path: root/trm/pretrain.py

from typing import Optional, Any, Sequence, List
from dataclasses import dataclass, replace
import os
import math
import yaml
import shutil
import copy

import torch
import torch.distributed as dist
from torch import nn
from torch.utils.data import DataLoader

import tqdm
import wandb
import coolname
import hydra
import pydantic
from omegaconf import DictConfig
from adam_atan2 import AdamATan2

from puzzle_dataset import PuzzleDataset, PuzzleDatasetConfig, PuzzleDatasetMetadata
from utils.functions import load_model_class, get_model_source_path
from models.sparse_embedding import CastedSparseEmbeddingSignSGD_Distributed
from models.ema import EMAHelper


class LossConfig(pydantic.BaseModel):
    model_config = pydantic.ConfigDict(extra='allow')
    name: str


class ArchConfig(pydantic.BaseModel):
    model_config = pydantic.ConfigDict(extra='allow')
    name: str
    loss: LossConfig


class EvaluatorConfig(pydantic.BaseModel):
    model_config = pydantic.ConfigDict(extra="allow")
    name: str


class PretrainConfig(pydantic.BaseModel):
    # Config
    arch: ArchConfig
    # Data
    data_paths: List[str]
    data_paths_test: List[str] = []
    # Evaluators
    evaluators: List[EvaluatorConfig] = []

    # Hyperparams
    global_batch_size: int
    epochs: int

    lr: float
    lr_min_ratio: float
    lr_warmup_steps: int

    weight_decay: float
    beta1: float
    beta2: float

    # Puzzle embedding
    puzzle_emb_lr: float
    puzzle_emb_weight_decay: float

    # Names
    project_name: Optional[str] = None
    run_name: Optional[str] = None
    load_checkpoint: Optional[str] = None
    checkpoint_path: Optional[str] = None

    # Extras
    seed: int = 0
    checkpoint_every_eval: bool = False
    eval_interval: Optional[int] = None
    min_eval_interval: Optional[int] = 0 # when to start eval
    eval_save_outputs: List[str] = []

    ema: bool = False # use Exponential-Moving-Average
    ema_rate: float = 0.999 # EMA-rate
    freeze_weights: bool = False # If True, freeze weights and only learn the embeddings

    trajectory_augment: bool = False
    trajectory_n: int = 4
    trajectory_noise_std: float = 1e-3
    trajectory_noise_min: Optional[float] = None
    trajectory_noise_max: Optional[float] = None
    trajectory_noise_sampling: str = "loguniform"
    trajectory_sigma_start: Optional[float] = 0.0
    trajectory_sigma_ramp_steps: int = 5000
    trajectory_perturb: str = "both"
    trajectory_micro_batch: int = 0
    trajectory_parallel: bool = False
    trajectory_directional: bool = False
    trajectory_directional_candidates: int = 4
    trajectory_directional_fd_eps: float = 3e-2
    trajectory_directional_horizon: int = 0
    trajectory_directional_sign: str = "alternate"

@dataclass
class TrainState:
    model: nn.Module
    optimizers: Sequence[torch.optim.Optimizer]
    optimizer_lrs: Sequence[float]
    carry: Any

    step: int
    total_steps: int


def create_dataloader(config: PretrainConfig, split: str, rank: int, world_size: int, **kwargs):
    dataset = PuzzleDataset(PuzzleDatasetConfig(
        seed=config.seed,
        dataset_paths=config.data_paths_test if len(config.data_paths_test)>0 and split=="test" else config.data_paths,
        rank=rank,
        num_replicas=world_size,
        **kwargs
    ), split=split)
    dataloader = DataLoader(
        dataset,
        batch_size=None,
        num_workers=1,
        prefetch_factor=8,
        pin_memory=True,
        persistent_workers=True
    )
    return dataloader, dataset.metadata


def create_model(config: PretrainConfig, train_metadata: PuzzleDatasetMetadata, rank: int, world_size: int):
    model_batch_size = config.global_batch_size // world_size
    if config.trajectory_augment and config.trajectory_parallel:
        model_batch_size *= config.trajectory_n

    model_cfg = dict(
        **config.arch.__pydantic_extra__,  # type: ignore
        batch_size=model_batch_size,
        vocab_size=train_metadata.vocab_size,
        seq_len=train_metadata.seq_len,
        num_puzzle_identifiers=train_metadata.num_puzzle_identifiers,
        causal=False  # Non-autoregressive
    )

    # Instantiate model with loss head
    model_cls = load_model_class(config.arch.name)
    loss_head_cls = load_model_class(config.arch.loss.name)

    with torch.device("cuda"):
        model: nn.Module = model_cls(model_cfg)
        print(model)
        model = loss_head_cls(model, **config.arch.loss.__pydantic_extra__)  # type: ignore
        if "DISABLE_COMPILE" not in os.environ:
            model = torch.compile(model)  # type: ignore

        # Load checkpoint
        if rank == 0:
            load_checkpoint(model, config)

        # Broadcast parameters from rank 0
        if world_size > 1:
            with torch.no_grad():
                for param in list(model.parameters()) + list(model.buffers()):
                    dist.broadcast(param, src=0)

    # Optimizers and lr
    if config.arch.puzzle_emb_ndim == 0:
        optimizers = [
            AdamATan2(
                model.parameters(),
                lr=0,  # Needs to be set by scheduler
                weight_decay=config.weight_decay,
                betas=(config.beta1, config.beta2)
            )
        ]
        optimizer_lrs = [
            config.lr
        ]
    elif config.freeze_weights:
        optimizers = [
            CastedSparseEmbeddingSignSGD_Distributed(
                model.model.puzzle_emb.buffers(),  # type: ignore
                lr=0,  # Needs to be set by scheduler
                weight_decay=config.puzzle_emb_weight_decay,
                world_size=world_size
            )
        ]
        optimizer_lrs = [
            config.puzzle_emb_lr
        ]
    else:
        optimizers = [
            CastedSparseEmbeddingSignSGD_Distributed(
                model.model.puzzle_emb.buffers(),  # type: ignore
                lr=0,  # Needs to be set by scheduler
                weight_decay=config.puzzle_emb_weight_decay,
                world_size=world_size
            ),
            AdamATan2(
                model.parameters(),
                lr=0,  # Needs to be set by scheduler
                weight_decay=config.weight_decay,
                betas=(config.beta1, config.beta2)
            )
        ]
        optimizer_lrs = [
            config.puzzle_emb_lr,
            config.lr
        ]

    return model, optimizers, optimizer_lrs

def mix_weights_direct(device, alpha, net, nets):
    sd = []
    for i in range(len(nets)):
        sd += [nets[i].state_dict()]
    sd_alpha = {}
    for k in sd[0].keys():
        comb_net = alpha[0]*sd[0][k].to(device)
        for i in range(1,len(nets)):
            comb_net += alpha[i]*sd[i][k].to(device)
        sd_alpha[k] =  comb_net
    net.load_state_dict(sd_alpha)
    return net

def cosine_schedule_with_warmup_lr_lambda(
    current_step: int, *, base_lr: float, num_warmup_steps: int, num_training_steps: int, min_ratio: float = 0.0, num_cycles: float = 0.5
):
    if current_step < num_warmup_steps:
        return base_lr * float(current_step) / float(max(1, num_warmup_steps))

    progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))
    return base_lr * (min_ratio + max(0.0, (1 - min_ratio) * 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress))))


def init_train_state(config: PretrainConfig, train_metadata: PuzzleDatasetMetadata, rank: int, world_size: int):
    # Estimated total training steps
    total_steps = int(config.epochs * train_metadata.total_groups * train_metadata.mean_puzzle_examples / config.global_batch_size)

    # Model
    model, optimizers, optimizer_lrs = create_model(config, train_metadata, rank=rank, world_size=world_size)

    return TrainState(
        step=0,
        total_steps=total_steps,

        model=model,
        optimizers=optimizers,
        optimizer_lrs=optimizer_lrs,
        carry=None
    )


def save_train_state(config: PretrainConfig, train_state: TrainState):
    # FIXME: Only saved model.
    if config.checkpoint_path is None:
        return

    os.makedirs(config.checkpoint_path, exist_ok=True)
    torch.save(train_state.model.state_dict(), os.path.join(config.checkpoint_path, f"step_{train_state.step}"))


def load_checkpoint(model: nn.Module, config: PretrainConfig):
    if config.load_checkpoint is not None:
        print(f"Loading checkpoint {config.load_checkpoint}")

        # Load state dict
        state_dict = torch.load(config.load_checkpoint, map_location="cuda")

        # Resize and reset puzzle emb if needed
        puzzle_emb_name = "_orig_mod.model.inner.puzzle_emb.weights"
        expected_shape: torch.Size = model.model.puzzle_emb.weights.shape  # type: ignore
        if puzzle_emb_name in state_dict:
            puzzle_emb = state_dict[puzzle_emb_name]
            if puzzle_emb.shape != expected_shape:
                print(f"Resetting puzzle embedding as shape is different. Found {puzzle_emb.shape}, Expected {expected_shape}")
                # Re-initialize using mean
                state_dict[puzzle_emb_name] = (
                    torch.mean(puzzle_emb, dim=0, keepdim=True).expand(expected_shape).contiguous()
                )
        model.load_state_dict(state_dict, assign=True)


def compute_lr(base_lr: float, config: PretrainConfig, train_state: TrainState):
    return cosine_schedule_with_warmup_lr_lambda(
        current_step=train_state.step,
        base_lr=base_lr,
        num_warmup_steps=round(config.lr_warmup_steps),
        num_training_steps=train_state.total_steps,
        min_ratio=config.lr_min_ratio
    )



def _unwrap_loss_head(model: nn.Module):
    return getattr(model, "_orig_mod", model)


def _unit_noise_like(tensor: torch.Tensor, sampling: str):
    if sampling == "uniform":
        return ((2.0 * torch.rand(tensor.shape, device=tensor.device, dtype=torch.float32) - 1.0) * math.sqrt(3.0)).to(tensor.dtype)
    return torch.randn(tensor.shape, device=tensor.device, dtype=torch.float32).to(tensor.dtype)


def _trajectory_noise_target(config: PretrainConfig, step: int) -> float:
    if config.trajectory_sigma_ramp_steps <= 0:
        return config.trajectory_noise_std
    start = config.trajectory_sigma_start if config.trajectory_sigma_start is not None else config.trajectory_noise_std
    frac = min(max(step / config.trajectory_sigma_ramp_steps, 0.0), 1.0)
    return float(start + frac * (config.trajectory_noise_std - start))


def _sample_noise_stds(config: PretrainConfig, batch_size: int, device: torch.device, step: int):
    target = _trajectory_noise_target(config, step)
    if target <= 0:
        return torch.zeros(batch_size, device=device, dtype=torch.float32)
    if config.trajectory_noise_sampling == "loguniform":
        scale = 1.0 if config.trajectory_noise_std <= 0 else target / config.trajectory_noise_std
        hi = float(config.trajectory_noise_max if config.trajectory_noise_max is not None else config.trajectory_noise_std) * scale
        lo = float(config.trajectory_noise_min if config.trajectory_noise_min is not None else max(hi / 10.0, 1e-8)) * scale
        hi = max(hi, 1e-12)
        lo = max(min(lo, hi), 1e-12)
        u = torch.rand(batch_size, device=device, dtype=torch.float32)
        return torch.exp(math.log(lo) + u * (math.log(hi) - math.log(lo)))
    return torch.full((batch_size,), target, device=device, dtype=torch.float32)


def _add_initial_noise(config: PretrainConfig, inner_carry: Any, noise_stds: torch.Tensor):
    if noise_stds.numel() == 0 or float(noise_stds.max().item()) <= 0:
        return inner_carry
    view_shape = (noise_stds.shape[0],) + (1,) * (inner_carry.z_H.ndim - 1)
    scale = noise_stds.view(view_shape)
    z_h = inner_carry.z_H
    z_l = inner_carry.z_L
    if config.trajectory_perturb in ("h", "both"):
        z_h = z_h + scale.to(z_h.dtype) * _unit_noise_like(z_h, config.trajectory_noise_sampling)
    if config.trajectory_perturb in ("l", "both"):
        z_l = z_l + scale.to(z_l.dtype) * _unit_noise_like(z_l, config.trajectory_noise_sampling)
    return replace(inner_carry, z_H=z_h, z_L=z_l)


def _directional_fields(config: PretrainConfig):
    if config.trajectory_perturb == "h":
        return ("z_H",)
    if config.trajectory_perturb == "l":
        return ("z_L",)
    return ("z_H", "z_L")


def _index_inner_carry(inner_carry: Any, indices: torch.Tensor):
    inner_type = type(inner_carry)
    return inner_type(**{
        name: getattr(inner_carry, name).index_select(0, indices).contiguous()
        for name in inner_carry.__dataclass_fields__.keys()
    })


def _cat_inner_carries(carries: Sequence[Any]):
    inner_type = type(carries[0])
    return inner_type(**{
        name: torch.cat([getattr(c, name) for c in carries], dim=0)
        for name in carries[0].__dataclass_fields__.keys()
    })


def _split_inner_carry(inner_carry: Any, first_size: int):
    inner_type = type(inner_carry)
    first = {}
    second = {}
    for name in inner_carry.__dataclass_fields__.keys():
        value = getattr(inner_carry, name)
        first[name] = value[:first_size].contiguous()
        second[name] = value[first_size:].contiguous()
    return inner_type(**first), inner_type(**second)


def _repeat_batch_interleave(batch: Any, repeats: int):
    return {k: v.repeat_interleave(repeats, dim=0) for k, v in batch.items()}


def _cat_batches(a: Any, b: Any):
    return {k: torch.cat([a[k], b[k]], dim=0) for k in a}


def _rand_unit_direction_candidates(config: PretrainConfig, inner_carry: Any, candidates: int):
    fields = _directional_fields(config)
    dirs = {}
    norm_sq = None
    for name in inner_carry.__dataclass_fields__.keys():
        value = getattr(inner_carry, name)
        if name in fields:
            direction = torch.randn(
                (value.shape[0], candidates) + tuple(value.shape[1:]),
                device=value.device,
                dtype=torch.float32,
            )
            term = direction.flatten(2).square().sum(-1)
            norm_sq = term if norm_sq is None else norm_sq + term
            dirs[name] = direction
        else:
            dirs[name] = torch.zeros(
                (value.shape[0], candidates) + tuple(value.shape[1:]),
                device=value.device,
                dtype=torch.float32,
            )

    if norm_sq is None:
        raise ValueError(f"unknown trajectory_perturb={config.trajectory_perturb!r}")

    norm = torch.sqrt(norm_sq).clamp_min(1e-30)
    out = {}
    for name, direction in dirs.items():
        view_shape = direction.shape[:2] + (1,) * (direction.ndim - 2)
        out[name] = (direction / norm.view(view_shape)).to(getattr(inner_carry, name).dtype)
    return out


def _make_shadow_inner(inner_carry: Any, dirs: dict[str, torch.Tensor], eps: float):
    inner_type = type(inner_carry)
    fields = {}
    for name in inner_carry.__dataclass_fields__.keys():
        value = getattr(inner_carry, name)
        perturbed = value[:, None] + eps * dirs[name]
        fields[name] = perturbed.reshape((value.shape[0] * dirs[name].shape[1],) + tuple(value.shape[1:])).detach()
    return inner_type(**fields)


def _separation(main: Any, shadow: Any, candidates: int):
    sep_sq = None
    batch_size = getattr(main, next(iter(main.__dataclass_fields__.keys()))).shape[0]
    for name in main.__dataclass_fields__.keys():
        main_value = getattr(main, name)
        shadow_value = getattr(shadow, name).reshape((batch_size, candidates) + tuple(main_value.shape[1:]))
        diff_sq = (shadow_value.float() - main_value[:, None].float()).flatten(2).square().sum(-1)
        sep_sq = diff_sq if sep_sq is None else sep_sq + diff_sq
    return torch.sqrt(sep_sq).clamp_min(1e-30)


def _gather_direction(dirs: dict[str, torch.Tensor], best_idx: torch.Tensor):
    batch_size = best_idx.shape[0]
    row_idx = torch.arange(batch_size, device=best_idx.device)
    return {name: value[row_idx, best_idx].contiguous() for name, value in dirs.items()}


def _directional_signs(config: PretrainConfig, branch_idx: int, count: int, device: torch.device):
    mode = config.trajectory_directional_sign
    if mode == "random":
        return torch.where(
            torch.rand(count, device=device, dtype=torch.float32) < 0.5,
            torch.full((count,), -1.0, device=device, dtype=torch.float32),
            torch.full((count,), 1.0, device=device, dtype=torch.float32),
        )
    if mode == "negative":
        return torch.full((count,), -1.0, device=device, dtype=torch.float32)
    if mode == "positive":
        return torch.full((count,), 1.0, device=device, dtype=torch.float32)
    if mode != "alternate":
        raise ValueError(f"unknown trajectory_directional_sign={mode!r}")
    sign = 1.0 if branch_idx % 2 == 1 else -1.0
    return torch.full((count,), sign, device=device, dtype=torch.float32)


@torch.no_grad()
def _select_directional_perturbation(
    config: PretrainConfig,
    base: nn.Module,
    inner_carry: Any,
    batch: Any,
    active_mask: torch.Tensor,
):
    active_indices = torch.nonzero(active_mask, as_tuple=False).flatten()
    active_count = int(active_indices.numel())
    if active_count == 0:
        return None, None, {
            "active": 0,
            "growth_sum": 0.0,
            "growth_count": 0,
        }

    candidates = max(int(config.trajectory_directional_candidates), 1)
    fd_eps = float(config.trajectory_directional_fd_eps)
    horizon = int(config.trajectory_directional_horizon)
    if horizon <= 0:
        horizon = int(base.config.halt_max_steps)  # type: ignore[attr-defined]
    horizon = max(horizon, 1)

    active_inner = _index_inner_carry(inner_carry, active_indices)
    active_batch = {k: v.index_select(0, active_indices).contiguous() for k, v in batch.items()}
    dirs = _rand_unit_direction_candidates(config, active_inner, candidates)
    main = type(active_inner)(**{
        name: getattr(active_inner, name).detach()
        for name in active_inner.__dataclass_fields__.keys()
    })
    shadow = _make_shadow_inner(main, dirs, fd_eps)
    combined = _cat_inner_carries([main, shadow])
    combined_batch = _cat_batches(active_batch, _repeat_batch_interleave(active_batch, candidates))

    was_training = base.inner.training  # type: ignore[attr-defined]
    base.inner.eval()  # type: ignore[attr-defined]
    try:
        for _ in range(horizon):
            combined, _logits, _q = base.inner(combined, combined_batch)  # type: ignore[attr-defined]
    finally:
        if was_training:
            base.inner.train()  # type: ignore[attr-defined]

    main_final, shadow_final = _split_inner_carry(combined, active_count)
    sep = _separation(main_final, shadow_final, candidates)
    best_idx = sep.argmax(dim=1)
    best_sep = sep.gather(1, best_idx[:, None]).squeeze(1)
    best_dirs = _gather_direction(dirs, best_idx)
    growth = torch.log(best_sep / max(fd_eps, 1e-30)).float() / horizon
    return active_indices, best_dirs, {
        "active": active_count,
        "growth_sum": float(growth.sum().item()),
        "growth_count": active_count,
    }


def _add_directional_initial_noise(
    config: PretrainConfig,
    base: nn.Module,
    inner_carry: Any,
    batch: Any,
    noise_stds: torch.Tensor,
    active_mask: torch.Tensor,
    branch_idx: int,
):
    selected_indices, selected_dirs, stats = _select_directional_perturbation(
        config,
        base,
        inner_carry,
        batch,
        active_mask,
    )
    if selected_indices is None or selected_dirs is None:
        return inner_carry, stats

    signs = _directional_signs(config, branch_idx, selected_indices.shape[0], noise_stds.device)
    scales = noise_stds.index_select(0, selected_indices) * signs
    fields = {}
    for name in inner_carry.__dataclass_fields__.keys():
        value = getattr(inner_carry, name)
        updated = value.clone()
        direction = selected_dirs[name].to(value.dtype)
        view_shape = (scales.shape[0],) + (1,) * (direction.ndim - 1)
        updated.index_copy_(0, selected_indices, value.index_select(0, selected_indices) + scales.view(view_shape).to(value.dtype) * direction)
        fields[name] = updated
    return type(inner_carry)(**fields), stats


def _token_loss(loss_fn, logits, labels, mask):
    kwargs = {"ignore_index": -100}
    code = getattr(loss_fn, "__code__", None)
    arg_names = code.co_varnames[: code.co_argcount + code.co_kwonlyargcount] if code is not None else ()
    if "valid_mask" in arg_names:
        kwargs["valid_mask"] = mask
    return loss_fn(logits, labels, **kwargs)


def _fixed_unroll_branch_loss(config: PretrainConfig, head: nn.Module, batch: Any, noise_stds: Optional[torch.Tensor]):
    base = head.model  # type: ignore[attr-defined]
    with torch.device("cuda"):
        carry = base.initial_carry(batch)
    reset_flag = torch.ones_like(carry.halted)
    inner_carry = base.inner.reset_carry(reset_flag, carry.inner_carry)
    if noise_stds is not None:
        inner_carry = _add_initial_noise(config, inner_carry, noise_stds)

    batch_size = batch["inputs"].shape[0]
    loss_sum = torch.zeros((), device=batch["inputs"].device, dtype=torch.float32)
    last_exact = torch.zeros((), device=batch["inputs"].device, dtype=torch.float32)
    for act_step in range(base.config.halt_max_steps):
        inner_carry, logits, (q_halt_logits, q_continue_logits) = base.inner(inner_carry, batch)
        labels = batch["labels"]
        with torch.no_grad():
            mask = labels != -100
            loss_counts = mask.sum(-1)
            loss_divisor = loss_counts.clamp_min(1).unsqueeze(-1)
            is_correct = mask & (torch.argmax(logits, dim=-1) == labels)
            seq_is_correct = is_correct.sum(-1) == loss_counts
            last_exact = seq_is_correct.to(torch.float32).sum()

        lm_loss = (_token_loss(head.loss_fn, logits, labels, mask) / loss_divisor).sum()
        q_halt_loss = torch.nn.functional.binary_cross_entropy_with_logits(
            q_halt_logits,
            seq_is_correct.to(q_halt_logits.dtype),
            reduction="sum",
        )
        q_continue_loss = torch.zeros_like(q_halt_loss)
        if not getattr(base.config, "no_ACT_continue", False):
            with torch.no_grad():
                next_q_halt_logits, next_q_continue_logits = base.inner(inner_carry, batch)[-1]
                is_last = act_step + 1 >= base.config.halt_max_steps
                target_q_continue = torch.sigmoid(
                    torch.where(
                        torch.full_like(next_q_halt_logits, is_last, dtype=torch.bool),
                        next_q_halt_logits,
                        torch.maximum(next_q_halt_logits, next_q_continue_logits),
                    )
                )
            q_continue_loss = torch.nn.functional.binary_cross_entropy_with_logits(q_continue_logits, target_q_continue, reduction="sum")
        loss_sum = loss_sum + lm_loss + 0.5 * (q_halt_loss + q_continue_loss)

    return loss_sum / max(base.config.halt_max_steps, 1), last_exact, batch_size


def _prepare_noisy_stream_carry(config: PretrainConfig, base: nn.Module, carry: Any, batch: Any, step: int, branch_idx: int):
    reset_mask = carry.halted
    new_inner = base.inner.reset_carry(reset_mask, carry.inner_carry)
    noise_stds = _sample_noise_stds(config, batch["inputs"].shape[0], batch["inputs"].device, step)
    noise_stds = torch.where(reset_mask, noise_stds, torch.zeros_like(noise_stds))
    directional_stats = {
        "active": int((noise_stds > 0).sum().detach().cpu().item()),
        "growth_sum": 0.0,
        "growth_count": 0,
    }
    if config.trajectory_directional and float(noise_stds.max().item()) > 0:
        new_inner, directional_stats = _add_directional_initial_noise(
            config,
            base,
            new_inner,
            batch,
            noise_stds,
            noise_stds > 0,
            branch_idx,
        )
    else:
        new_inner = _add_initial_noise(config, new_inner, noise_stds)
    new_steps = torch.where(reset_mask, 0, carry.steps)
    new_current_data = {
        k: torch.where(reset_mask.view((-1,) + (1,) * (batch[k].ndim - 1)), batch[k], v)
        for k, v in carry.current_data.items()
    }
    return (
        replace(
            carry,
            inner_carry=new_inner,
            steps=new_steps,
            halted=torch.zeros_like(reset_mask),
            current_data=new_current_data,
        ),
        directional_stats,
    )


def _merge_trajectory_carries(carries: Sequence[Any]):
    inner_type = type(carries[0].inner_carry)
    inner_fields = carries[0].inner_carry.__dataclass_fields__.keys()
    inner_carry = inner_type(**{
        name: torch.cat([getattr(c.inner_carry, name) for c in carries], dim=0)
        for name in inner_fields
    })
    current_data = {
        k: torch.cat([c.current_data[k] for c in carries], dim=0)
        for k in carries[0].current_data
    }
    return replace(
        carries[0],
        inner_carry=inner_carry,
        steps=torch.cat([c.steps for c in carries], dim=0),
        halted=torch.cat([c.halted for c in carries], dim=0),
        current_data=current_data,
    )


def _split_trajectory_carry(carry: Any, branch_size: int, num_branches: int):
    inner_type = type(carry.inner_carry)
    inner_fields = carry.inner_carry.__dataclass_fields__.keys()
    inner_chunks = {
        name: [chunk.contiguous() for chunk in getattr(carry.inner_carry, name).split(branch_size, dim=0)]
        for name in inner_fields
    }
    steps_chunks = [chunk.contiguous() for chunk in carry.steps.split(branch_size, dim=0)]
    halted_chunks = [chunk.contiguous() for chunk in carry.halted.split(branch_size, dim=0)]
    data_chunks = {
        k: [chunk.contiguous() for chunk in v.split(branch_size, dim=0)]
        for k, v in carry.current_data.items()
    }
    return [
        replace(
            carry,
            inner_carry=inner_type(**{name: inner_chunks[name][idx] for name in inner_fields}),
            steps=steps_chunks[idx],
            halted=halted_chunks[idx],
            current_data={k: chunks[idx] for k, chunks in data_chunks.items()},
        )
        for idx in range(num_branches)
    ]


def _repeat_batch_for_trajectories(batch: Any, num_branches: int):
    return {
        k: torch.cat([v for _ in range(num_branches)], dim=0)
        for k, v in batch.items()
    }


def _split_outputs(outputs: Any, branch_size: int, num_branches: int):
    return [
        {
            k: v.narrow(0, idx * branch_size, branch_size)
            for k, v in outputs.items()
        }
        for idx in range(num_branches)
    ]


def _branch_metrics_and_loss(head: nn.Module, carry: Any, outputs: Any):
    labels = carry.current_data["labels"]
    with torch.no_grad():
        mask = labels != -100
        loss_counts = mask.sum(-1)
        loss_divisor = loss_counts.clamp_min(1).unsqueeze(-1)
        is_correct = mask & (torch.argmax(outputs["logits"], dim=-1) == labels)
        seq_is_correct = is_correct.sum(-1) == loss_counts
        valid_metrics = carry.halted & (loss_counts > 0)
        metrics = {
            "count": valid_metrics.sum(),
            "accuracy": torch.where(valid_metrics, (is_correct.to(torch.float32) / loss_divisor).sum(-1), 0).sum(),
            "exact_accuracy": (valid_metrics & seq_is_correct).sum(),
            "q_halt_accuracy": (valid_metrics & ((outputs["q_halt_logits"] >= 0) == seq_is_correct)).sum(),
            "steps": torch.where(valid_metrics, carry.steps, 0).sum(),
        }

        lm_loss = (head.loss_fn(outputs["logits"], labels, ignore_index=-100) / loss_divisor).sum()
        q_halt_loss = torch.nn.functional.binary_cross_entropy_with_logits(
            outputs["q_halt_logits"],
            seq_is_correct.to(outputs["q_halt_logits"].dtype),
            reduction="sum",
        )
        q_continue_loss = torch.zeros_like(q_halt_loss)
        if "target_q_continue" in outputs:
            q_continue_loss = torch.nn.functional.binary_cross_entropy_with_logits(
                outputs["q_continue_logits"],
                outputs["target_q_continue"],
                reduction="sum",
            )
        metrics.update({
            "lm_loss": lm_loss,
            "q_halt_loss": q_halt_loss,
        })
        if "target_q_continue" in outputs:
            metrics["q_continue_loss"] = q_continue_loss
        total_loss = lm_loss + 0.5 * (q_halt_loss + q_continue_loss)
    return metrics, total_loss


def train_batch_trajectory_aug_parallel(config: PretrainConfig, train_state: TrainState, batch: Any, global_batch_size: int, rank: int, world_size: int):
    if world_size != 1:
        raise NotImplementedError("trajectory_parallel currently expects single-GPU baseline configs")
    train_state.step += 1
    if train_state.step > train_state.total_steps:
        return

    batch = {k: v.cuda() for k, v in batch.items()}
    head = _unwrap_loss_head(train_state.model)
    base = head.model  # type: ignore[attr-defined]
    num_branches = max(config.trajectory_n, 1)
    branch_size = batch["inputs"].shape[0]

    if train_state.carry is None or not isinstance(train_state.carry, list):
        with torch.device("cuda"):
            train_state.carry = [train_state.model.initial_carry(batch) for _ in range(num_branches)]  # type: ignore

    for optim in train_state.optimizers:
        optim.zero_grad()

    carries = []
    directional_active = 0
    directional_growth_sum = 0.0
    directional_growth_count = 0
    for branch_idx, carry in enumerate(train_state.carry):
        if branch_idx > 0:
            carry, stats = _prepare_noisy_stream_carry(config, base, carry, batch, train_state.step, branch_idx)
            directional_active += int(stats["active"])
            directional_growth_sum += float(stats["growth_sum"])
            directional_growth_count += int(stats["growth_count"])
        carries.append(carry)

    parallel_carry = _merge_trajectory_carries(carries)
    parallel_batch = _repeat_batch_for_trajectories(batch, num_branches)
    return_keys = ["logits", "q_halt_logits", "q_continue_logits", "target_q_continue"]
    parallel_carry, loss, _metrics, outputs, _ = train_state.model(
        carry=parallel_carry,
        batch=parallel_batch,
        return_keys=return_keys,
    )
    train_state.carry = _split_trajectory_carry(parallel_carry, branch_size, num_branches)
    (loss / (global_batch_size * num_branches)).backward()

    lr_this_step = None
    for optim, base_lr in zip(train_state.optimizers, train_state.optimizer_lrs):
        lr_this_step = compute_lr(base_lr, config, train_state)
        for param_group in optim.param_groups:
            param_group["lr"] = lr_this_step
        optim.step()
        optim.zero_grad()

    if rank == 0 and outputs is not None:
        split_outputs = _split_outputs(outputs, branch_size, num_branches)
        clean_metrics, clean_loss = _branch_metrics_and_loss(head, train_state.carry[0], split_outputs[0])
        noisy_loss = 0.0
        for branch_idx in range(1, num_branches):
            _, branch_loss = _branch_metrics_and_loss(head, train_state.carry[branch_idx], split_outputs[branch_idx])
            noisy_loss += float(branch_loss.detach().cpu())

        reduced_metrics = {f"train/{k}": float(v.detach().cpu()) for k, v in clean_metrics.items()}
        count = max(reduced_metrics.get("train/count", 1.0), 1.0)
        reduced_metrics = {
            k: v / (global_batch_size if k.endswith("loss") else count)
            for k, v in reduced_metrics.items()
        }
        reduced_metrics["train/lr"] = lr_this_step
        reduced_metrics["train/trajectory_clean_loss"] = float(clean_loss.detach().cpu()) / global_batch_size
        reduced_metrics["train/trajectory_noisy_loss"] = noisy_loss / max(num_branches - 1, 1) / global_batch_size
        reduced_metrics["train/trajectory_sigma"] = _trajectory_noise_target(config, train_state.step)
        reduced_metrics["train/trajectory_parallel"] = 1.0
        reduced_metrics["train/trajectory_directional"] = 1.0 if config.trajectory_directional else 0.0
        reduced_metrics["train/trajectory_directional_active"] = float(directional_active)
        if directional_growth_count > 0:
            reduced_metrics["train/trajectory_directional_growth"] = directional_growth_sum / directional_growth_count
        return reduced_metrics


def train_batch_trajectory_aug(config: PretrainConfig, train_state: TrainState, batch: Any, global_batch_size: int, rank: int, world_size: int):
    if config.trajectory_parallel:
        return train_batch_trajectory_aug_parallel(config, train_state, batch, global_batch_size, rank, world_size)
    if world_size != 1:
        raise NotImplementedError("trajectory_augment currently expects single-GPU baseline configs")
    train_state.step += 1
    if train_state.step > train_state.total_steps:
        return

    batch = {k: v.cuda() for k, v in batch.items()}
    head = _unwrap_loss_head(train_state.model)
    base = head.model  # type: ignore[attr-defined]

    if train_state.carry is None or not isinstance(train_state.carry, list):
        with torch.device("cuda"):
            train_state.carry = [train_state.model.initial_carry(batch) for _ in range(config.trajectory_n)]  # type: ignore

    for optim in train_state.optimizers:
        optim.zero_grad()

    clean_metrics = None
    clean_loss_value = 0.0
    noisy_loss_value = 0.0
    directional_active = 0
    directional_growth_sum = 0.0
    directional_growth_count = 0
    for branch_idx in range(config.trajectory_n):
        carry = train_state.carry[branch_idx]
        if branch_idx > 0:
            carry, stats = _prepare_noisy_stream_carry(config, base, carry, batch, train_state.step, branch_idx)
            directional_active += int(stats["active"])
            directional_growth_sum += float(stats["growth_sum"])
            directional_growth_count += int(stats["growth_count"])
        carry, loss, metrics, _, _ = train_state.model(carry=carry, batch=batch, return_keys=[])
        train_state.carry[branch_idx] = carry
        (loss / (global_batch_size * max(config.trajectory_n, 1))).backward()
        if branch_idx == 0:
            clean_metrics = metrics
            clean_loss_value = float(loss.detach().cpu())
        else:
            noisy_loss_value += float(loss.detach().cpu())

    lr_this_step = None
    for optim, base_lr in zip(train_state.optimizers, train_state.optimizer_lrs):
        lr_this_step = compute_lr(base_lr, config, train_state)
        for param_group in optim.param_groups:
            param_group["lr"] = lr_this_step
        optim.step()
        optim.zero_grad()

    if rank == 0 and clean_metrics is not None:
        reduced_metrics = {f"train/{k}": float(v.detach().cpu()) for k, v in clean_metrics.items()}
        count = max(reduced_metrics.get("train/count", 1.0), 1.0)
        reduced_metrics = {
            k: v / (global_batch_size if k.endswith("loss") else count)
            for k, v in reduced_metrics.items()
        }
        reduced_metrics["train/lr"] = lr_this_step
        reduced_metrics["train/trajectory_clean_loss"] = clean_loss_value / global_batch_size
        reduced_metrics["train/trajectory_noisy_loss"] = noisy_loss_value / max(config.trajectory_n - 1, 1) / global_batch_size
        reduced_metrics["train/trajectory_sigma"] = _trajectory_noise_target(config, train_state.step)
        reduced_metrics["train/trajectory_directional"] = 1.0 if config.trajectory_directional else 0.0
        reduced_metrics["train/trajectory_directional_active"] = float(directional_active)
        if directional_growth_count > 0:
            reduced_metrics["train/trajectory_directional_growth"] = directional_growth_sum / directional_growth_count
        return reduced_metrics


def create_evaluators(config: PretrainConfig, eval_metadata: PuzzleDatasetMetadata) -> List[Any]:
    data_paths =config.data_paths_test if len(config.data_paths_test)>0 else config.data_paths
    # Initialize evaluators
    evaluators = []
    for cfg in config.evaluators:
        for data_path in data_paths:
            cls = load_model_class(cfg.name, "evaluators.")(
                data_path=data_path, eval_metadata=eval_metadata, **cfg.__pydantic_extra__
            )  # type: ignore
            evaluators.append(cls)

    return evaluators

def train_batch(config: PretrainConfig, train_state: TrainState, batch: Any, global_batch_size: int, rank: int, world_size: int):
    if config.trajectory_augment:
        return train_batch_trajectory_aug(config, train_state, batch, global_batch_size, rank, world_size)

    train_state.step += 1
    if train_state.step > train_state.total_steps:  # At most train_total_steps
        return

    # To device
    batch = {k: v.cuda() for k, v in batch.items()}

    # Init carry if it is None
    if train_state.carry is None:
        with torch.device("cuda"):
            train_state.carry = train_state.model.initial_carry(batch)  # type: ignore

    # Forward
    train_state.carry, loss, metrics, _, _ = train_state.model(carry=train_state.carry, batch=batch, return_keys=[])

    ((1 / global_batch_size) * loss).backward()

    # Allreduce
    if world_size > 1:
        for param in train_state.model.parameters():
            if param.grad is not None:
                dist.all_reduce(param.grad)
            
    # Apply optimizer
    lr_this_step = None    
    for optim, base_lr in zip(train_state.optimizers, train_state.optimizer_lrs):
        lr_this_step = compute_lr(base_lr, config, train_state)

        for param_group in optim.param_groups:
            param_group['lr'] = lr_this_step
            
        optim.step()
        optim.zero_grad()

    # Reduce metrics
    if len(metrics):
        assert not any(v.requires_grad for v in metrics.values())

        metric_keys = list(sorted(metrics.keys()))  # Sort keys to guarantee all processes use the same order.
        # Reduce and reconstruct
        metric_values = torch.stack([metrics[k] for k in metric_keys])
        if world_size > 1:
            dist.reduce(metric_values, dst=0)

        if rank == 0:
            metric_values = metric_values.cpu().numpy()
            reduced_metrics = {k: metric_values[i] for i, k in enumerate(metric_keys)}
            
            # Postprocess
            count = max(reduced_metrics["count"], 1)  # Avoid NaNs
            reduced_metrics = {f"train/{k}": v / (global_batch_size if k.endswith("loss") else count) for k, v in reduced_metrics.items()}

            reduced_metrics["train/lr"] = lr_this_step
            return reduced_metrics

def evaluate(
    config: PretrainConfig,
    train_state: TrainState,
    eval_loader: torch.utils.data.DataLoader,
    eval_metadata: PuzzleDatasetMetadata,
    evaluators: List[Any],
    rank: int,
    world_size: int,
    cpu_group: Optional[dist.ProcessGroup],
):
    reduced_metrics = None

    with torch.inference_mode():
        return_keys = set(config.eval_save_outputs)
        for evaluator in evaluators:
            evaluator.begin_eval()
            return_keys.update(evaluator.required_outputs)

        # Run evaluation
        set_ids = {k: idx for idx, k in enumerate(eval_metadata.sets)}

        save_preds = {}

        metric_keys = []
        metric_values = None

        carry = None
        processed_batches = 0
        
        for set_name, batch, global_batch_size in eval_loader:
            processed_batches += 1
            if rank == 0:
                print(f"Processing batch {processed_batches}: {set_name}")
            
            # To device
            batch = {k: v.cuda() for k, v in batch.items()}
            with torch.device("cuda"):
                carry = train_state.model.initial_carry(batch)  # type: ignore

            # Forward
            inference_steps = 0
            while True:
                carry, loss, metrics, preds, all_finish = train_state.model(
                    carry=carry, batch=batch, return_keys=return_keys
                )
                inference_steps += 1

                if all_finish:
                    break

            if rank == 0:
                print(f"  Completed inference in {inference_steps} steps")

            for collection in (batch, preds):
                for k, v in collection.items():
                    if k in config.eval_save_outputs:
                        save_preds.setdefault(k, [])
                        save_preds[k].append(v.cpu())  # Move to CPU for saving GPU memory

            for evaluator in evaluators:
                evaluator.update_batch(batch, preds)

            del carry, loss, preds, batch, all_finish

            # Aggregate metrics
            set_id = set_ids[set_name]

            if metric_values is None:
                metric_keys = list(
                    sorted(metrics.keys())
                )  # Sort keys to guarantee all processes use the same order.
                metric_values = torch.zeros(
                    (len(set_ids), len(metrics.values())), dtype=torch.float32, device="cuda"
                )

            metric_values[set_id] += torch.stack([metrics[k] for k in metric_keys])

            del metrics

        # concatenate save preds
        save_preds = {k: torch.cat(v, dim=0) for k, v in save_preds.items()}

        # Save preds
        if config.checkpoint_path is not None and len(save_preds):
            # Each rank save predictions independently
            os.makedirs(os.path.dirname(config.checkpoint_path), exist_ok=True)
            torch.save(
                save_preds, os.path.join(config.checkpoint_path, f"step_{train_state.step}_all_preds.{rank}")
            )

        del save_preds

        # Reduce to rank 0
        if metric_values is not None:
            if world_size > 1:
                dist.reduce(metric_values, dst=0)

            if rank == 0:
                reduced_metrics = metric_values.cpu().numpy()
                reduced_metrics = {
                    set_name: {
                        metric_name: reduced_metrics[set_id, metric_id]
                        for metric_id, metric_name in enumerate(metric_keys)
                    }
                    for set_id, set_name in enumerate(set_ids)
                }

                # Postprocess
                for set_name, m in reduced_metrics.items():
                    count = m.pop("count")
                    reduced_metrics[set_name] = {k: v / count for k, v in m.items()}

        # Run evaluators
        if rank == 0:
            print(f"\nRunning {len(evaluators)} evaluator(s)...")
            
        for i, evaluator in enumerate(evaluators):
            if rank == 0:
                print(f"Running evaluator {i+1}/{len(evaluators)}: {evaluator.__class__.__name__}")
                
            # Path for saving
            evaluator_save_path = None
            if config.checkpoint_path is not None:
                evaluator_save_path = os.path.join(
                    config.checkpoint_path,
                    f"evaluator_{evaluator.__class__.__name__}_step_{train_state.step}",
                )
                os.makedirs(evaluator_save_path, exist_ok=True)

            # Run and log
            metrics = evaluator.result(evaluator_save_path, rank=rank, world_size=world_size, group=cpu_group)
            if rank == 0 and metrics is not None:
                if reduced_metrics is None:
                    reduced_metrics = {}

                reduced_metrics.update(metrics)
                print(f"  Completed {evaluator.__class__.__name__}")
                
        if rank == 0:
            print("All evaluators completed!")

    return reduced_metrics

def save_code_and_config(config: PretrainConfig):
    if config.checkpoint_path is None or wandb.run is None:
        return

    os.makedirs(config.checkpoint_path, exist_ok=True)

    # Copy code
    code_list = [
        get_model_source_path(config.arch.name),
        get_model_source_path(config.arch.loss.name)
    ]
    for code_file in code_list:
        if code_file is not None:
            code_name = os.path.basename(code_file)

            shutil.copy(code_file, os.path.join(config.checkpoint_path, code_name))

    # Dump config as yaml
    config_file = os.path.join(config.checkpoint_path, "all_config.yaml")
    with open(config_file, "wt") as f:
        yaml.dump(config.model_dump(), f)

    # Log code
    wandb.run.log_code(config.checkpoint_path)


def load_synced_config(hydra_config: DictConfig, rank: int, world_size: int) -> PretrainConfig:
    objects = [None]
    if rank == 0:
        config = PretrainConfig(**hydra_config)  # type: ignore

        # Naming
        if config.project_name is None:
            config.project_name = f"{os.path.basename(config.data_paths[0]).capitalize()}-ACT-torch"
        if config.run_name is None:
            config.run_name = f"{config.arch.name.split('@')[-1]} {coolname.generate_slug(2)}"
        if config.checkpoint_path is None:
            config.checkpoint_path = os.path.join("checkpoints", config.project_name, config.run_name)

        objects = [config]

    if world_size > 1:
        dist.broadcast_object_list(objects, src=0)

    return objects[0]  # type: ignore


@hydra.main(config_path="config", config_name="cfg_pretrain", version_base=None)
def launch(hydra_config: DictConfig):
    RANK = 0
    WORLD_SIZE = 1
    CPU_PROCESS_GROUP = None

    # Initialize distributed training if in distributed environment (e.g. torchrun)
    if "LOCAL_RANK" in os.environ:
        # Initialize distributed, default device and dtype
        dist.init_process_group(backend="nccl")

        RANK = dist.get_rank()
        WORLD_SIZE = dist.get_world_size()

        torch.cuda.set_device(int(os.environ["LOCAL_RANK"]))
        
        # CPU GLOO process group
        CPU_PROCESS_GROUP = dist.new_group(backend="gloo")
        assert (
            dist.get_rank(CPU_PROCESS_GROUP) == RANK and dist.get_world_size(CPU_PROCESS_GROUP) == WORLD_SIZE
        )

    # Load sync'ed config
    config = load_synced_config(hydra_config, rank=RANK, world_size=WORLD_SIZE)

    # Seed RNGs to ensure consistency
    torch.random.manual_seed(config.seed + RANK)

    # Dataset
    train_epochs_per_iter = config.eval_interval if config.eval_interval is not None else config.epochs
    total_iters = config.epochs // train_epochs_per_iter

    assert config.epochs % train_epochs_per_iter == 0, "Eval interval must be a divisor of total epochs."

    train_loader, train_metadata = create_dataloader(config, "train", test_set_mode=False, epochs_per_iter=train_epochs_per_iter, global_batch_size=config.global_batch_size, rank=RANK, world_size=WORLD_SIZE)
    try:
        eval_loader,  eval_metadata  = create_dataloader(config, "test", test_set_mode=True, epochs_per_iter=1, global_batch_size=config.global_batch_size, rank=RANK, world_size=WORLD_SIZE)
    except:
        print("NO EVAL DATA FOUND")
        eval_loader = eval_metadata = None

    try:
        evaluators = create_evaluators(config, eval_metadata)
    except:
        print("No evaluator found")
        evaluators = []

    # Train state
    train_state = init_train_state(config, train_metadata, rank=RANK, world_size=WORLD_SIZE)

    # Progress bar and logger
    progress_bar = None
    ema_helper = None
    if RANK == 0:
        progress_bar = tqdm.tqdm(total=train_state.total_steps)
        wandb.init(project=config.project_name, name=config.run_name, config=config.model_dump(), settings=wandb.Settings(_disable_stats=True))  # type: ignore
        wandb.log({"num_params": sum(x.numel() for x in train_state.model.parameters())}, step=0)
        save_code_and_config(config)
    if config.ema:
        print('Setup EMA')
        ema_helper = EMAHelper(mu=config.ema_rate)
        ema_helper.register(train_state.model)

    # Training Loop
    for _iter_id in range(total_iters):
        print (f"[Rank {RANK}, World Size {WORLD_SIZE}]: Epoch {_iter_id * train_epochs_per_iter}")

        ############ Train Iter
        if RANK == 0:
            print("TRAIN")
        train_state.model.train()
        for set_name, batch, global_batch_size in train_loader:
            metrics = train_batch(config, train_state, batch, global_batch_size, rank=RANK, world_size=WORLD_SIZE)

            if RANK == 0 and metrics is not None:
                wandb.log(metrics, step=train_state.step)
                progress_bar.update(train_state.step - progress_bar.n)  # type: ignore
            if config.ema:
                ema_helper.update(train_state.model)

        if _iter_id >= config.min_eval_interval:
            ############ Evaluation
            if RANK == 0:
                print("EVALUATE")
            if config.ema:
                print("SWITCH TO EMA")
                train_state_eval = copy.deepcopy(train_state)
                train_state_eval.model = ema_helper.ema_copy(train_state_eval.model)
            else:
                train_state_eval = train_state
            train_state_eval.model.eval()
            metrics = evaluate(config, 
                train_state_eval, 
                eval_loader, 
                eval_metadata, 
                evaluators,
                rank=RANK, 
                world_size=WORLD_SIZE,
                cpu_group=CPU_PROCESS_GROUP)

            if RANK == 0 and metrics is not None:
                wandb.log(metrics, step=train_state.step)
                
            ############ Checkpointing
            if RANK == 0:
                print("SAVE CHECKPOINT")
            if RANK == 0 and (config.checkpoint_every_eval or (_iter_id == total_iters - 1)):
                save_train_state(config, train_state_eval)

            if config.ema:
                del train_state_eval

    # finalize
    if dist.is_initialized():
        dist.destroy_process_group()
    wandb.finish()


if __name__ == "__main__":
    launch()