File size: 13,874 Bytes
251ba47
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
"""Pretraining loop: DDP + bf16, WSD schedule, deterministic resume.

One run = one (mixture, tokenizer) cell of the design. The mixture is driven
entirely by the run config's `langs`/`probs`; the tokenizer by `tok_name`. The
loader is globally deterministic and world-size-independent, so a run resumed on
a different node count sees the exact same token stream.

Cooldown branch: set `branch.from` to a `stable` checkpoint; the schedule then
has warmup=stable=0 and only decays, giving a cheap final model at a larger
token budget without retraining the trunk.
"""
import json
import os
import time
from pathlib import Path

import numpy as np
import torch

from .model import ModelConfig, Transformer
from .data.loader import MixedStream
from .schedule import lr_at, ckpt_interval, stable_end_tokens, total_tokens
from .paths import run_dir, ensure


def _ddp():
    if "RANK" in os.environ and int(os.environ.get("WORLD_SIZE", "1")) > 1:
        import torch.distributed as dist
        backend = "nccl" if torch.cuda.is_available() else "gloo"
        dist.init_process_group(backend=backend)
        rank = dist.get_rank()
        world = dist.get_world_size()
        local = int(os.environ.get("LOCAL_RANK", "0"))
        if torch.cuda.is_available():
            torch.cuda.set_device(local)
        return dist.is_initialized(), rank, world, local
    return False, 0, 1, 0


def _log(rank, path, rec):
    if rank == 0:
        with open(path, "a") as f:
            f.write(json.dumps(rec) + "\n")


class Trainer:
    def __init__(self, cfg: dict):
        self.cfg = cfg
        self.dist, self.rank, self.world, self.local = _ddp()
        self.device = torch.device(f"cuda:{self.local}" if torch.cuda.is_available() else "cpu")
        self.is_cuda = self.device.type == "cuda"
        torch.manual_seed(cfg.get("seed", 0))
        np.random.seed(cfg.get("seed", 0))

        mc = cfg["model"]
        self.mcfg = ModelConfig(**mc)
        self.seq_len = self.mcfg.max_seq_len
        raw_model = Transformer(self.mcfg).to(self.device)
        if self.rank == 0:
            print(f"[train] params: {raw_model.num_params(False)/1e6:.1f}M non-embedding, "
                  f"{raw_model.num_params(True)/1e6:.1f}M total")
        # Keep the canonical module unwrapped for portable state_dict keys.
        # Compilation is only the forward/backward execution path.
        model = raw_model
        if cfg.get("compile", False) and self.is_cuda:
            model = torch.compile(model)
        self.raw_model = raw_model
        if self.dist:
            from torch.nn.parallel import DistributedDataParallel as DDP
            model = DDP(model, device_ids=[self.local] if self.is_cuda else None)
        self.model = model

        opt = cfg.get("optim", {})
        self.optim = torch.optim.AdamW(
            self.raw_model.parameters(),
            lr=cfg["schedule"]["peak_lr"],
            betas=tuple(opt.get("betas", (0.9, 0.95))),
            weight_decay=opt.get("weight_decay", 0.1),
            eps=opt.get("eps", 1e-8),
        )
        self.grad_clip = opt.get("grad_clip", 1.0)

        # global batch bookkeeping
        self.micro_bsz = cfg["train"]["micro_batch_size"]
        gbt = cfg["train"]["global_batch_tokens"]
        per_step_windows = max(1, round(gbt / self.seq_len))
        # round up to a multiple of micro_bsz*world so each rank gets equal work
        unit = self.micro_bsz * self.world
        self.global_windows = max(unit, (per_step_windows // unit) * unit)
        self.grad_accum = self.global_windows // unit
        self.tokens_per_step = self.global_windows * self.seq_len
        if self.rank == 0:
            print(f"[train] global batch: {self.global_windows} windows "
                  f"({self.tokens_per_step/1e6:.3f}M tokens), grad_accum={self.grad_accum}")

        # schedule (branch collapses warmup/stable)
        self.sched = dict(cfg["schedule"])
        self.branch = cfg.get("branch")
        if self.branch:
            self.sched["warmup_tokens"] = 0
            self.sched["stable_tokens"] = 0
        self.target_tokens = total_tokens(self.sched)
        self.ckpt_table = cfg["train"].get("ckpt_schedule",
                                           [[2e9, 250e6], [5e9, 500e6],
                                            [15e9, 1e9], [1e15, 2e9]])
        # token budgets at which a stable trunk saves a named branch point
        self.stable_marks = sorted(cfg["train"].get("stable_marks", []))
        self.marks_done = set()

        # data mixer
        self.mixer = MixedStream(cfg["langs"], cfg["tok_name"], self.seq_len,
                                 seed=cfg.get("data_seed", 1234),
                                 probs=cfg.get("probs"))

        self.rdir = ensure(run_dir(cfg["name"]))
        self.log_path = self.rdir / "train.jsonl"
        self.tokens = 0
        self.step = 0
        self.last_ckpt_tokens = 0
        self.saved_stable = False

        self.wandb = None
        if self.rank == 0:
            try:
                import wandb
                self.wandb = wandb.init(
                    project="XScript-Pretraining", name=cfg["name"],
                    id=cfg.get("wandb_id", cfg["name"]),
                    resume="allow", config=cfg,
                )
                self.wandb.summary["params_total_M"] = self.raw_model.num_params(True) / 1e6
                self.wandb.summary["params_non_embed_M"] = self.raw_model.num_params(False) / 1e6
            except Exception as exc:
                print(f"[train] wandb disabled ({exc})")

    # ---- checkpoint io ----
    def _ckpt_path(self, tag):
        return ensure(self.rdir / "checkpoints") / f"{tag}.pt"

    def save(self, tag, resumable=True):
        if self.rank != 0:
            return
        payload = {
            "model": self.raw_model.state_dict(),
            "step": self.step, "tokens": self.tokens,
            "cfg": self.cfg,
        }
        if resumable:
            payload.update({
                "optim": self.optim.state_dict(),
                "mixer": self.mixer.state_dict(),
                "last_ckpt_tokens": self.last_ckpt_tokens,
                "saved_stable": self.saved_stable,
                "torch_rng": torch.get_rng_state(),
            })
        torch.save(payload, self._ckpt_path(tag))
        kind = "full" if resumable else "model-only"
        print(f"[train] saved {tag} ({kind}) @ {self.tokens/1e9:.3f}B tokens")

    def maybe_resume(self):
        last = self._ckpt_path("last")
        if self.branch and not last.exists():
            ck = torch.load(self.branch["from"], map_location="cpu", weights_only=False)
            self.raw_model.load_state_dict(ck["model"])
            if self.branch.get("load_optim", True):
                self.optim.load_state_dict(ck["optim"])
            if self.rank == 0:
                print(f"[train] branched from {self.branch['from']} "
                      f"@ {ck['tokens']/1e9:.3f}B (cooldown {self.target_tokens/1e9:.1f}B)")
            return
        if last.exists():
            ck = torch.load(last, map_location="cpu", weights_only=False)
            self.raw_model.load_state_dict(ck["model"])
            self.optim.load_state_dict(ck["optim"])
            self.mixer.load_state_dict(ck["mixer"])
            self.step = ck["step"]; self.tokens = ck["tokens"]
            self.last_ckpt_tokens = ck["last_ckpt_tokens"]
            self.saved_stable = ck.get("saved_stable", False)
            self.marks_done = {m for m in self.stable_marks if m <= self.tokens}
            torch.set_rng_state(ck["torch_rng"])
            if self.rank == 0:
                print(f"[train] resumed @ step {self.step}, {self.tokens/1e9:.3f}B tokens")

    # ---- data ----
    def _next_micro_batches(self):
        """Return grad_accum micro-batches of (x, y) on device for this rank."""
        arr, counts = self.mixer.rank_batch(self.global_windows, self.rank, self.world)
        # arr: (global_windows/world, seq_len+1)
        t = torch.from_numpy(arr.astype(np.int64))
        x = t[:, :-1].to(self.device, non_blocking=True)
        y = t[:, 1:].to(self.device, non_blocking=True)
        micros = [(x[i:i + self.micro_bsz], y[i:i + self.micro_bsz])
                  for i in range(0, x.size(0), self.micro_bsz)]
        return micros, counts

    # ---- eval ----
    def _eval_sources(self):
        from .langs import LANGS
        srcs = {}
        for l in self.cfg["langs"]:
            try:
                from .eval.bpb import load_holdout
                h = load_holdout(l, self.cfg["train"].get("eval_docs", 500))
                if h:
                    srcs[f"holdout_{l}"] = h
            except Exception:
                pass
        try:
            from . import flores
            par = flores.load_parallel(list(self.cfg["langs"]), "dev")
            for l, sents in par.items():
                srcs[f"flores_{l}"] = sents
        except Exception as e:
            if self.rank == 0:
                print(f"[train] flores eval skipped: {e}")
        return srcs

    def evaluate(self):
        if self.rank != 0:
            return {}
        from .eval.bpb import eval_sources
        from .tok.wrapper import Tok
        from .paths import tokenizer_dir
        tok = Tok(tokenizer_dir(self.cfg["tok_name"]))
        res = eval_sources(self.raw_model, tok, self._eval_sources(),
                           self.device, self.seq_len)
        self.model.train()
        return res

    # ---- loop ----
    def train(self):
        self.maybe_resume()
        self.model.train()
        t0 = time.time()
        log_every = self.cfg["train"].get("log_every", 20)
        while self.tokens < self.target_tokens:
            lr = lr_at(self.tokens, self.sched)
            for g in self.optim.param_groups:
                g["lr"] = lr

            micros, counts = self._next_micro_batches()
            self.optim.zero_grad(set_to_none=True)
            loss_acc = 0.0
            for j, (x, y) in enumerate(micros):
                sync = (not self.dist) or (j == len(micros) - 1)
                ctx = self.model.no_sync() if (self.dist and not sync) else _null()
                with ctx:
                    with torch.autocast("cuda", dtype=torch.bfloat16) if self.is_cuda else _null():
                        _, loss = self.model(x, y)
                    (loss / len(micros)).backward()
                loss_acc += loss.detach().item() / len(micros)
            torch.nn.utils.clip_grad_norm_(self.raw_model.parameters(), self.grad_clip)
            self.optim.step()

            self.tokens += self.tokens_per_step
            self.step += 1

            if self.step % log_every == 0:
                dt = time.time() - t0
                tps = self.tokens_per_step * log_every / dt if dt > 0 else 0
                rec = {
                    "step": self.step, "tokens": self.tokens, "lr": lr,
                    "loss": loss_acc, "tok_per_s": round(tps),
                    "mix": self.mixer.stats(),
                }
                _log(self.rank, self.log_path, rec)
                if self.wandb:
                    self.wandb.log({**rec, "tokens_b": self.tokens / 1e9}, step=self.step)
                if self.rank == 0:
                    print(f"[train] step {self.step} | {self.tokens/1e9:.2f}B | "
                          f"loss {loss_acc:.4f} | lr {lr:.2e} | {tps/1e3:.0f}k tok/s")
                t0 = time.time()

            # stable checkpoint exactly once, at the trunk's decay boundary
            if (not self.branch and not self.saved_stable
                    and self.tokens >= stable_end_tokens(self.sched)):
                self.save("stable")
                self.saved_stable = True

            # named branch points for cooldown extensions (100B runs)
            for mark in self.stable_marks:
                if mark not in self.marks_done and self.tokens >= mark:
                    self.save(f"stable_{int(mark/1e6)}M")
                    self.marks_done.add(mark)

            # log-spaced checkpoint + eval
            if self.tokens - self.last_ckpt_tokens >= ckpt_interval(self.tokens, self.ckpt_table):
                self.last_ckpt_tokens = self.tokens
                self.save("last")
                self.save(f"step{self.step}_{int(self.tokens/1e6)}M", resumable=False)
                res = self.evaluate()
                _log(self.rank, self.log_path,
                     {"step": self.step, "tokens": self.tokens, "eval": res})
                if self.wandb and res:
                    self.wandb.log({f"eval/{k}_bpb": v["bpb"] for k, v in res.items()} |
                                   {f"eval/{k}_ppl": v["ppl_token"] for k, v in res.items()},
                                   step=self.step)
                if self.rank == 0 and res:
                    brief = {k: round(v["bpb"], 4) for k, v in res.items()}
                    print(f"[eval] {self.tokens/1e9:.2f}B: {brief}")

        self.save("last")
        self.save("final", resumable=False)
        res = self.evaluate()
        _log(self.rank, self.log_path,
             {"step": self.step, "tokens": self.tokens, "eval_final": res})
        if self.wandb:
            if res:
                self.wandb.log({f"eval_final/{k}_bpb": v["bpb"] for k, v in res.items()} |
                               {f"eval_final/{k}_ppl": v["ppl_token"] for k, v in res.items()},
                               step=self.step)
            self.wandb.finish()
        if self.rank == 0:
            print(f"[train] DONE {self.cfg['name']} @ {self.tokens/1e9:.2f}B tokens")
        if self.dist:
            import torch.distributed as dist
            dist.destroy_process_group()


class _null:
    def __enter__(self): return self
    def __exit__(self, *a): return False


def run_from_config(cfg: dict):
    Trainer(cfg).train()