src/stage_b.py

#!/usr/bin/env python3
"""E3 Stage B: the FULL multitask mixture that defeated all seven prior models.

Same corpus, same objective mixture, same head system, same budget and the
same loss weights as the failed runs. The ONLY difference is the state
mechanism (addressed binding writes vs unstructured). Per the frozen
preregistration: no enum weighting, no continuation, no retention tricks, no
post-hoc tuning.

Parameter-matched: enc_layers=8 -> 24.27M vs PNSR's 24.61M (-1.4%, inside the
+/-3% policy).
"""
import argparse
import json
import sys
import time
from pathlib import Path

import numpy as np
import torch

sys.path.insert(0, str(Path(__file__).resolve().parent))
from pns.common import atomic_write_json, ckpt_root, eval_root, logs_root, shards_root  # noqa
from pns.model import modules  # noqa: E402
from pns.model.bind import BindConfig, PNSBind  # noqa: E402
from pns.train.loader import LifetimeBatcher  # noqa: E402

TBPTT = 32


def to_dev(b, dev):
    return {k: torch.from_numpy(np.ascontiguousarray(
        v.view(np.int64) if v.dtype == np.uint64 else v.astype(np.int64))).to(dev)
        for k, v in b.items()}


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--mode", choices=["bind", "unbound"], required=True)
    ap.add_argument("--run", required=True)
    ap.add_argument("--seed", type=int, default=1)
    ap.add_argument("--updates", type=int, default=6000)
    ap.add_argument("--batch", type=int, default=96)
    ap.add_argument("--lr", type=float, default=4e-4)
    args = ap.parse_args()
    dev = "cuda"
    torch.manual_seed(args.seed)
    cfg = BindConfig(d=384, heads=8, enc_layers=8, mode=args.mode, full_heads=True)
    model = PNSBind(cfg).to(dev)
    opt = torch.optim.AdamW(model.parameters(), lr=args.lr, betas=(0.9, 0.95),
                            weight_decay=0.01)
    rd = ckpt_root() / args.run
    rd.mkdir(parents=True, exist_ok=True)
    log = open(logs_root() / f"{args.run}.jsonl", "a")
    print(f"{args.run}: mode={args.mode} params="
          f"{sum(p.numel() for p in model.parameters())/1e6:.2f}M", flush=True)

    batcher = LifetimeBatcher("e3_train", shards_root(), args.batch, seed=args.seed)
    d = cfg.d
    u, t0 = 0, time.time()
    for lt in batcher:
        if u >= args.updates:
            break
        g = to_dev(lt, dev)
        B, L = g["etype"].shape
        state = model.initial_state(B, dev)
        for c0 in range(0, L, TBPTT):
            if u >= args.updates:
                break
            c1 = min(c0 + TBPTT, L)
            state = state.detach()
            for pg in opt.param_groups:
                pg["lr"] = args.lr * (min(1.0, (u + 1) / 300) if u < 300
                                      else (0.1 + 0.9 * max(0.0, 1 - (u - 300)
                                                            / max(1, args.updates - 300))))
            opt.zero_grad(set_to_none=True)
            outs = {k: [] for k in ("mode", "enum", "ptr", "op", "args")}
            with torch.autocast("cuda", dtype=torch.bfloat16):
                h_static = model.recenc.static(
                    g["rec_val_toks"], g["rec_key_toks"], g["rec_store"],
                    g["rec_kind"], g["rec_key"], g["rec_ent"])
                for t in range(c0, c1):
                    live = g["live"][:, t]
                    mask = live >= 0
                    rows = live.clamp(min=0)
                    bank = torch.gather(h_static, 1, rows.unsqueeze(-1).expand(-1, -1, d))
                    bank = model.recenc.finalize(
                        bank, (t - torch.gather(g["rec_birth"], 1, rows)).clamp(min=0))
                    state, out = model.step(
                        state, g["tok"][:, t], g["etype"][:, t], g["dt"][:, t],
                        g["bind_write"][:, t], g["bind_read"][:, t],
                        g["bind_slot_ent"], g["bind_slot_attr"],
                        rec_bank=bank * mask.unsqueeze(-1), live_mask=mask)
                    for k in outs:
                        outs[k].append(out[k])
                flat = {k: torch.cat([o.unsqueeze(1) for o in v], 1).flatten(0, 1)
                        for k, v in outs.items()}
                sup = {k: g[k][:, c0:c1].flatten(0, 1) for k in
                       ("mode_gold", "enum_gold", "ptr_gold_slot", "op_gold")}
                sup["enum_legal"] = g["enum_legal"][:, c0:c1].flatten(0, 1)
                sup["op_arg_slots"] = g["op_arg_slots"][:, c0:c1].flatten(0, 1)
                # IDENTICAL loss weights to the failed runs (w_enum default 2.0)
                L_parts = modules.losses(flat, sup)
                loss = sum(L_parts.values())
            loss.backward()
            gn = torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
            if torch.isfinite(loss):
                opt.step()
            u += 1
            if u % 200 == 0:
                s = {"u": u, "loss": round(float(loss), 4), "gn": round(float(gn), 2),
                     "s": round(time.time() - t0)}
                log.write(json.dumps(s) + "\n"); log.flush()
                print(json.dumps(s), flush=True)
    torch.save({"model": model.state_dict(), "cfg": vars(cfg), "args": vars(args)},
               rd / "final.pt")
    atomic_write_json(eval_root() / f"STAGEB_{args.run}_done.json",
                      {"run": args.run, "updates": u})
    print(f"done {u} updates", flush=True)


if __name__ == "__main__":
    main()