File size: 11,157 Bytes
73dd4cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
train_qa_link.py -- train a question->answer latent bridge (the upgrade the panel's
"Tell Math a secret" demo points at).

Task: an arithmetic question ("23 + 54 =") is shown ONLY to the frozen Math/reasoning
specialist, which encodes it to its 256-d output latent. A NEW RecursiveLink + a
fine-tuned Language asker must emit the ANSWER ("077", zero-padded digits) reading
nothing but that latent: asker input is just "ANS> " + answer digits (teacher-forced
in training; decoded autoregressively at eval). 8% of (a, op, b) problems are HELD OUT
of training, so eval accuracy on them is generalization, not memorization. Ablating
the latent removes the question entirely -> accuracy collapses to the digit prior.

Saves links/qa__language__from__reasoning.safetensors in the same key style as the
key-recall bridge (link./ali./asker. + metadata) for moe_gradio.py to load.

Run:  python agents/modmind/train_qa_link.py  [--steps 4000] [--device cuda]
"""
from __future__ import annotations

import argparse
import hashlib
import json
import os
import random
import sys
import time

import torch
import torch.nn.functional as F

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)

from model import RecursiveLink, SpikeWhaleLM            # noqa: E402
from specialist_presets import specialist_config          # noqa: E402
from spike_tokenizer import SpikeTokenizer                # noqa: E402

ASKER, CONSULTANT = "language", "reasoning"
D_LATENT = 256
PROMPT = "ANS> "
ANS_LEN = 3                       # answers zero-padded to 3 digits ("077")
HOLDOUT_PCT = 8                   # % of problems held out of training entirely
OUT = os.path.join(HERE, "links", f"qa__{ASKER}__from__{CONSULTANT}.safetensors")


# ---- the problem space --------------------------------------------------------
def all_problems():
    """Every (a, op, b) the bridge is trained/evaluated on. Answers are 0..198."""
    probs = []
    for a in range(10, 100):
        for b in range(10, 100):
            probs.append((a, "+", b))
            if a >= b:
                probs.append((a, "-", b))
    for a in range(2, 13):
        for b in range(2, 13):
            probs.append((a, "*", b))
    return probs


def answer(a, op, b):
    return {"+": a + b, "-": a - b, "*": a * b}[op]


def is_holdout(a, op, b, pct):
    if pct <= 0:
        return False
    h = hashlib.md5(f"{a}{op}{b}".encode()).digest()[0]
    return h % 100 < pct


def render(a, op, b):
    return f"{a} {op} {b} ="


# ---- model loading (same pattern as moe_gradio.py) ------------------------------
def load_specialist(domain, device):
    from safetensors.torch import load_file
    ck = os.path.join(HERE, domain, "checkpoints", "model.safetensors")
    cfg = specialist_config(domain)
    m = SpikeWhaleLM(cfg).to(device)
    sd = load_file(ck, device=device)
    sd = {k: (v.float() if v.is_floating_point() else v) for k, v in sd.items()}
    m.load_state_dict(sd)
    tok = SpikeTokenizer(vocab_file=os.path.join(HERE, domain, "tokenizer.json"))
    return m, tok


# ---- training -------------------------------------------------------------------
def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--steps", type=int, default=4000)
    ap.add_argument("--batch", type=int, default=128)
    ap.add_argument("--link-lr", type=float, default=1e-3)
    ap.add_argument("--asker-lr", type=float, default=1e-4)
    ap.add_argument("--asker-wd", type=float, default=0.0)
    ap.add_argument("--holdout", type=int, default=0,
                    help="%% of problems held out of training (0 = train on ALL, the lookup-table demo)")
    ap.add_argument("--eval-every", type=int, default=200)
    ap.add_argument("--eval-n", type=int, default=256)
    ap.add_argument("--eval-chunk", type=int, default=64)     # keep eval VRAM peaks small
    ap.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu")
    ap.add_argument("--seed", type=int, default=0)
    ap.add_argument("--fresh", action="store_true", help="ignore last.pt and start over")
    args = ap.parse_args()
    dev = args.device
    random.seed(args.seed); torch.manual_seed(args.seed)

    print(f"[qa-link] device={dev}", flush=True)
    consultant, c_tok = load_specialist(CONSULTANT, dev)
    asker, a_tok = load_specialist(ASKER, dev)
    consultant.eval()
    for p in consultant.parameters():
        p.requires_grad_(False)

    # answer digits must be single tokens for the asker (position-aligned readout)
    digit_ids = []
    for d in "0123456789":
        ids = a_tok.encode(d, add_special_tokens=False)
        assert len(ids) == 1, f"digit {d!r} is not a single token: {ids}"
        digit_ids.append(ids[0])
    prompt_ids = a_tok.encode(PROMPT, add_special_tokens=False)
    plen = len(prompt_ids)
    print(f"[qa-link] prompt {PROMPT!r} = {plen} tokens; digits map to single tokens", flush=True)

    link = RecursiveLink(d_latent=D_LATENT).to(dev)
    opt = torch.optim.AdamW([
        {"params": list(link.parameters()), "lr": args.link_lr, "weight_decay": 0.0},
        {"params": list(asker.parameters()), "lr": args.asker_lr, "weight_decay": args.asker_wd},
    ])

    probs = all_problems()
    train_pool = [p for p in probs if not is_holdout(*p, args.holdout)]
    eval_pool = [p for p in probs if is_holdout(*p, args.holdout)]
    memorize = args.holdout <= 0
    if memorize:
        eval_pool = train_pool          # no holdout: "accuracy" = coverage of the whole table
        print(f"[qa-link] MEMORIZE mode: training on ALL {len(train_pool)} problems (no holdout)", flush=True)
    else:
        print(f"[qa-link] {len(train_pool)} train problems, {len(eval_pool)} held out", flush=True)
    label = "accuracy" if memorize else "held-out exact"

    @torch.no_grad()
    def encode_questions(batch):
        """Frozen consultant -> latents. Bucketed by token length (latent is a
        mean-pool over positions, so padding would corrupt it)."""
        idss = [c_tok.encode(render(*p), add_special_tokens=False) for p in batch]
        lat = torch.zeros(len(batch), D_LATENT, device=dev)
        by_len = {}
        for i, ids in enumerate(idss):
            by_len.setdefault(len(ids), []).append(i)
        for L, idx in by_len.items():
            c_ids = torch.tensor([idss[i] for i in idx], device=dev)
            lat[idx] = consultant(input_ids=c_ids).latent
        return lat

    def ans_tokens(p):
        return [digit_ids[int(ch)] for ch in f"{answer(*p):0{ANS_LEN}d}"]

    @torch.no_grad()
    def evaluate(pool, n, ablate=False):
        """Autoregressive 3-digit decode (full-vocab argmax, no teacher forcing).
        Chunked to keep VRAM peaks small."""
        asker.eval()
        sample = random.sample(pool, min(n, len(pool)))
        hit_e = hit_d = 0
        for o in range(0, len(sample), args.eval_chunk):
            chunk = sample[o:o + args.eval_chunk]
            lat = encode_questions(chunk)
            inj = torch.zeros_like(link(lat)) if ablate else link(lat)
            ids = torch.tensor([prompt_ids] * len(chunk), device=dev)
            for _ in range(ANS_LEN):
                logits = asker(input_ids=ids, inject_latent=inj).logits[:, -1, :]
                ids = torch.cat([ids, logits.argmax(-1, keepdim=True)], dim=1)
            pred = ids[:, plen:]
            tgt = torch.tensor([ans_tokens(p) for p in chunk], device=dev)
            hit_e += int((pred == tgt).all(dim=1).sum())
            hit_d += int((pred == tgt).sum())
        asker.train()
        return hit_e / len(sample), hit_d / (len(sample) * ANS_LEN)

    # resume from last.pt if a previous run died mid-flight
    last_pt = OUT + ".last.pt"
    best, start_step = -1.0, 0
    if os.path.exists(last_pt) and not args.fresh:
        st = torch.load(last_pt, map_location=dev, weights_only=False)
        link.load_state_dict(st["link"]); asker.load_state_dict(st["asker"])
        opt.load_state_dict(st["opt"]); best, start_step = st["best"], st["step"]
        print(f"[qa-link] resumed from step {start_step} (best held-out {best*100:.1f}%)", flush=True)

    t0 = time.time()
    asker.train()
    for step in range(start_step + 1, args.steps + 1):
        batch = random.sample(train_pool, args.batch)
        lat = encode_questions(batch)
        inj = link(lat)
        a_ids = torch.tensor([prompt_ids + ans_tokens(p) for p in batch], device=dev)
        labels = a_ids.clone()
        labels[:, :plen] = -100          # loss only on the answer digits
        out = asker(input_ids=a_ids, labels=labels, inject_latent=inj)
        opt.zero_grad(); out.loss.backward(); opt.step()

        if step % args.eval_every == 0 or step == args.steps:
            ex, pd = evaluate(eval_pool, args.eval_n)
            extra = "" if memorize else f"  train exact {evaluate(train_pool, args.eval_n)[0]*100:5.1f}%"
            print(f"[qa-link] step {step:5d}  loss {out.loss.item():.4f}  "
                  f"{label} {ex*100:5.1f}% (digits {pd*100:5.1f}%){extra}  "
                  f"[{time.time()-t0:.0f}s]", flush=True)
            if ex > best:
                best = ex
                save(link, asker, ex, step, args, memorize)
                print(f"[qa-link]   saved -> {OUT} ({label} {ex*100:.1f}%)", flush=True)
            # resume checkpoint every eval, so a crash never loses more than eval_every steps
            torch.save({"link": link.state_dict(), "asker": asker.state_dict(),
                        "opt": opt.state_dict(), "best": best, "step": step}, last_pt + ".tmp")
            os.replace(last_pt + ".tmp", last_pt)

    # final ablation numbers from the BEST saved bridge are written at save();
    # report the last-step ablation here for the log.
    ex_a, pd_a = evaluate(eval_pool, args.eval_n, ablate=True)
    print(f"[qa-link] ablated (latent cut): exact {ex_a*100:.1f}% / digits {pd_a*100:.1f}%", flush=True)
    print(f"[qa-link] done. best {label} {best*100:.1f}%", flush=True)


def save(link, asker, acc, step, args, memorize):
    from safetensors.torch import save_file
    os.makedirs(os.path.dirname(OUT), exist_ok=True)
    t = {}
    for k, v in link.state_dict().items():
        t["link." + k] = v.detach().to("cpu", torch.float16).contiguous()
    for k, v in asker.model.latent_inject.state_dict().items():
        t["ali." + k] = v.detach().to("cpu", torch.float16).contiguous()
    for k, v in asker.state_dict().items():
        t["asker." + k] = (v.detach().to("cpu", torch.float16).contiguous()
                           if v.is_floating_point() else v.detach().cpu().contiguous())
    tmp = OUT + ".tmp"
    save_file(t, tmp, metadata={
        "kind": "qa", "ans_len": str(ANS_LEN), "prompt": PROMPT,
        "asker": ASKER, "consultant": CONSULTANT,
        "mode": "memorize" if memorize else "generalize",
        "holdout_pct": str(args.holdout), "step": str(step),
        # accuracy over the whole table (memorize) or held-out set (generalize)
        "holdout_exact": f"{acc:.4f}",
        "ops": json.dumps(["+", "-", "*"]),
    })
    os.replace(tmp, OUT)   # atomic: the panel hot-reloads this file while we train


if __name__ == "__main__":
    main()