"""GSM plan-distillation control: KL from CoT-context teacher into FF adapter.

The width/depth law's falsification test: on code, distilling the model's own
plans into a feedforward adapter beat every recurrent variant. If the law is
right, the same recipe FAILS on GSM — sequential arithmetic cannot be
compressed into weight-space plan-priming. Interpretation caveat (recorded):
GSM answer spans are ~3 tokens vs ~100 for code, so supervision density is
also lower here; a null is law-consistent but not law-proving.
"""

import argparse
import json
import math
import os
import random
import sys
import time
from pathlib import Path

import torch
import torch.nn.functional as F

from loop_common import (COT_SUFFIX, DIRECT_SUFFIX, BandLooper, MergeAdapter,
                         chat_prompt, last_number, num_eq)
from prep_mbpp import batch_generate

sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from jlens.core import load_model  # noqa: E402

OUT = Path(os.environ.get("LOOP_OUT",
    Path(__file__).resolve().parent.parent / "results-loop"))
STEPS = 600
BATCH = 4
LR = 1e-3
WARMUP = 20
KL_T = 1.0


def lr_at(step):
    if step < WARMUP:
        return LR * (step + 1) / WARMUP
    t = (step - WARMUP) / max(1, STEPS - WARMUP)
    return 1e-4 + 0.5 * (LR - 1e-4) * (1 + math.cos(math.pi * t))


def ensure_cots(model, tok, items):
    path = OUT / "gsm_cots.json"
    if path.exists():
        return json.load(open(path))
    print(f"generating CoTs for {len(items)} items", flush=True)
    gens = batch_generate(model, tok,
                          [chat_prompt(tok, it["question"], COT_SUFFIX)
                           for it in items], max_new_tokens=320,
                          batch_size=32)
    cots = {}
    for it, g in zip(items, gens):
        if num_eq(last_number(g), it["gold"]):   # keep only correct traces
            cots[str(it["idx"])] = g
    json.dump(cots, open(path, "w"), indent=1)
    return cots


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--seed", type=int, default=0)
    args = ap.parse_args()
    rng = random.Random(args.seed)
    torch.manual_seed(args.seed)

    model, tok = load_model(dtype=torch.bfloat16)
    tok.padding_side = "left"
    for p in model.parameters():
        p.requires_grad_(False)
    looper = BandLooper(model)
    d = model.config.get_text_config().hidden_size
    adapter = MergeAdapter(d=d).cuda()
    opt = torch.optim.AdamW(adapter.parameters(), lr=LR, weight_decay=0.01)

    data = [it for it in json.load(open(OUT / "star_data.json"))
            if it["split"] == "train" and it["label"] != "drop"]
    cots = ensure_cots(model, tok, data)
    tok.padding_side = "right"
    train = [it for it in data if str(it["idx"]) in cots
             and len(tok(it["question"])["input_ids"]) <= 350]
    print(f"distill pool: {len(train)}", flush=True)

    pad = tok.pad_token_id or 0
    t0 = time.time()
    for step in range(STEPS):
        batch = rng.sample(train, BATCH)
        pairs = []
        for it in batch:
            a = tok(it["gold"] + "<end_of_turn>",
                    add_special_tokens=False)["input_ids"]
            sp = tok(chat_prompt(tok, it["question"], DIRECT_SUFFIX),
                     add_special_tokens=False)["input_ids"]
            tp = tok(chat_prompt(tok, it["question"], COT_SUFFIX)
                     + cots[str(it["idx"])] + "\n",
                     add_special_tokens=False)["input_ids"]
            pairs.append((sp, tp, a))
        Ts = max(len(sp) + len(a) for sp, _, a in pairs)
        Tt = max(len(tp) + len(a) for _, tp, a in pairs)
        s_ids = torch.full((BATCH, Ts), pad, dtype=torch.long)
        s_msk = torch.zeros((BATCH, Ts), dtype=torch.long)
        t_ids = torch.full((BATCH, Tt), pad, dtype=torch.long)
        t_msk = torch.zeros((BATCH, Tt), dtype=torch.long)
        lab = torch.full((BATCH, Ts), -100, dtype=torch.long)
        spans = []
        for i, (sp, tp, a) in enumerate(pairs):
            s_ids[i, : len(sp) + len(a)] = torch.tensor(sp + a)
            s_msk[i, : len(sp) + len(a)] = 1
            lab[i, len(sp): len(sp) + len(a)] = torch.tensor(a)
            t_ids[i, : len(tp) + len(a)] = torch.tensor(tp + a)
            t_msk[i, : len(tp) + len(a)] = 1
            spans.append((len(sp), len(tp), len(a)))
        s_ids, s_msk, lab = s_ids.cuda(), s_msk.cuda(), lab.cuda()
        t_ids, t_msk = t_ids.cuda(), t_msk.cuda()
        lmask = (lab == -100) & (s_msk == 1)

        with torch.no_grad():
            t_logits = model(input_ids=t_ids, attention_mask=t_msk,
                             use_cache=False).logits
        for g in opt.param_groups:
            g["lr"] = lr_at(step)
        s_logits = looper.loop_logits(adapter, s_ids, 1, attention_mask=s_msk,
                                      loop_mask=lmask, feedforward=True,
                                      use_checkpoint=True)
        kl = torch.zeros((), device="cuda")
        n_tok = 0
        for i, (ls, lt, la) in enumerate(spans):
            sl = s_logits[i, ls - 1: ls + la - 1].float()
            tl = t_logits[i, lt - 1: lt + la - 1].float()
            kl = kl + F.kl_div(F.log_softmax(sl / KL_T, -1),
                               F.log_softmax(tl / KL_T, -1),
                               log_target=True, reduction="sum")
            n_tok += la
        kl = kl / n_tok
        ce = F.cross_entropy(s_logits[:, :-1].flatten(0, 1).float(),
                             lab[:, 1:].flatten(), ignore_index=-100)
        loss = kl + 0.5 * ce
        opt.zero_grad(set_to_none=True)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(adapter.parameters(), 1.0)
        opt.step()
        if step % 10 == 0:
            print(f"step {step:4d}  kl={kl.item():.4f}  ce={ce.item():.4f}  "
                  f"({(time.time()-t0)/(step+1):.1f}s/step)", flush=True)
        if step % 200 == 199 or step == STEPS - 1:
            torch.save(adapter.state_dict(),
                       OUT / f"adapter_gsm_distill_e{step+1}.pt")
    print("done", flush=True)


if __name__ == "__main__":
    main()