into whichever pairs_w*.csv holds each pair. Idempotent."""
    n = 0
    for shard in sorted(OUT.glob("pairs_w*.csv")):
        d = pd.read_csv(shard)
        if "pair_id" not in d.columns:
            continue
        touched = False
        for col in TV_COLS:
            if col not in d.columns:
                d[col] = np.nan
        for i, pid in enumerate(d.pair_id):
            if pid in tv:
                for col, v in tv[pid].items():
                    d.at[i, col] = v
                touched = True
                n += 1
        if touched:
            d.to_csv(shard, index=False)
            log(f"  merged into {shard.name}")
    return n


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--families", nargs="*", default=None,
                    help="default: every family that already has rows on disk")
    ap.add_argument("--keep-weights", action="store_true")
    a = ap.parse_args()
    import torch
    device = "cuda" if torch.cuda.is_available() else "cpu"
    doc = SU.enumerate_suite()

    fams = a.families
    if not fams:
        have = set()
        for shard in sorted(OUT.glob("pairs_w*.csv")):
            d = pd.read_csv(shard)
            have |= set(d.family.dropna().unique())
        fams = [f for f in SU.families(doc) if f in have]
    log(f"task vectors for {fams}")
    for f in fams:
        t0 = time.time()
        try:
            tv = compute_family(f, device=device, keep=a.keep_weights, doc=doc)
            cache_write(f, tv)                    # persist BEFORE merging, never after
            n = merge_into_shards(tv)
            log(f"FAMILY {f}: {len(tv)} pairs, {n} rows updated in {time.time()-t0:.0f}s")
        except Exception as e:
            import traceback
            log(f"FAMILY {f} FAILED: {type(e).__name__}: {e}")
            log(traceback.format_exc().splitlines()[-1])
    log("done")


if __name__ == "__main__":
    main()