add train.py for VPS SFT

d57e955 verified 14 days ago

22.2 kB

#!/usr/bin/env python3
"""SN38 full-parameter continued pretraining (CPT) / SFT for sn38-nanochrono.
Run from the sn38 repo env so the architecture is registered:
  cd /opt/sn38 && PYTHONPATH=/opt/sn38 uv run python /opt/work/train.py --mode cpt ...
What is implemented (see README section "What the research says"):
  * CPT: sources sampled by --mix; documents packed BEST-FIT (whole docs, no needless truncation),
    each sequence gets a block-diagonal DOCUMENT MASK so tokens never attend across documents (--doc-mask).
  * Loss computed in chunks (never materialises the full fp32 [B,T,32768] logits), -100 labels respected.
  * LR: linear warm-up then cosine re-decay to 10% (re-warm + re-decay recipe for CPT).
  * Held-out validation per source group (new-2024 vs replay) and a probe set (year,sentence,phrase csv)
    scored exactly like the validator's leak.py (BOS+prompt, phrase with leading space, sum log-prob, epsilon).
  * SFT: loss only on response + EOS (position-based mask, so EOS==PAD is NOT a problem), asserts the EOS is a target.
  * bf16 autocast with fp32 master weights, fused AdamW, grad clip, optional activation checkpointing / torch.compile.
Saves HF layout (config.json, model.safetensors [bf16], tokenizer, generation_config.json) + a symlink `latest`.
"""
import argparse, csv, glob, gzip, json, math, os, random, time
# Multi-GPU: torchrun --standalone --nproc_per_node 8 train.py ... (data parallel, grads averaged after accumulation;
# --global-tokens is the TOTAL per optimizer step across GPUs). --resume restarts from <out>/resume.pt (spot instances).
import torch, torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
import sn38.architectures  # noqa: registers sn38-nanochrono / nanoexpand / miniformer
sn38.architectures.register_all()

IGN = -100


# ------------------------------------------------------------------ data: sources
def docs_from_source(src_dir, seed, skip=0):
    """Infinite shuffled stream of texts; the first `skip` docs of the first shard are held out (validation)."""
    files = sorted(glob.glob(os.path.join(src_dir, "part-*.jsonl.gz")))
    if not files:
        raise SystemExit(f"no shards in {src_dir}")
    rng = random.Random(seed)
    epoch = 0
    while True:
        order = files[:]; rng.shuffle(order)
        for f in order:
            with gzip.open(f, "rt", encoding="utf-8") as fh:
                for i, line in enumerate(fh):
                    if f == files[0] and i < skip:
                        continue
                    yield json.loads(line)["text"]
        epoch += 1
        print(f"[data] {os.path.basename(src_dir)} epoch {epoch} done", flush=True)


def heldout_docs(src_dir, n):
    files = sorted(glob.glob(os.path.join(src_dir, "part-*.jsonl.gz")))
    out = []
    with gzip.open(files[0], "rt", encoding="utf-8") as fh:
        for i, line in enumerate(fh):
            if i >= n: break
            out.append(json.loads(line)["text"])
    return out


# ------------------------------------------------------------------ data: best-fit packing + doc mask
class Packer:
    """Best-fit-decreasing packing of whole documents into seq+1 windows.
    Docs longer than seq+1 are split into seq+1 chunks (unavoidable); everything else is never truncated.
    Yields (ids[B,T+1], labels[B,T+1], docid[B,T+1]) ; pad positions have label -100."""

    def __init__(self, tok, seq, micro_bs, buffer_docs=512):
        self.tok, self.L, self.B, self.buf_n = tok, seq + 1, micro_bs, buffer_docs
        self.bos, self.eos, self.pad = tok.bos_token_id, tok.eos_token_id, tok.pad_token_id

    def _chunks(self, texts):
        for ids in self.tok(texts, add_special_tokens=False)["input_ids"]:
            ids = [self.bos] + ids + [self.eos]
            for i in range(0, len(ids), self.L):
                yield ids[i:i + self.L]

    def pack(self, chunks):
        chunks.sort(key=len, reverse=True)
        bins = []  # list of [free, [chunk,...]]
        for c in chunks:
            best = None
            for b in bins:
                if b[0] >= len(c) and (best is None or b[0] < best[0]):
                    best = b
            if best is None:
                bins.append([self.L - len(c), [c]])
            else:
                best[0] -= len(c); best[1].append(c)
        return bins

    def stream(self, doc_iter):
        carry = []
        while True:
            texts = [next(doc_iter) for _ in range(self.buf_n)]
            bins = self.pack(carry + list(self._chunks(texts)))
            bins.sort(key=lambda b: b[0])              # fullest bins first
            keep = len(bins) - len(bins) % self.B
            carry = [c for b in bins[keep:] for c in b[1]]   # sparse leftovers are re-packed next round
            for i in range(0, keep, self.B):
                yield self._tensorise(bins[i:i + self.B])

    def _tensorise(self, bins):
        ids = torch.full((self.B, self.L), self.pad, dtype=torch.long)
        lab = torch.full((self.B, self.L), IGN, dtype=torch.long)
        doc = torch.zeros((self.B, self.L), dtype=torch.long)
        for r, (_, chunks) in enumerate(bins):
            p = 0
            for d, c in enumerate(chunks, start=1):
                ids[r, p:p + len(c)] = torch.tensor(c); lab[r, p:p + len(c)] = torch.tensor(c)
                doc[r, p:p + len(c)] = d; p += len(c)
        return ids, lab, doc


def mixed_docs(corpus, mix, seed, skip):
    names = list(mix); weights = [mix[n] for n in names]
    gens = {n: docs_from_source(os.path.join(corpus, n), seed + i, skip) for i, n in enumerate(names)}
    rng = random.Random(seed)
    while True:
        yield next(gens[rng.choices(names, weights)[0]])


_DOC_MASK = {"m": None}


def install_doc_mask(model):
    """Patch nanochrono attention so a block-diagonal document mask is AND-ed with causal/sliding masks."""
    import sn38.architectures.nanochrono.modeling_nanochrono as mn
    orig = mn.NanochronoAttention._attend

    def _attend(self, q, k, v, window, mask, past_len):
        dm = _DOC_MASK["m"]
        if dm is None or q.size(2) != k.size(2):
            return orig(self, q, k, v, window, mask, past_len)
        T = q.size(2)
        row = torch.arange(T, device=q.device).unsqueeze(1); col = torch.arange(T, device=q.device).unsqueeze(0)
        m = col <= row
        if 0 <= window < T:
            m = m & ((row - col) <= window)
        m = m.unsqueeze(0).unsqueeze(0) & dm
        return F.scaled_dot_product_attention(q, k, v, attn_mask=m)
    mn.NanochronoAttention._attend = _attend


def doc_mask_from_ids(doc):  # doc [B,T] ints, 0 = padding -> [B,1,T,T] bool
    m = (doc.unsqueeze(2) == doc.unsqueeze(1)) & (doc.unsqueeze(2) > 0)
    eye = torch.eye(doc.size(1), dtype=torch.bool, device=doc.device).unsqueeze(0)
    return (m | eye).unsqueeze(1)  # padding rows attend to themselves so softmax is defined


# ------------------------------------------------------------------ data: SFT
def sft_batches(path, tok, seq, micro_bs, seed, shard=(0, 1)):
    """path may be a comma list of jsonl files (prompt/response rows). The response is tokenized with ONE leading space
    so the Q->A boundary matches natural text ("...? The answer") -- at eval the model continues straight after the raw
    question with no separator, so whatever boundary we train is what it will emit."""
    rows = []
    for f in path.split(","):
        rows += [json.loads(l) for l in open(f, encoding="utf-8")]
    rows = rows[shard[0]::shard[1]]                     # each DDP rank owns a disjoint slice
    if shard[0] == 0:
        print(f"[sft] {len(rows) * shard[1]} rows from {path} ({shard[1]} shard(s))", flush=True)
    rng = random.Random(seed); bos, eos, pad = tok.bos_token_id, tok.eos_token_id, tok.pad_token_id
    while True:
        rng.shuffle(rows)
        for i in range(0, len(rows) - micro_bs + 1, micro_bs):
            xs, ys = [], []
            for r in rows[i:i + micro_bs]:
                p = [bos] + tok(r["prompt"], add_special_tokens=False)["input_ids"]
                a = tok(" " + r["response"].strip(), add_special_tokens=False)["input_ids"] + [eos]
                if len(p) + len(a) > seq + 1:            # never cut the EOS off: trim the prompt from the left
                    p = p[-(seq + 1 - len(a)):] if len(a) < seq + 1 else p[:1]
                    a = a[: seq + 1 - len(p)]
                ids = p + a; lab = [IGN] * len(p) + a
                assert lab[-1] == eos, "EOS must be a training target"
                padn = seq + 1 - len(ids)
                xs.append(ids + [pad] * padn); ys.append(lab + [IGN] * padn)
            x, y = torch.tensor(xs), torch.tensor(ys)
            yield x, y, None


# ------------------------------------------------------------------ loss (chunked, softcapped, -100 aware)
def lm_loss(model, ids, labels, chunk=512):
    """ids/labels [B,T+1]. Returns (sum_nll, n_targets). Only the hidden states are stored; logits are
    recomputed per chunk in backward (checkpoint), so the fp32 [B,T,V] tensor never exists at once."""
    hidden = model.model(input_ids=ids[:, :-1], use_cache=False).last_hidden_state
    tgt = labels[:, 1:]
    cap = model.config.final_logit_softcapping
    W = model.lm_head.weight

    def piece(h, t):
        lg = F.linear(h, W).float()
        if cap:
            lg = cap * torch.tanh(lg / cap)
        return F.cross_entropy(lg.view(-1, lg.size(-1)), t.reshape(-1), ignore_index=IGN, reduction="sum")

    total = hidden.new_zeros((), dtype=torch.float32)
    for s in range(0, hidden.size(1), chunk):
        total = total + checkpoint(piece, hidden[:, s:s + chunk], tgt[:, s:s + chunk], use_reentrant=False)
    return total, (tgt != IGN).sum()


# ------------------------------------------------------------------ evaluation
@torch.no_grad()
def eval_loss(model, tok, texts, seq, dev, bs=8):
    model.eval(); tot = 0.0; n = 0
    for i in range(0, len(texts), bs):
        enc = tok(texts[i:i + bs], add_special_tokens=False, truncation=True, max_length=seq)["input_ids"]
        L = max(len(e) for e in enc) + 1
        ids = torch.full((len(enc), L), tok.pad_token_id); lab = torch.full((len(enc), L), IGN)
        for r, e in enumerate(enc):
            e = [tok.bos_token_id] + e
            ids[r, :len(e)] = torch.tensor(e); lab[r, :len(e)] = torch.tensor(e)
        with torch.autocast(dev.type, dtype=torch.bfloat16, enabled=dev.type == "cuda"):
            s, k = lm_loss(model, ids.to(dev), lab.to(dev))
        tot += s.item(); n += k.item()
    model.train()
    return tot / max(1, n)


def load_probes(path):
    rows = list(csv.DictReader(open(path, encoding="utf-8")))
    out = []
    for r in rows:
        s, p = r["sentence"], r["phrase"]
        i = s.lower().rfind(p.lower())
        out.append({"year": int(r["year"]), "prompt": (s[:i].rstrip() if i >= 0 else s), "phrase": p})
    return out


@torch.no_grad()
def probe_scores(model, tok, probes, dev):
    """Sum of log-probs of the phrase tokens after BOS+prompt — identical arithmetic to sn38.template.leak."""
    model.eval(); scores = []
    for it in probes:
        p = [tok.bos_token_id] + tok(it["prompt"], add_special_tokens=False)["input_ids"]
        ph = tok(" " + it["phrase"], add_special_tokens=False)["input_ids"]
        ids = torch.tensor([p + ph], device=dev)
        with torch.autocast(dev.type, dtype=torch.bfloat16, enabled=dev.type == "cuda"):
            lg = model(input_ids=ids, use_cache=False).logits.float()
        lp = F.log_softmax(lg[0, len(p) - 1:len(p) - 1 + len(ph)], -1)
        scores.append(sum(lp[j, t].item() for j, t in enumerate(ph)))
    model.train()
    return scores


def run_eval(model, tok, dev, val_sets, probes, seq, epsilon, log):
    msg = []
    for name, texts in val_sets.items():
        msg.append(f"val/{name}={eval_loss(model, tok, texts, seq, dev):.4f}")
    if probes:
        sc = probe_scores(model, tok, probes, dev)
        known = sum(s > epsilon for s in sc) / len(sc)
        med = sorted(sc)[len(sc) // 2]
        msg.append(f"probes: known_rate={known:.2%} (need >=70%) median={med:.2f}")
    log("[eval] " + "  ".join(msg))


# ------------------------------------------------------------------ model helpers
def enable_checkpointing(model):
    """nanochrono has supports_gradient_checkpointing=False; wrap decoder layers manually."""
    for layer in model.model.layers:
        fwd = layer.forward
        layer.forward = (lambda f: lambda *a, **k: checkpoint(f, *a, use_reentrant=False, **k))(fwd)


def lr_at(step, total, base, warmup, floor=0.1):
    if step < warmup:
        return base * (step + 1) / warmup
    p = (step - warmup) / max(1, total - warmup)
    return base * (floor + (1 - floor) * 0.5 * (1 + math.cos(math.pi * p)))


def save(model, tok, out, step, rep_pen):
    d = os.path.join(out, f"step-{step:06d}")
    compiled = getattr(model.model, "_orig_mod", None)
    if compiled is not None:
        model.model = compiled                      # save clean keys, not "_orig_mod.*"
    model.to(torch.bfloat16).save_pretrained(d, safe_serialization=True)
    model.to(torch.float32)
    if compiled is not None:
        model.model = torch.compile(model.model)
    tok.save_pretrained(d)
    GenerationConfig(do_sample=False, repetition_penalty=rep_pen, max_new_tokens=100,
                     eos_token_id=tok.eos_token_id, pad_token_id=tok.pad_token_id, bos_token_id=tok.bos_token_id).save_pretrained(d)
    latest = os.path.join(out, "latest")
    if os.path.islink(latest): os.remove(latest)
    os.symlink(os.path.basename(d), latest)
    print(f"[save] {d}", flush=True)
    return d


# ------------------------------------------------------------------ main
def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--mode", choices=["cpt", "sft"], default="cpt")
    ap.add_argument("--base", required=True, help="dir of the starting checkpoint (HF layout)")
    ap.add_argument("--out", required=True)
    ap.add_argument("--corpus", help="output dir of build_corpus.py (cpt)")
    ap.add_argument("--mix", default="", help="cpt: name=weight,name=weight,...")
    ap.add_argument("--replay-sources", default="fineweb_edu_replay,fnspid_news_1999_2023", help="names counted as replay for val")
    ap.add_argument("--sft-jsonl", help="sft: jsonl (or comma list) with prompt/response rows")
    ap.add_argument("--tokens", type=float, default=3e9, help="total training tokens")
    ap.add_argument("--seq", type=int, default=2048)
    ap.add_argument("--micro-bs", type=int, default=8)
    ap.add_argument("--global-tokens", type=int, default=1_048_576, help="tokens per optimizer step")
    ap.add_argument("--lr", type=float, default=1e-4)
    ap.add_argument("--warmup", type=int, default=100)
    ap.add_argument("--wd", type=float, default=0.1)
    ap.add_argument("--save-every", type=int, default=500)
    ap.add_argument("--eval-every", type=int, default=100)
    ap.add_argument("--val-docs", type=int, default=64, help="held-out docs per source for val loss (cpt)")
    ap.add_argument("--probes", default="", help="csv year,sentence,phrase scored like leak.py")
    ap.add_argument("--epsilon", type=float, default=-11.51)
    ap.add_argument("--doc-mask", action="store_true", help="block-diagonal attention across packed documents")
    ap.add_argument("--ckpt", action="store_true", help="activation checkpointing (80GB GPUs)")
    ap.add_argument("--compile", action="store_true", help="torch.compile the backbone")
    ap.add_argument("--rep-penalty", type=float, default=1.1, help="written to generation_config.json")
    ap.add_argument("--seed", type=int, default=1337)
    ap.add_argument("--device", default="cuda")
    ap.add_argument("--resume", action="store_true", help="continue from <out>/resume.pt (fp32 weights + optimizer + step); spot-instance safe")
    a = ap.parse_args()

    # ---- multi-GPU (torchrun): data parallel with a manual all-reduce after accumulation. Each rank streams its own shuffle
    # of the mix (seed+rank) and --global-tokens is split across ranks, so the optimizer sees the same tokens/step as 1 GPU.
    rank = int(os.environ.get("RANK", 0)); world = int(os.environ.get("WORLD_SIZE", 1)); local = int(os.environ.get("LOCAL_RANK", 0))
    if world > 1:
        import torch.distributed as dist
        dist.init_process_group("nccl" if a.device.startswith("cuda") else "gloo")
        if a.device.startswith("cuda"):
            a.device = f"cuda:{local}"; torch.cuda.set_device(local)
    is_main = rank == 0
    import builtins
    def print(*args, **kw):            # only rank 0 talks (shadows the builtin inside main)
        if is_main: builtins.print(*args, **kw)

    torch.manual_seed(a.seed + rank)
    dev = torch.device(a.device)
    if dev.type == "cuda":
        torch.backends.cuda.matmul.allow_tf32 = True
    tok = AutoTokenizer.from_pretrained(a.base)
    model = AutoModelForCausalLM.from_pretrained(a.base, dtype=torch.float32).to(dev)
    model.config.use_cache = False
    model.train()
    if a.ckpt:
        enable_checkpointing(model)
    if a.doc_mask:
        install_doc_mask(model)
    if a.compile:
        model.model = torch.compile(model.model)
    n_params = sum(p.numel() for p in model.parameters())
    print(f"[model] {n_params/1e9:.3f}B params  (limit 2.2B)", flush=True)
    assert n_params <= 2_200_000_000

    decay = [p for n, p in model.named_parameters() if p.ndim >= 2]
    no_decay = [p for n, p in model.named_parameters() if p.ndim < 2]
    opt = torch.optim.AdamW([{"params": decay, "weight_decay": a.wd}, {"params": no_decay, "weight_decay": 0.0}],
                            lr=a.lr, betas=(0.9, 0.95), eps=1e-8, fused=(dev.type == "cuda"))

    accum = max(1, a.global_tokens // (a.micro_bs * a.seq * world))
    step_tokens = accum * a.micro_bs * a.seq * world
    total_steps = int(a.tokens // step_tokens)
    print(f"[plan] steps={total_steps} accum={accum} x {world} GPU(s) micro_bs={a.micro_bs} seq={a.seq} -> {step_tokens:,} tok/step "
          f"lr={a.lr} warmup={a.warmup} doc_mask={a.doc_mask}", flush=True)

    start_step = 0
    resume_path = os.path.join(a.out, "resume.pt")
    if a.resume and os.path.exists(resume_path):
        st = torch.load(resume_path, map_location=dev, weights_only=False)
        model.load_state_dict(st["model"]); opt.load_state_dict(st["opt"]); start_step = st["step"]
        print(f"[resume] {resume_path} -> step {start_step}/{total_steps} (data streams re-seeded with seed+step)", flush=True)
    data_seed = a.seed + rank + 1000 * start_step

    val_sets = {}
    if a.mode == "cpt":
        mix = {kv.split("=")[0]: float(kv.split("=")[1]) for kv in a.mix.split(",") if kv}
        if not mix:
            raise SystemExit("--mix required for cpt, e.g. fineweb_2024=0.3,fineweb_edu_2024=0.25,...")
        replay = set(a.replay_sources.split(","))
        new_txt, rep_txt = [], []
        for n in mix:
            (rep_txt if n in replay else new_txt).extend(heldout_docs(os.path.join(a.corpus, n), a.val_docs))
        val_sets = {"new2024": new_txt[:256], "replay": rep_txt[:128]}
        data = Packer(tok, a.seq, a.micro_bs).stream(mixed_docs(a.corpus, mix, data_seed, a.val_docs))
    else:
        data = sft_batches(a.sft_jsonl, tok, a.seq, a.micro_bs, data_seed, shard=(rank, world))
    probes = load_probes(a.probes) if a.probes else []

    os.makedirs(a.out, exist_ok=True)
    log = lambda m: print(f"[{time.strftime('%H:%M:%S')}] {m}", flush=True)

    def save_resume(step):
        if not is_main: return
        tmp = resume_path + ".tmp"
        torch.save({"model": model.state_dict(), "opt": opt.state_dict(), "step": step}, tmp)
        os.replace(tmp, resume_path)

    if is_main and start_step == 0:
        run_eval(model, tok, dev, val_sets, probes, a.seq, a.epsilon, log)
    t0 = time.time(); seen = 0; pad_frac = 0.0
    for step in range(start_step, total_steps):
        for g in opt.param_groups: g["lr"] = lr_at(step, total_steps, a.lr, a.warmup)
        loss_sum = 0.0; tgt_sum = 0
        for _ in range(accum):
            x, y, doc = next(data)
            x, y = x.to(dev, non_blocking=True), y.to(dev, non_blocking=True)
            _DOC_MASK["m"] = doc_mask_from_ids(doc[:, :-1].to(dev)) if (a.doc_mask and doc is not None) else None
            with torch.autocast(dev.type, dtype=torch.bfloat16, enabled=(dev.type == "cuda")):
                s, k = lm_loss(model, x, y)
            (s / (accum * a.micro_bs * a.seq)).backward()   # normalise by nominal tokens (constant) -> stable step size
            loss_sum += s.item(); tgt_sum += k.item()
            seen += x.numel(); pad_frac = 0.98 * pad_frac + 0.02 * (1 - k.item() / (x.numel() - x.size(0)))
        _DOC_MASK["m"] = None
        if world > 1:                                    # average gradients across ranks (every rank then holds identical grads)
            for p in model.parameters():
                if p.grad is not None:
                    dist.all_reduce(p.grad, op=dist.ReduceOp.AVG)
        gn = torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        opt.step(); opt.zero_grad(set_to_none=True)
        if step % 10 == 0 or step == total_steps - 1:
            el = time.time() - t0; done = step + 1 - start_step
            log(f"step {step}/{total_steps} loss {loss_sum/max(1,tgt_sum):.4f} gnorm {gn:.2f} lr {opt.param_groups[0]['lr']:.2e} "
                f"pad {pad_frac:.1%} tok/s {seen*world/el:,.0f} eta {el/done*(total_steps-step-1)/3600:.1f}h")
        if (step + 1) % a.eval_every == 0 and is_main:
            run_eval(model, tok, dev, val_sets, probes, a.seq, a.epsilon, log)
        if (step + 1) % a.save_every == 0 and step + 1 < total_steps:
            if is_main:
                save(model, tok, a.out, step + 1, a.rep_penalty)
            save_resume(step + 1)
    if is_main:
        run_eval(model, tok, dev, val_sets, probes, a.seq, a.epsilon, log)
        final = save(model, tok, a.out, total_steps, a.rep_penalty)
        print(f"[done] final checkpoint: {final}")
    if world > 1:
        dist.barrier(); dist.destroy_process_group()


if __name__ == "__main__":
    main()

	#!/usr/bin/env python3
	"""SN38 full-parameter continued pretraining (CPT) / SFT for sn38-nanochrono.

	Run from the sn38 repo env so the architecture is registered:
	cd /opt/sn38 && PYTHONPATH=/opt/sn38 uv run python /opt/work/train.py --mode cpt ...

	What is implemented (see README section "What the research says"):
	* CPT: sources sampled by --mix; documents packed BEST-FIT (whole docs, no needless truncation),
	each sequence gets a block-diagonal DOCUMENT MASK so tokens never attend across documents (--doc-mask).
	* Loss computed in chunks (never materialises the full fp32 [B,T,32768] logits), -100 labels respected.
	* LR: linear warm-up then cosine re-decay to 10% (re-warm + re-decay recipe for CPT).
	* Held-out validation per source group (new-2024 vs replay) and a probe set (year,sentence,phrase csv)
	scored exactly like the validator's leak.py (BOS+prompt, phrase with leading space, sum log-prob, epsilon).
	* SFT: loss only on response + EOS (position-based mask, so EOS==PAD is NOT a problem), asserts the EOS is a target.
	* bf16 autocast with fp32 master weights, fused AdamW, grad clip, optional activation checkpointing / torch.compile.
	Saves HF layout (config.json, model.safetensors [bf16], tokenizer, generation_config.json) + a symlink `latest`.
	"""
	import argparse, csv, glob, gzip, json, math, os, random, time
	# Multi-GPU: torchrun --standalone --nproc_per_node 8 train.py ... (data parallel, grads averaged after accumulation;
	# --global-tokens is the TOTAL per optimizer step across GPUs). --resume restarts from <out>/resume.pt (spot instances).
	import torch, torch.nn.functional as F
	from torch.utils.checkpoint import checkpoint
	from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
	import sn38.architectures # noqa: registers sn38-nanochrono / nanoexpand / miniformer
	sn38.architectures.register_all()

	IGN = -100


	# ------------------------------------------------------------------ data: sources
	def docs_from_source(src_dir, seed, skip=0):
	"""Infinite shuffled stream of texts; the first `skip` docs of the first shard are held out (validation)."""
	files = sorted(glob.glob(os.path.join(src_dir, "part-*.jsonl.gz")))
	if not files:
	raise SystemExit(f"no shards in {src_dir}")
	rng = random.Random(seed)
	epoch = 0
	while True:
	order = files[:]; rng.shuffle(order)
	for f in order:
	with gzip.open(f, "rt", encoding="utf-8") as fh:
	for i, line in enumerate(fh):
	if f == files[0] and i < skip:
	continue
	yield json.loads(line)["text"]
	epoch += 1
	print(f"[data] {os.path.basename(src_dir)} epoch {epoch} done", flush=True)


	def heldout_docs(src_dir, n):
	files = sorted(glob.glob(os.path.join(src_dir, "part-*.jsonl.gz")))
	out = []
	with gzip.open(files[0], "rt", encoding="utf-8") as fh:
	for i, line in enumerate(fh):
	if i >= n: break
	out.append(json.loads(line)["text"])
	return out


	# ------------------------------------------------------------------ data: best-fit packing + doc mask
	class Packer:
	"""Best-fit-decreasing packing of whole documents into seq+1 windows.
	Docs longer than seq+1 are split into seq+1 chunks (unavoidable); everything else is never truncated.
	Yields (ids[B,T+1], labels[B,T+1], docid[B,T+1]) ; pad positions have label -100."""

	def __init__(self, tok, seq, micro_bs, buffer_docs=512):
	self.tok, self.L, self.B, self.buf_n = tok, seq + 1, micro_bs, buffer_docs
	self.bos, self.eos, self.pad = tok.bos_token_id, tok.eos_token_id, tok.pad_token_id

	def _chunks(self, texts):
	for ids in self.tok(texts, add_special_tokens=False)["input_ids"]:
	ids = [self.bos] + ids + [self.eos]
	for i in range(0, len(ids), self.L):
	yield ids[i:i + self.L]

	def pack(self, chunks):
	chunks.sort(key=len, reverse=True)
	bins = [] # list of [free, [chunk,...]]
	for c in chunks:
	best = None
	for b in bins:
	if b[0] >= len(c) and (best is None or b[0] < best[0]):
	best = b
	if best is None:
	bins.append([self.L - len(c), [c]])
	else:
	best[0] -= len(c); best[1].append(c)
	return bins

	def stream(self, doc_iter):
	carry = []
	while True:
	texts = [next(doc_iter) for _ in range(self.buf_n)]
	bins = self.pack(carry + list(self._chunks(texts)))
	bins.sort(key=lambda b: b[0]) # fullest bins first
	keep = len(bins) - len(bins) % self.B
	carry = [c for b in bins[keep:] for c in b[1]] # sparse leftovers are re-packed next round
	for i in range(0, keep, self.B):
	yield self._tensorise(bins[i:i + self.B])

	def _tensorise(self, bins):
	ids = torch.full((self.B, self.L), self.pad, dtype=torch.long)
	lab = torch.full((self.B, self.L), IGN, dtype=torch.long)
	doc = torch.zeros((self.B, self.L), dtype=torch.long)
	for r, (_, chunks) in enumerate(bins):
	p = 0
	for d, c in enumerate(chunks, start=1):
	ids[r, p:p + len(c)] = torch.tensor(c); lab[r, p:p + len(c)] = torch.tensor(c)
	doc[r, p:p + len(c)] = d; p += len(c)
	return ids, lab, doc


	def mixed_docs(corpus, mix, seed, skip):
	names = list(mix); weights = [mix[n] for n in names]
	gens = {n: docs_from_source(os.path.join(corpus, n), seed + i, skip) for i, n in enumerate(names)}
	rng = random.Random(seed)
	while True:
	yield next(gens[rng.choices(names, weights)[0]])


	_DOC_MASK = {"m": None}


	def install_doc_mask(model):
	"""Patch nanochrono attention so a block-diagonal document mask is AND-ed with causal/sliding masks."""
	import sn38.architectures.nanochrono.modeling_nanochrono as mn
	orig = mn.NanochronoAttention._attend

	def _attend(self, q, k, v, window, mask, past_len):
	dm = _DOC_MASK["m"]
	if dm is None or q.size(2) != k.size(2):
	return orig(self, q, k, v, window, mask, past_len)
	T = q.size(2)
	row = torch.arange(T, device=q.device).unsqueeze(1); col = torch.arange(T, device=q.device).unsqueeze(0)
	m = col <= row
	if 0 <= window < T:
	m = m & ((row - col) <= window)
	m = m.unsqueeze(0).unsqueeze(0) & dm
	return F.scaled_dot_product_attention(q, k, v, attn_mask=m)
	mn.NanochronoAttention._attend = _attend


	def doc_mask_from_ids(doc): # doc [B,T] ints, 0 = padding -> [B,1,T,T] bool
	m = (doc.unsqueeze(2) == doc.unsqueeze(1)) & (doc.unsqueeze(2) > 0)
	eye = torch.eye(doc.size(1), dtype=torch.bool, device=doc.device).unsqueeze(0)
	return (m \| eye).unsqueeze(1) # padding rows attend to themselves so softmax is defined


	# ------------------------------------------------------------------ data: SFT
	def sft_batches(path, tok, seq, micro_bs, seed, shard=(0, 1)):
	"""path may be a comma list of jsonl files (prompt/response rows). The response is tokenized with ONE leading space
	so the Q->A boundary matches natural text ("...? The answer") -- at eval the model continues straight after the raw
	question with no separator, so whatever boundary we train is what it will emit."""
	rows = []
	for f in path.split(","):
	rows += [json.loads(l) for l in open(f, encoding="utf-8")]
	rows = rows[shard[0]::shard[1]] # each DDP rank owns a disjoint slice
	if shard[0] == 0:
	print(f"[sft] {len(rows) * shard[1]} rows from {path} ({shard[1]} shard(s))", flush=True)
	rng = random.Random(seed); bos, eos, pad = tok.bos_token_id, tok.eos_token_id, tok.pad_token_id
	while True:
	rng.shuffle(rows)
	for i in range(0, len(rows) - micro_bs + 1, micro_bs):
	xs, ys = [], []
	for r in rows[i:i + micro_bs]:
	p = [bos] + tok(r["prompt"], add_special_tokens=False)["input_ids"]
	a = tok(" " + r["response"].strip(), add_special_tokens=False)["input_ids"] + [eos]
	if len(p) + len(a) > seq + 1: # never cut the EOS off: trim the prompt from the left
	p = p[-(seq + 1 - len(a)):] if len(a) < seq + 1 else p[:1]
	a = a[: seq + 1 - len(p)]
	ids = p + a; lab = [IGN] * len(p) + a
	assert lab[-1] == eos, "EOS must be a training target"
	padn = seq + 1 - len(ids)
	xs.append(ids + [pad] * padn); ys.append(lab + [IGN] * padn)
	x, y = torch.tensor(xs), torch.tensor(ys)
	yield x, y, None


	# ------------------------------------------------------------------ loss (chunked, softcapped, -100 aware)
	def lm_loss(model, ids, labels, chunk=512):
	"""ids/labels [B,T+1]. Returns (sum_nll, n_targets). Only the hidden states are stored; logits are
	recomputed per chunk in backward (checkpoint), so the fp32 [B,T,V] tensor never exists at once."""
	hidden = model.model(input_ids=ids[:, :-1], use_cache=False).last_hidden_state
	tgt = labels[:, 1:]
	cap = model.config.final_logit_softcapping
	W = model.lm_head.weight

	def piece(h, t):
	lg = F.linear(h, W).float()
	if cap:
	lg = cap * torch.tanh(lg / cap)
	return F.cross_entropy(lg.view(-1, lg.size(-1)), t.reshape(-1), ignore_index=IGN, reduction="sum")

	total = hidden.new_zeros((), dtype=torch.float32)
	for s in range(0, hidden.size(1), chunk):
	total = total + checkpoint(piece, hidden[:, s:s + chunk], tgt[:, s:s + chunk], use_reentrant=False)
	return total, (tgt != IGN).sum()


	# ------------------------------------------------------------------ evaluation
	@torch.no_grad()
	def eval_loss(model, tok, texts, seq, dev, bs=8):
	model.eval(); tot = 0.0; n = 0
	for i in range(0, len(texts), bs):
	enc = tok(texts[i:i + bs], add_special_tokens=False, truncation=True, max_length=seq)["input_ids"]
	L = max(len(e) for e in enc) + 1
	ids = torch.full((len(enc), L), tok.pad_token_id); lab = torch.full((len(enc), L), IGN)
	for r, e in enumerate(enc):
	e = [tok.bos_token_id] + e
	ids[r, :len(e)] = torch.tensor(e); lab[r, :len(e)] = torch.tensor(e)
	with torch.autocast(dev.type, dtype=torch.bfloat16, enabled=dev.type == "cuda"):
	s, k = lm_loss(model, ids.to(dev), lab.to(dev))
	tot += s.item(); n += k.item()
	model.train()
	return tot / max(1, n)


	def load_probes(path):
	rows = list(csv.DictReader(open(path, encoding="utf-8")))
	out = []
	for r in rows:
	s, p = r["sentence"], r["phrase"]
	i = s.lower().rfind(p.lower())
	out.append({"year": int(r["year"]), "prompt": (s[:i].rstrip() if i >= 0 else s), "phrase": p})
	return out


	@torch.no_grad()
	def probe_scores(model, tok, probes, dev):
	"""Sum of log-probs of the phrase tokens after BOS+prompt — identical arithmetic to sn38.template.leak."""
	model.eval(); scores = []
	for it in probes:
	p = [tok.bos_token_id] + tok(it["prompt"], add_special_tokens=False)["input_ids"]
	ph = tok(" " + it["phrase"], add_special_tokens=False)["input_ids"]
	ids = torch.tensor([p + ph], device=dev)
	with torch.autocast(dev.type, dtype=torch.bfloat16, enabled=dev.type == "cuda"):
	lg = model(input_ids=ids, use_cache=False).logits.float()
	lp = F.log_softmax(lg[0, len(p) - 1:len(p) - 1 + len(ph)], -1)
	scores.append(sum(lp[j, t].item() for j, t in enumerate(ph)))
	model.train()
	return scores


	def run_eval(model, tok, dev, val_sets, probes, seq, epsilon, log):
	msg = []
	for name, texts in val_sets.items():
	msg.append(f"val/{name}={eval_loss(model, tok, texts, seq, dev):.4f}")
	if probes:
	sc = probe_scores(model, tok, probes, dev)
	known = sum(s > epsilon for s in sc) / len(sc)
	med = sorted(sc)[len(sc) // 2]
	msg.append(f"probes: known_rate={known:.2%} (need >=70%) median={med:.2f}")
	log("[eval] " + " ".join(msg))


	# ------------------------------------------------------------------ model helpers
	def enable_checkpointing(model):
	"""nanochrono has supports_gradient_checkpointing=False; wrap decoder layers manually."""
	for layer in model.model.layers:
	fwd = layer.forward
	layer.forward = (lambda f: lambda a, k: checkpoint(f, a, use_reentrant=False, **k))(fwd)


	def lr_at(step, total, base, warmup, floor=0.1):
	if step < warmup:
	return base * (step + 1) / warmup
	p = (step - warmup) / max(1, total - warmup)
	return base * (floor + (1 - floor) * 0.5 * (1 + math.cos(math.pi * p)))


	def save(model, tok, out, step, rep_pen):
	d = os.path.join(out, f"step-{step:06d}")
	compiled = getattr(model.model, "_orig_mod", None)
	if compiled is not None:
	model.model = compiled # save clean keys, not "_orig_mod.*"
	model.to(torch.bfloat16).save_pretrained(d, safe_serialization=True)
	model.to(torch.float32)
	if compiled is not None:
	model.model = torch.compile(model.model)
	tok.save_pretrained(d)
	GenerationConfig(do_sample=False, repetition_penalty=rep_pen, max_new_tokens=100,
	eos_token_id=tok.eos_token_id, pad_token_id=tok.pad_token_id, bos_token_id=tok.bos_token_id).save_pretrained(d)
	latest = os.path.join(out, "latest")
	if os.path.islink(latest): os.remove(latest)
	os.symlink(os.path.basename(d), latest)
	print(f"[save] {d}", flush=True)
	return d


	# ------------------------------------------------------------------ main
	def main():
	ap = argparse.ArgumentParser()
	ap.add_argument("--mode", choices=["cpt", "sft"], default="cpt")
	ap.add_argument("--base", required=True, help="dir of the starting checkpoint (HF layout)")
	ap.add_argument("--out", required=True)
	ap.add_argument("--corpus", help="output dir of build_corpus.py (cpt)")
	ap.add_argument("--mix", default="", help="cpt: name=weight,name=weight,...")
	ap.add_argument("--replay-sources", default="fineweb_edu_replay,fnspid_news_1999_2023", help="names counted as replay for val")
	ap.add_argument("--sft-jsonl", help="sft: jsonl (or comma list) with prompt/response rows")
	ap.add_argument("--tokens", type=float, default=3e9, help="total training tokens")
	ap.add_argument("--seq", type=int, default=2048)
	ap.add_argument("--micro-bs", type=int, default=8)
	ap.add_argument("--global-tokens", type=int, default=1_048_576, help="tokens per optimizer step")
	ap.add_argument("--lr", type=float, default=1e-4)
	ap.add_argument("--warmup", type=int, default=100)
	ap.add_argument("--wd", type=float, default=0.1)
	ap.add_argument("--save-every", type=int, default=500)
	ap.add_argument("--eval-every", type=int, default=100)
	ap.add_argument("--val-docs", type=int, default=64, help="held-out docs per source for val loss (cpt)")
	ap.add_argument("--probes", default="", help="csv year,sentence,phrase scored like leak.py")
	ap.add_argument("--epsilon", type=float, default=-11.51)
	ap.add_argument("--doc-mask", action="store_true", help="block-diagonal attention across packed documents")
	ap.add_argument("--ckpt", action="store_true", help="activation checkpointing (80GB GPUs)")
	ap.add_argument("--compile", action="store_true", help="torch.compile the backbone")
	ap.add_argument("--rep-penalty", type=float, default=1.1, help="written to generation_config.json")
	ap.add_argument("--seed", type=int, default=1337)
	ap.add_argument("--device", default="cuda")
	ap.add_argument("--resume", action="store_true", help="continue from <out>/resume.pt (fp32 weights + optimizer + step); spot-instance safe")
	a = ap.parse_args()

	# ---- multi-GPU (torchrun): data parallel with a manual all-reduce after accumulation. Each rank streams its own shuffle
	# of the mix (seed+rank) and --global-tokens is split across ranks, so the optimizer sees the same tokens/step as 1 GPU.
	rank = int(os.environ.get("RANK", 0)); world = int(os.environ.get("WORLD_SIZE", 1)); local = int(os.environ.get("LOCAL_RANK", 0))
	if world > 1:
	import torch.distributed as dist
	dist.init_process_group("nccl" if a.device.startswith("cuda") else "gloo")
	if a.device.startswith("cuda"):
	a.device = f"cuda:{local}"; torch.cuda.set_device(local)
	is_main = rank == 0
	import builtins
	def print(args, *kw): # only rank 0 talks (shadows the builtin inside main)
	if is_main: builtins.print(args, *kw)

	torch.manual_seed(a.seed + rank)
	dev = torch.device(a.device)
	if dev.type == "cuda":
	torch.backends.cuda.matmul.allow_tf32 = True
	tok = AutoTokenizer.from_pretrained(a.base)
	model = AutoModelForCausalLM.from_pretrained(a.base, dtype=torch.float32).to(dev)
	model.config.use_cache = False
	model.train()
	if a.ckpt:
	enable_checkpointing(model)
	if a.doc_mask:
	install_doc_mask(model)
	if a.compile:
	model.model = torch.compile(model.model)
	n_params = sum(p.numel() for p in model.parameters())
	print(f"[model] {n_params/1e9:.3f}B params (limit 2.2B)", flush=True)
	assert n_params <= 2_200_000_000

	decay = [p for n, p in model.named_parameters() if p.ndim >= 2]
	no_decay = [p for n, p in model.named_parameters() if p.ndim < 2]
	opt = torch.optim.AdamW([{"params": decay, "weight_decay": a.wd}, {"params": no_decay, "weight_decay": 0.0}],
	lr=a.lr, betas=(0.9, 0.95), eps=1e-8, fused=(dev.type == "cuda"))

	accum = max(1, a.global_tokens // (a.micro_bs * a.seq * world))
	step_tokens = accum * a.micro_bs * a.seq * world
	total_steps = int(a.tokens // step_tokens)
	print(f"[plan] steps={total_steps} accum={accum} x {world} GPU(s) micro_bs={a.micro_bs} seq={a.seq} -> {step_tokens:,} tok/step "
	f"lr={a.lr} warmup={a.warmup} doc_mask={a.doc_mask}", flush=True)

	start_step = 0
	resume_path = os.path.join(a.out, "resume.pt")
	if a.resume and os.path.exists(resume_path):
	st = torch.load(resume_path, map_location=dev, weights_only=False)
	model.load_state_dict(st["model"]); opt.load_state_dict(st["opt"]); start_step = st["step"]
	print(f"[resume] {resume_path} -> step {start_step}/{total_steps} (data streams re-seeded with seed+step)", flush=True)
	data_seed = a.seed + rank + 1000 * start_step

	val_sets = {}
	if a.mode == "cpt":
	mix = {kv.split("=")[0]: float(kv.split("=")[1]) for kv in a.mix.split(",") if kv}
	if not mix:
	raise SystemExit("--mix required for cpt, e.g. fineweb_2024=0.3,fineweb_edu_2024=0.25,...")
	replay = set(a.replay_sources.split(","))
	new_txt, rep_txt = [], []
	for n in mix:
	(rep_txt if n in replay else new_txt).extend(heldout_docs(os.path.join(a.corpus, n), a.val_docs))
	val_sets = {"new2024": new_txt[:256], "replay": rep_txt[:128]}
	data = Packer(tok, a.seq, a.micro_bs).stream(mixed_docs(a.corpus, mix, data_seed, a.val_docs))
	else:
	data = sft_batches(a.sft_jsonl, tok, a.seq, a.micro_bs, data_seed, shard=(rank, world))
	probes = load_probes(a.probes) if a.probes else []

	os.makedirs(a.out, exist_ok=True)
	log = lambda m: print(f"[{time.strftime('%H:%M:%S')}] {m}", flush=True)

	def save_resume(step):
	if not is_main: return
	tmp = resume_path + ".tmp"
	torch.save({"model": model.state_dict(), "opt": opt.state_dict(), "step": step}, tmp)
	os.replace(tmp, resume_path)

	if is_main and start_step == 0:
	run_eval(model, tok, dev, val_sets, probes, a.seq, a.epsilon, log)
	t0 = time.time(); seen = 0; pad_frac = 0.0
	for step in range(start_step, total_steps):
	for g in opt.param_groups: g["lr"] = lr_at(step, total_steps, a.lr, a.warmup)
	loss_sum = 0.0; tgt_sum = 0
	for _ in range(accum):
	x, y, doc = next(data)
	x, y = x.to(dev, non_blocking=True), y.to(dev, non_blocking=True)
	_DOC_MASK["m"] = doc_mask_from_ids(doc[:, :-1].to(dev)) if (a.doc_mask and doc is not None) else None
	with torch.autocast(dev.type, dtype=torch.bfloat16, enabled=(dev.type == "cuda")):
	s, k = lm_loss(model, x, y)
	(s / (accum * a.micro_bs * a.seq)).backward() # normalise by nominal tokens (constant) -> stable step size
	loss_sum += s.item(); tgt_sum += k.item()
	seen += x.numel(); pad_frac = 0.98 * pad_frac + 0.02 * (1 - k.item() / (x.numel() - x.size(0)))
	_DOC_MASK["m"] = None
	if world > 1: # average gradients across ranks (every rank then holds identical grads)
	for p in model.parameters():
	if p.grad is not None:
	dist.all_reduce(p.grad, op=dist.ReduceOp.AVG)
	gn = torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
	opt.step(); opt.zero_grad(set_to_none=True)
	if step % 10 == 0 or step == total_steps - 1:
	el = time.time() - t0; done = step + 1 - start_step
	log(f"step {step}/{total_steps} loss {loss_sum/max(1,tgt_sum):.4f} gnorm {gn:.2f} lr {opt.param_groups[0]['lr']:.2e} "
	f"pad {pad_frac:.1%} tok/s {seenworld/el:,.0f} eta {el/done(total_steps-step-1)/3600:.1f}h")
	if (step + 1) % a.eval_every == 0 and is_main:
	run_eval(model, tok, dev, val_sets, probes, a.seq, a.epsilon, log)
	if (step + 1) % a.save_every == 0 and step + 1 < total_steps:
	if is_main:
	save(model, tok, a.out, step + 1, a.rep_penalty)
	save_resume(step + 1)
	if is_main:
	run_eval(model, tok, dev, val_sets, probes, a.seq, a.epsilon, log)
	final = save(model, tok, a.out, total_steps, a.rep_penalty)
	print(f"[done] final checkpoint: {final}")
	if world > 1:
	dist.barrier(); dist.destroy_process_group()


	if __name__ == "__main__":
	main()