diff --git a/.gitignore b/.gitignore
index 6de8fd2..bbb955e 100644
--- a/.gitignore
+++ b/.gitignore
@@ -2,3 +2,4 @@ __pycache__
 /data
 /logs
 /ckpts
+/.cache
diff --git a/README.md b/README.md
index eb7d707..2a4f723 100644
--- a/README.md
+++ b/README.md
@@ -2,9 +2,12 @@
 
 An unofficial (toy) implementation of VALL-E, based on the [encodec](https://github.com/facebookresearch/encodec) tokenizer.
 
+[!["Buy Me A Coffee"](https://www.buymeacoffee.com/assets/img/custom_images/orange_img.png)](https://www.buymeacoffee.com/enhuiz)
+
 ## TODO
 
 - [x] AR model for the first quantizer.
 - [x] Audio decoding from tokens.
 - [x] NAR model for the rest quantizers.
-- [ ] Trainers for both models.
+- [x] Trainers for both models.
+- [ ] Pre-trained checkpoint.
diff --git a/config/ar.yml b/config/ar.yml
new file mode 100644
index 0000000..ad0266e
--- /dev/null
+++ b/config/ar.yml
@@ -0,0 +1,4 @@
+data_dirs: [data/test]
+
+model: ar
+batch_size: 1
diff --git a/config/nar.yml b/config/nar.yml
new file mode 100644
index 0000000..f5a78e7
--- /dev/null
+++ b/config/nar.yml
@@ -0,0 +1,4 @@
+data_dirs: [data/test]
+
+model: nar
+batch_size: 1
diff --git a/data/test/test.ar.recon.wav b/data/test/test.ar.recon.wav
index 3fd7f77..5ee5fe2 100644
Binary files a/data/test/test.ar.recon.wav and b/data/test/test.ar.recon.wav differ
diff --git a/data/test/test.nar.init.wav b/data/test/test.nar.init.wav
index 7709797..eb71667 100644
Binary files a/data/test/test.nar.init.wav and b/data/test/test.nar.init.wav differ
diff --git a/data/test/test.nar.recon.wav b/data/test/test.nar.recon.wav
index 7de0c4b..bbc09f0 100644
Binary files a/data/test/test.nar.recon.wav and b/data/test/test.nar.recon.wav differ
diff --git a/vall_e/config.py b/vall_e/config.py
new file mode 100644
index 0000000..bf3d6c0
--- /dev/null
+++ b/vall_e/config.py
@@ -0,0 +1,77 @@
+from dataclasses import dataclass, field
+from functools import cached_property
+from pathlib import Path
+
+import diskcache
+
+from .utils import Config as ConfigBase
+
+
+@dataclass(frozen=True)
+class Config(ConfigBase):
+    data_root: Path = Path("data")
+    data_dirs: list[Path] = field(default_factory=lambda: [])
+    test_data_dirs: list[Path] = field(default_factory=lambda: [])
+
+    batch_size: int = 24
+    eval_batch_size: int = 12
+    nj: int = 8
+
+    @property
+    def sample_rate(self):
+        return 24_000
+
+    p_additional_prompt: float = 0.5
+
+    token_dim: int = 256
+    num_tokens: int = 1024
+
+    batch_size: int = 128
+    eval_batch_size: int = 512
+    warmup_min_lr: float = 1e-6
+    warmup_max_lr: float = 2e-4
+    dis_warmup_max_lr: float = 4e-4
+    warmup_num_steps: int = 1_000
+    max_iter: int = 10_000
+    gradient_clipping: float = 100
+    eval_every: int = 2_000
+    save_ckpt_every: int = 10_000
+
+    model: str = "ar"
+    d_model: int = 512
+    n_heads: int = 8
+    n_layers: int = 12
+    p_dropout: float = 0.1
+
+    @property
+    def ds_cfg(self):
+        return {
+            "train_micro_batch_size_per_gpu": self.batch_size,
+            "gradient_accumulation_steps": 1,
+            "optimizer": {"type": "Adam"},
+            "scheduler": {
+                "type": "WarmupDecayLR",
+                "params": {
+                    "warmup_min_lr": self.warmup_min_lr,
+                    "warmup_max_lr": self.warmup_max_lr,
+                    "warmup_num_steps": self.warmup_num_steps,
+                    "total_num_steps": self.max_iter,
+                    "warmup_type": "linear",
+                },
+            },
+            "gradient_clipping": self.gradient_clipping,
+        }
+
+    @property
+    def cache_dir(self):
+        return ".cache" / self.relpath
+
+    @cached_property
+    def diskcache(self):
+        return diskcache.Cache(self.cache_dir).memoize
+
+
+cfg = Config.from_cli()
+
+if __name__ == "__main__":
+    print(cfg)
diff --git a/vall_e/data.py b/vall_e/data.py
new file mode 100644
index 0000000..8dde4a9
--- /dev/null
+++ b/vall_e/data.py
@@ -0,0 +1,294 @@
+import copy
+import logging
+import random
+from collections import defaultdict
+from functools import cache, cached_property
+from itertools import groupby, zip_longest
+from typing import Any
+
+import numpy as np
+import torch
+from torch import Tensor
+from torch.utils.data import DataLoader, Dataset
+from tqdm import tqdm
+
+from .config import cfg
+from .sampler import Sampler
+
+torch.multiprocessing.set_sharing_strategy("file_system")
+
+_logger = logging.getLogger(__name__)
+
+
+def _replace_file_extension(path, suffix):
+    return (path.parent / path.name.split(".")[0]).with_suffix(suffix)
+
+
+def _get_quant_path(path):
+    return _replace_file_extension(path, ".qnt.pt")
+
+
+def _load_quants(path) -> Tensor:
+    """
+    Returns:
+        quants: (t q)
+    """
+    path = _get_quant_path(path)
+    return torch.load(path)[0].t()
+
+
+@cache
+def _get_phones(path):
+    path = _replace_file_extension(path, ".phn.txt")
+    with open(path, "r", encoding="utf8") as f:
+        content = f.read()
+    return ["<s>"] + content.split() + ["</s>"]
+
+
+def _interleaved_reorder(l, fn):
+    groups = defaultdict(list)
+    for e in l:
+        groups[fn(e)].append(e)
+    groups = {k: groups[k] for k in sorted(groups)}
+    for interleaved in zip_longest(*groups.values()):
+        for value in interleaved:
+            if value is not None:
+                yield value
+
+
+@cache
+def _validate(path, min_phones, max_phones):
+    phones = _get_phones(path)
+    unique_phones = list(set(phones))
+    if len(unique_phones) == 0:
+        return False
+    if len(unique_phones) == 1 and unique_phones[0] == "_":
+        return False
+    if len(phones) < min_phones:
+        return False
+    if len(phones) > max_phones:
+        return False
+    return True
+
+
+def _get_spkr_name(path) -> str:
+    return path.parts[-2]  # spkr/*.wav
+
+
+class VALLEDatset(Dataset):
+    def __init__(
+        self,
+        paths,
+        phone_symmap=None,
+        spkr_symmap=None,
+        min_phones=10,
+        max_phones=100,
+        training=False,
+        extra_paths_by_spkr_name: dict[str, list] = {},
+    ):
+        super().__init__()
+        self._head = None
+        self.min_phones = min_phones
+        self.max_phones = max_phones
+        self.paths = [
+            path for path in paths if _validate(path, self.min_phones, self.max_phones)
+        ]
+        self.spkr_symmap = spkr_symmap or self._get_spkr_symmap()
+        self.phone_symmap = phone_symmap or self._get_phone_symmap()
+        self.training = training
+
+        self.paths_by_spkr_name = self._get_paths_by_spkr_name(extra_paths_by_spkr_name)
+
+        if training:
+            self.sampler = Sampler(self.paths, [_get_spkr_name])
+        else:
+            self.sampler = None
+
+    def _get_paths_by_spkr_name(self, extra_paths_by_spkr_name: dict[str, list]):
+        ret = defaultdict(list)
+        for path in self.paths:
+            if _get_quant_path(path).exists():
+                ret[_get_spkr_name(path)].append(path)
+        for k, v in extra_paths_by_spkr_name.items():
+            ret[k].extend(v)
+        return {**ret}
+
+    @cached_property
+    def phones(self):
+        return sorted(set().union(*[_get_phones(path) for path in self.paths]))
+
+    def _get_phone_symmap(self):
+        # Note that we use phone symmap starting from 1 so that we can safely pad 0.
+        return {s: i for i, s in enumerate(self.phones, 1)}
+
+    @cached_property
+    def spkrs(self):
+        return sorted({_get_spkr_name(path) for path in self.paths})
+
+    def _get_spkr_symmap(self):
+        return {s: i for i, s in enumerate(self.spkrs)}
+
+    def sample_prompts(self, spkr_name):
+        prom_list = []
+
+        while (
+            len(prom_list) == 0
+            or random.random() < cfg.p_additional_prompt
+            and len(prom_list) < 10
+        ):
+            path = random.choice(self.paths_by_spkr_name[spkr_name])
+            prom_list.append(_load_quants(path))
+
+        prom = torch.cat(prom_list)
+
+        return prom
+
+    def __getitem__(self, index):
+        if self.training:
+            assert self.sampler is not None
+            path = self.sampler.sample()
+        else:
+            path = self.paths[index]
+
+        spkr_name = _get_spkr_name(path)
+        text = torch.tensor([*map(self.phone_symmap.get, _get_phones(path))])
+        proms = self.sample_prompts(spkr_name)
+        resps = _load_quants(path)
+        resp = resps[..., 0]
+
+        return dict(
+            path=path,
+            spkr_name=spkr_name,
+            text=text,
+            proms=proms,
+            resps=resps,
+            resp=resp,
+        )
+
+    def head_(self, n):
+        self._head = n
+
+    def training_(self, value):
+        self.training = value
+
+    def interleaved_reorder_(self, fn):
+        self.paths = [*_interleaved_reorder(self.paths, fn)]
+
+    def __len__(self):
+        return min(len(self.paths), self._head or len(self.paths))
+
+
+def collate_fn(samples: list[dict]):
+    batch: dict[str, Any] = {k: [s[k] for s in samples] for k in samples[0]}
+    return batch
+
+
+def _seed_worker(worker_id):
+    worker_seed = torch.initial_seed() % 2**32
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
+
+
+def _create_dl(dataset, training):
+    return DataLoader(
+        dataset=dataset,
+        batch_size=cfg.batch_size if training else cfg.eval_batch_size,
+        shuffle=training,
+        drop_last=training,
+        num_workers=cfg.nj,
+        collate_fn=collate_fn,
+        persistent_workers=True,
+        worker_init_fn=_seed_worker,
+    )
+
+
+def _load_train_val_paths():
+    paths = []
+    train_paths = []
+    val_paths = []
+
+    for data_dir in cfg.data_dirs:
+        paths.extend(tqdm(data_dir.rglob("**/*.qnt.pt")))
+
+    if len(paths) == 0:
+        raise RuntimeError(f"Failed to find any .qnt.pt file in {cfg.data_dirs}.")
+
+    pairs = sorted([(_get_spkr_name(p), p) for p in paths])
+    del paths
+
+    for _, group in groupby(pairs, lambda pair: pair[0]):
+        paths = sorted([p for _, p in group])
+        random.seed(0)
+        random.shuffle(paths)
+        n = round(len(paths) * 0.95)
+        train_paths.extend(paths[:n])
+        val_paths.extend(paths[n:])
+
+    train_paths, val_paths = map(sorted, [train_paths, val_paths])
+
+    return train_paths, val_paths
+
+
+def _load_test_paths():
+    test_paths = []
+    for data_dir in cfg.test_data_dirs:
+        test_paths.extend(data_dir.rglob("**/*.asr.txt"))
+    test_paths = sorted(test_paths)
+    return test_paths
+
+
+@cfg.diskcache()
+def create_datasets():
+    train_paths, val_paths = _load_train_val_paths()
+    test_paths = _load_test_paths()
+
+    train_dataset = VALLEDatset(train_paths, training=True)
+
+    val_dataset = VALLEDatset(
+        val_paths,
+        train_dataset.phone_symmap,
+        train_dataset.spkr_symmap,
+        extra_paths_by_spkr_name=train_dataset.paths_by_spkr_name,
+    )
+
+    val_dataset.interleaved_reorder_(_get_spkr_name)
+    val_dataset.head_(200)
+
+    test_dataset = VALLEDatset(
+        test_paths,
+        train_dataset.phone_symmap,
+        train_dataset.spkr_symmap,
+        extra_paths_by_spkr_name=train_dataset.paths_by_spkr_name,
+    )
+
+    return train_dataset, val_dataset, test_dataset
+
+
+def create_train_val_dataloader():
+    train_dataset, val_dataset, test_dataset = create_datasets()
+
+    train_dl = _create_dl(train_dataset, training=True)
+    val_dl = _create_dl(val_dataset, training=False)
+    test_dl = _create_dl(test_dataset, training=False)
+
+    _logger.info(str(train_dataset.phone_symmap))
+    _logger.info(str(train_dataset.spkr_symmap))
+
+    _logger.info(f"#samples (train): {len(train_dataset)}.")
+    _logger.info(f"#samples (val): {len(val_dataset)}.")
+    _logger.info(f"#samples (test): {len(test_dataset)}.")
+
+    train200_dataset = copy.deepcopy(train_dataset)
+    train200_dataset.interleaved_reorder_(_get_spkr_name)
+    train200_dataset.head_(200)
+    train200_dataset.training_(False)
+    train200_dl = _create_dl(train200_dataset, training=False)
+    assert isinstance(train200_dl.dataset, VALLEDatset)
+
+    return train_dl, train200_dl, val_dl, test_dl
+
+
+if __name__ == "__main__":
+    train_dl, train200_dl, val_dl, test_dl = create_train_val_dataloader()
+    sample = train_dl.dataset[0]
+    print(sample)
diff --git a/vall_e/emb/g2p.py b/vall_e/emb/g2p.py
new file mode 100644
index 0000000..c98758a
--- /dev/null
+++ b/vall_e/emb/g2p.py
@@ -0,0 +1,50 @@
+import argparse
+import random
+import string
+from functools import cache
+from pathlib import Path
+
+import torch
+from g2p_en import G2p
+from tqdm import tqdm
+
+
+@cache
+def _get_model():
+    return G2p()
+
+
+@cache
+def _get_graphs(path):
+    with open(path, "r") as f:
+        graphs = f.read()
+    return graphs
+
+
+def encode(graphs: str) -> list[str]:
+    g2p = _get_model()
+    phones = g2p(graphs)
+    ignored = {" ", *string.punctuation}
+    return ["_" if p in ignored else p for p in phones]
+
+
+@torch.no_grad()
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("folder", type=Path)
+    parser.add_argument("--suffix", type=str, default=".normalized.txt")
+    args = parser.parse_args()
+
+    paths = list(args.folder.rglob(f"*{args.suffix}"))
+    random.shuffle(paths)
+
+    for path in tqdm(paths):
+        phone_path = path.with_name(path.stem.split(".")[0] + ".phn.txt")
+        graphs = _get_graphs(path)
+        phones = encode(graphs)
+        with open(phone_path, "w") as f:
+            f.write(" ".join(phones))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/vall_e/emb/qnt.py b/vall_e/emb/qnt.py
index 1cba87e..b70951a 100644
--- a/vall_e/emb/qnt.py
+++ b/vall_e/emb/qnt.py
@@ -2,35 +2,51 @@ import argparse
 from functools import cache
 from pathlib import Path
 
+import soundfile
 import torch
 import torchaudio
+from einops import rearrange
 from encodec import EncodecModel
 from encodec.utils import convert_audio
 from torch import Tensor
 from tqdm import tqdm
 
+from ..config import cfg
+
 
 @cache
 def _load_model(device="cuda"):
     # Instantiate a pretrained EnCodec model
+    assert cfg.sample_rate == 24_000
     model = EncodecModel.encodec_model_24khz()
     model.set_target_bandwidth(6.0)
     model.to(device)
     return model
 
 
+def unload_model():
+    return _load_model.cache_clear()
+
+
 @torch.inference_mode()
 def decode(codes: Tensor, device="cuda"):
     """
     Args:
-        codes: (b k t)
+        codes: (b q t)
     """
     assert codes.dim() == 3
     model = _load_model(device)
     return model.decode([(codes, None)]), model.sample_rate
 
 
-def replace_file_extension(path, suffix):
+def decode_to_file(resps: Tensor, path: Path):
+    assert resps.dim() == 2, f"Require shape (t q), but got {resps.shape}."
+    resps = rearrange(resps, "t q -> 1 q t")
+    wavs, sr = decode(resps)
+    soundfile.write(str(path), wavs.cpu()[0, 0], sr)
+
+
+def _replace_file_extension(path, suffix):
     return (path.parent / path.name.split(".")[0]).with_suffix(suffix)
 
 
@@ -46,7 +62,7 @@ def encode(wav, sr, device="cuda"):
     wav = convert_audio(wav, sr, model.sample_rate, model.channels)
     wav = wav.to(device)
     encoded_frames = model.encode(wav)
-    qnt = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1)  # (b k t)
+    qnt = torch.cat([encoded[0] for encoded in encoded_frames], dim=-1)  # (b q t)
     return qnt
 
 
@@ -59,7 +75,7 @@ def main():
     paths = [*args.folder.rglob(f"*{args.suffix}")]
 
     for path in tqdm(paths):
-        out_path = replace_file_extension(path, ".qnt.pt")
+        out_path = _replace_file_extension(path, ".qnt.pt")
         wav, sr = torchaudio.load(path)
         if wav.shape[0] == 2:
             wav = wav[:1]
diff --git a/vall_e/sampler.py b/vall_e/sampler.py
new file mode 100644
index 0000000..b979ebb
--- /dev/null
+++ b/vall_e/sampler.py
@@ -0,0 +1,48 @@
+"""
+A sampler that balances data by key_fns.
+
+MIT License
+
+Copyright (c) 2023 Zhe Niu
+
+niuzhe.nz@outlook.com
+"""
+
+import random
+
+
+class Sampler:
+    def __init__(self, l, key_fns):
+        self.tree = self._build(l, key_fns)
+
+    def _build(self, l, key_fns) -> dict[dict, list]:
+        if not key_fns:
+            return l
+
+        tree = {}
+
+        key_fn, *key_fns = key_fns
+
+        for x in l:
+            k = key_fn(x)
+
+            if k in tree:
+                tree[k].append(x)
+            else:
+                tree[k] = [x]
+
+        for k in tree:
+            tree[k] = self._build(tree[k], key_fns)
+
+        return tree
+
+    def _sample(self, tree: dict | list):
+        if isinstance(tree, list):
+            ret = random.choice(tree)
+        else:
+            key = random.choice([*tree.keys()])
+            ret = self._sample(tree[key])
+        return ret
+
+    def sample(self):
+        return self._sample(self.tree)
diff --git a/vall_e/train.py b/vall_e/train.py
new file mode 100644
index 0000000..a5a721f
--- /dev/null
+++ b/vall_e/train.py
@@ -0,0 +1,137 @@
+import json
+import logging
+from collections import defaultdict
+
+import torch
+from tqdm import tqdm
+
+from .config import cfg
+from .data import create_train_val_dataloader
+from .emb import qnt
+from .utils import setup_logging, to_device, trainer
+from .vall_e import AR, NAR
+
+_logger = logging.getLogger(__name__)
+
+
+def load_engines():
+    if cfg.model.lower() == "ar":
+        model = AR(
+            cfg.num_tokens,
+            cfg.d_model,
+            cfg.n_heads,
+            cfg.n_layers,
+            cfg.p_dropout,
+        )
+    elif cfg.model.lower() == "nar":
+        model = NAR(
+            cfg.num_tokens,
+            cfg.d_model,
+            cfg.n_heads,
+            cfg.n_layers,
+            cfg.p_dropout,
+        )
+    else:
+        raise NotImplementedError(cfg.model)
+
+    engines = dict(
+        model=trainer.Engine(
+            model=model,
+            config=cfg.ds_cfg,
+        ),
+    )
+
+    return trainer.load_engines(engines, cfg)
+
+
+def main():
+    setup_logging(cfg.log_dir)
+
+    train_dl, train200_dl, val_dl, test_dl = create_train_val_dataloader()
+
+    def train_feeder(engines, batch, name):
+        model = engines["model"]
+
+        if cfg.model == "ar":
+            _ = model(
+                text_list=batch["text"],
+                proms_list=batch["proms"],
+                resp_list=batch["resp"],
+            )
+        elif cfg.model == "nar":
+            _ = model(
+                text_list=batch["text"],
+                proms_list=batch["proms"],
+                resps_list=batch["resps"],
+            )
+
+        losses = model.gather_attribute("loss")
+
+        loss = torch.stack([*losses.values()]).sum()
+
+        stats = {}
+        stats |= {k: v.item() for k, v in losses.items()}
+        stats |= engines.gather_attribute("scalar")
+
+        return loss, stats
+
+    @torch.inference_mode()
+    def run_eval(engines, name, dl):
+        log_dir = cfg.log_dir / str(engines.global_step) / name
+
+        model = engines["model"]
+        log_dir = cfg.log_dir / str(engines.global_step) / name
+        stats = defaultdict(list)
+        for batch in tqdm(dl):
+            batch: dict
+            batch = to_device(batch, cfg.device)
+
+            if cfg.model == "ar":
+                resp_list = model(text_list=batch["text"], proms_list=batch["proms"])
+                resps_list = [r.unsqueeze(-1) for r in resp_list]
+            elif cfg.model == "nar":
+                resps_list = model(
+                    text_list=batch["text"],
+                    proms_list=batch["proms"],
+                    resp_list=batch["resp"],
+                )
+            else:
+                raise NotImplementedError(cfg.model)
+
+            losses = model.gather_attribute("loss")
+            batch_stats = {k: v.item() for k, v in losses.items()}
+            for k, v in batch_stats.items():
+                stats[k].append(v)
+
+            for path, ref, hyp in zip(batch["path"], batch["resps"], resps_list):
+                relpath = path.relative_to(cfg.data_root)
+                hyp_path = (log_dir / "hyp" / relpath).with_suffix(".wav")
+                ref_path = (log_dir / "ref" / relpath).with_suffix(".wav")
+                hyp_path.parent.mkdir(parents=True, exist_ok=True)
+                ref_path.parent.mkdir(parents=True, exist_ok=True)
+                qnt.decode_to_file(ref, ref_path)
+                if len(hyp) > 0:
+                    qnt.decode_to_file(hyp, hyp_path)
+
+        stats = {k: sum(v) / len(v) for k, v in stats.items()}
+        stats["global_step"] = engines.global_step
+        stats["name"] = name
+        _logger.info(f"Eval: {stats}.")
+
+        _logger.info(f"{json.dumps(stats)}.")
+
+    def eval_fn(engines):
+        run_eval(engines, "train200", train200_dl)
+        run_eval(engines, "val", val_dl)
+        run_eval(engines, "test", test_dl)
+
+    trainer.train(
+        engines_loader=load_engines,
+        train_dl=train_dl,
+        train_feeder=train_feeder,
+        eval_fn=eval_fn,
+    )
+
+
+if __name__ == "__main__":
+    main()