src/train.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "torch>=2.0.0",
#     "transformers>=4.30.0",
#     "datasets>=2.14.0",
#     "click>=8.0.0",
#     "tqdm>=4.60.0",
#     "wandb>=0.15.0",
#     "python-dotenv>=1.0.0",
# ]
# ///
"""
Training script for Bamboo-1 Vietnamese Dependency Parser.

Supports multiple methods:
- baseline: BiLSTM + Biaffine (Dozat & Manning, 2017)
- trankit: XLM-RoBERTa + Biaffine (Nguyen et al., 2021)

Usage:
    uv run scripts/train.py                           # Default baseline
    uv run scripts/train.py --method trankit          # Reproduce Trankit
    uv run scripts/train.py --method trankit --dataset ud-vtb  # Trankit on VTB
"""

import sys
from pathlib import Path
from collections import Counter
from dataclasses import dataclass
from typing import List, Tuple, Optional

# Load environment variables
from dotenv import load_dotenv
load_dotenv()

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence, pad_sequence
from torch.utils.data import Dataset, DataLoader
from torch.optim import Adam, AdamW
from torch.optim.lr_scheduler import ExponentialLR
from tqdm import tqdm

import click

sys.path.insert(0, str(Path(__file__).parent.parent))
from src.corpus import UDD1Corpus
from src.ud_corpus import UDVietnameseVTB
from src.vndt_corpus import VnDTCorpus
from src.cost_estimate import CostTracker, detect_hardware


# ============================================================================
# Data Processing
# ============================================================================

@dataclass
class Sentence:
    """A dependency-parsed sentence."""
    words: List[str]
    heads: List[int]
    rels: List[str]


def read_conllu(path: str) -> List[Sentence]:
    """Read CoNLL-U file and return list of sentences."""
    sentences = []
    words, heads, rels = [], [], []

    with open(path, 'r', encoding='utf-8') as f:
        for line in f:
            line = line.strip()
            if not line:
                if words:
                    sentences.append(Sentence(words, heads, rels))
                    words, heads, rels = [], [], []
            elif line.startswith('#'):
                continue
            else:
                parts = line.split('\t')
                if '-' in parts[0] or '.' in parts[0]:  # Skip multi-word tokens
                    continue
                words.append(parts[1])  # FORM
                heads.append(int(parts[6]))  # HEAD
                rels.append(parts[7])  # DEPREL

        if words:
            sentences.append(Sentence(words, heads, rels))

    return sentences


class Vocabulary:
    """Vocabulary for words, characters, and relations."""
    PAD = '<pad>'
    UNK = '<unk>'

    def __init__(self, min_freq: int = 2):
        self.min_freq = min_freq
        self.word2idx = {self.PAD: 0, self.UNK: 1}
        self.char2idx = {self.PAD: 0, self.UNK: 1}
        self.rel2idx = {}
        self.idx2rel = {}

    def build(self, sentences: List[Sentence]):
        """Build vocabulary from sentences."""
        word_counts = Counter()
        char_counts = Counter()
        rel_counts = Counter()

        for sent in sentences:
            for word in sent.words:
                word_counts[word.lower()] += 1
                for char in word:
                    char_counts[char] += 1
            for rel in sent.rels:
                rel_counts[rel] += 1

        # Words
        for word, count in word_counts.items():
            if count >= self.min_freq and word not in self.word2idx:
                self.word2idx[word] = len(self.word2idx)

        # Characters
        for char, count in char_counts.items():
            if char not in self.char2idx:
                self.char2idx[char] = len(self.char2idx)

        # Relations
        for rel in rel_counts:
            if rel not in self.rel2idx:
                idx = len(self.rel2idx)
                self.rel2idx[rel] = idx
                self.idx2rel[idx] = rel

    def encode_word(self, word: str) -> int:
        return self.word2idx.get(word.lower(), self.word2idx[self.UNK])

    def encode_char(self, char: str) -> int:
        return self.char2idx.get(char, self.char2idx[self.UNK])

    def encode_rel(self, rel: str) -> int:
        return self.rel2idx.get(rel, 0)

    @property
    def n_words(self) -> int:
        return len(self.word2idx)

    @property
    def n_chars(self) -> int:
        return len(self.char2idx)

    @property
    def n_rels(self) -> int:
        return len(self.rel2idx)


class DependencyDataset(Dataset):
    """Dataset for dependency parsing."""

    def __init__(self, sentences: List[Sentence], vocab: Vocabulary):
        self.sentences = sentences
        self.vocab = vocab

    def __len__(self):
        return len(self.sentences)

    def __getitem__(self, idx):
        sent = self.sentences[idx]

        # Encode words
        word_ids = [self.vocab.encode_word(w) for w in sent.words]

        # Encode characters
        char_ids = [[self.vocab.encode_char(c) for c in w] for w in sent.words]

        # Heads and relations
        heads = sent.heads
        rels = [self.vocab.encode_rel(r) for r in sent.rels]

        return word_ids, char_ids, heads, rels


def collate_fn(batch):
    """Collate function for DataLoader."""
    word_ids, char_ids, heads, rels = zip(*batch)

    # Get lengths
    lengths = [len(w) for w in word_ids]
    max_len = max(lengths)

    # Pad words
    word_ids_padded = torch.zeros(len(batch), max_len, dtype=torch.long)
    for i, wids in enumerate(word_ids):
        word_ids_padded[i, :len(wids)] = torch.tensor(wids)

    # Pad characters
    max_word_len = max(max(len(c) for c in chars) for chars in char_ids)
    char_ids_padded = torch.zeros(len(batch), max_len, max_word_len, dtype=torch.long)
    for i, chars in enumerate(char_ids):
        for j, c in enumerate(chars):
            char_ids_padded[i, j, :len(c)] = torch.tensor(c)

    # Pad heads
    heads_padded = torch.zeros(len(batch), max_len, dtype=torch.long)
    for i, h in enumerate(heads):
        heads_padded[i, :len(h)] = torch.tensor(h)

    # Pad rels
    rels_padded = torch.zeros(len(batch), max_len, dtype=torch.long)
    for i, r in enumerate(rels):
        rels_padded[i, :len(r)] = torch.tensor(r)

    # Mask
    mask = torch.zeros(len(batch), max_len, dtype=torch.bool)
    for i, l in enumerate(lengths):
        mask[i, :l] = True

    lengths = torch.tensor(lengths)

    return word_ids_padded, char_ids_padded, heads_padded, rels_padded, mask, lengths


# ============================================================================
# Model
# ============================================================================

class CharLSTM(nn.Module):
    """Character-level LSTM embeddings."""

    def __init__(self, n_chars: int, char_dim: int = 50, hidden_dim: int = 100):
        super().__init__()
        self.embed = nn.Embedding(n_chars, char_dim, padding_idx=0)
        self.lstm = nn.LSTM(char_dim, hidden_dim // 2, batch_first=True, bidirectional=True)
        self.hidden_dim = hidden_dim

    def forward(self, chars):
        """
        Args:
            chars: (batch, seq_len, max_word_len)
        Returns:
            (batch, seq_len, hidden_dim)
        """
        batch, seq_len, max_word_len = chars.shape

        # Flatten
        chars_flat = chars.view(-1, max_word_len)  # (batch * seq_len, max_word_len)

        # Get word lengths
        word_lens = (chars_flat != 0).sum(dim=1)
        word_lens = word_lens.clamp(min=1)

        # Embed
        char_embeds = self.embed(chars_flat)  # (batch * seq_len, max_word_len, char_dim)

        # Pack and run LSTM
        packed = pack_padded_sequence(char_embeds, word_lens.cpu(), batch_first=True, enforce_sorted=False)
        _, (hidden, _) = self.lstm(packed)

        # Concatenate forward and backward hidden states
        hidden = torch.cat([hidden[0], hidden[1]], dim=-1)  # (batch * seq_len, hidden_dim)

        return hidden.view(batch, seq_len, self.hidden_dim)


class MLP(nn.Module):
    """Multi-layer perceptron."""

    def __init__(self, input_dim: int, hidden_dim: int, dropout: float = 0.33):
        super().__init__()
        self.linear = nn.Linear(input_dim, hidden_dim)
        self.activation = nn.LeakyReLU(0.1)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        return self.dropout(self.activation(self.linear(x)))


class Biaffine(nn.Module):
    """Biaffine attention layer."""

    def __init__(self, input_dim: int, output_dim: int = 1, bias_x: bool = True, bias_y: bool = True):
        super().__init__()
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.bias_x = bias_x
        self.bias_y = bias_y

        self.weight = nn.Parameter(torch.zeros(output_dim, input_dim + bias_x, input_dim + bias_y))
        nn.init.xavier_uniform_(self.weight)

    def forward(self, x, y):
        """
        Args:
            x: (batch, seq_len, input_dim) - dependent
            y: (batch, seq_len, input_dim) - head
        Returns:
            (batch, seq_len, seq_len, output_dim) or (batch, seq_len, seq_len) if output_dim=1
        """
        if self.bias_x:
            x = torch.cat([x, torch.ones_like(x[..., :1])], dim=-1)
        if self.bias_y:
            y = torch.cat([y, torch.ones_like(y[..., :1])], dim=-1)

        # (batch, seq_len, output_dim, input_dim+1)
        x = torch.einsum('bxi,oij->bxoj', x, self.weight)
        # (batch, seq_len, seq_len, output_dim)
        scores = torch.einsum('bxoj,byj->bxyo', x, y)

        if self.output_dim == 1:
            scores = scores.squeeze(-1)

        return scores


class BiaffineDependencyParser(nn.Module):
    """Biaffine Dependency Parser (Dozat & Manning, 2017)."""

    def __init__(
        self,
        n_words: int,
        n_chars: int,
        n_rels: int,
        word_dim: int = 100,
        char_dim: int = 50,
        char_hidden: int = 100,
        lstm_hidden: int = 400,
        lstm_layers: int = 3,
        arc_hidden: int = 500,
        rel_hidden: int = 100,
        dropout: float = 0.33,
    ):
        super().__init__()

        self.word_embed = nn.Embedding(n_words, word_dim, padding_idx=0)
        self.char_lstm = CharLSTM(n_chars, char_dim, char_hidden)

        input_dim = word_dim + char_hidden

        self.lstm = nn.LSTM(
            input_dim, lstm_hidden // 2,
            num_layers=lstm_layers,
            batch_first=True,
            bidirectional=True,
            dropout=dropout if lstm_layers > 1 else 0
        )

        self.mlp_arc_dep = MLP(lstm_hidden, arc_hidden, dropout)
        self.mlp_arc_head = MLP(lstm_hidden, arc_hidden, dropout)
        self.mlp_rel_dep = MLP(lstm_hidden, rel_hidden, dropout)
        self.mlp_rel_head = MLP(lstm_hidden, rel_hidden, dropout)

        self.arc_attn = Biaffine(arc_hidden, 1, bias_x=True, bias_y=False)
        self.rel_attn = Biaffine(rel_hidden, n_rels, bias_x=True, bias_y=True)

        self.dropout = nn.Dropout(dropout)
        self.n_rels = n_rels

    def forward(self, words, chars, mask):
        """
        Args:
            words: (batch, seq_len)
            chars: (batch, seq_len, max_word_len)
            mask: (batch, seq_len)
        Returns:
            arc_scores: (batch, seq_len, seq_len)
            rel_scores: (batch, seq_len, seq_len, n_rels)
        """
        # Embeddings
        word_embeds = self.word_embed(words)
        char_embeds = self.char_lstm(chars)
        embeds = torch.cat([word_embeds, char_embeds], dim=-1)
        embeds = self.dropout(embeds)

        # BiLSTM
        lengths = mask.sum(dim=1).cpu()
        packed = pack_padded_sequence(embeds, lengths, batch_first=True, enforce_sorted=False)
        lstm_out, _ = self.lstm(packed)
        lstm_out, _ = pad_packed_sequence(lstm_out, batch_first=True, total_length=mask.size(1))
        lstm_out = self.dropout(lstm_out)

        # MLP
        arc_dep = self.mlp_arc_dep(lstm_out)
        arc_head = self.mlp_arc_head(lstm_out)
        rel_dep = self.mlp_rel_dep(lstm_out)
        rel_head = self.mlp_rel_head(lstm_out)

        # Biaffine
        arc_scores = self.arc_attn(arc_dep, arc_head)  # (batch, seq_len, seq_len)
        rel_scores = self.rel_attn(rel_dep, rel_head)  # (batch, seq_len, seq_len, n_rels)

        return arc_scores, rel_scores

    def loss(self, arc_scores, rel_scores, heads, rels, mask):
        """Compute loss."""
        batch_size, seq_len = mask.shape

        # Arc loss
        arc_scores = arc_scores.masked_fill(~mask.unsqueeze(2), float('-inf'))
        arc_loss = F.cross_entropy(
            arc_scores[mask].view(-1, seq_len),
            heads[mask],
            reduction='mean'
        )

        # Rel loss - select scores for gold heads
        rel_scores_gold = rel_scores[torch.arange(batch_size).unsqueeze(1), torch.arange(seq_len), heads]
        rel_loss = F.cross_entropy(
            rel_scores_gold[mask],
            rels[mask],
            reduction='mean'
        )

        return arc_loss + rel_loss

    def decode(self, arc_scores, rel_scores, mask):
        """Decode predictions."""
        # Greedy decoding
        arc_preds = arc_scores.argmax(dim=-1)

        batch_size, seq_len = mask.shape
        rel_scores_pred = rel_scores[torch.arange(batch_size).unsqueeze(1), torch.arange(seq_len), arc_preds]
        rel_preds = rel_scores_pred.argmax(dim=-1)

        return arc_preds, rel_preds


# ============================================================================
# Trankit-style Transformer Parser (XLM-RoBERTa + Biaffine)
# ============================================================================

class TransformerDependencyParser(nn.Module):
    """
    Trankit-style dependency parser using XLM-RoBERTa.

    Architecture follows Nguyen et al. 2021 EACL:
    - XLM-RoBERTa encoder
    - Word-level pooling (first subword)
    - Biaffine attention for arc/rel prediction
    """

    def __init__(
        self,
        n_rels: int,
        encoder: str = "xlm-roberta-base",
        arc_hidden: int = 500,
        rel_hidden: int = 100,
        dropout: float = 0.33,
    ):
        super().__init__()
        from transformers import AutoModel, AutoTokenizer

        self.encoder_name = encoder
        self.tokenizer = AutoTokenizer.from_pretrained(encoder)
        self.encoder = AutoModel.from_pretrained(encoder)
        self.hidden_size = self.encoder.config.hidden_size

        # Biaffine layers
        self.mlp_arc_dep = MLP(self.hidden_size, arc_hidden, dropout)
        self.mlp_arc_head = MLP(self.hidden_size, arc_hidden, dropout)
        self.mlp_rel_dep = MLP(self.hidden_size, rel_hidden, dropout)
        self.mlp_rel_head = MLP(self.hidden_size, rel_hidden, dropout)

        self.arc_attn = Biaffine(arc_hidden, 1, bias_x=True, bias_y=False)
        self.rel_attn = Biaffine(rel_hidden, n_rels, bias_x=True, bias_y=True)

        self.dropout = nn.Dropout(dropout)
        self.n_rels = n_rels

    def encode_pretokenized(self, input_ids, attention_mask, word_starts, word_mask):
        """Encode pre-tokenized batch (fast path - no tokenization overhead)."""
        outputs = self.encoder(input_ids, attention_mask=attention_mask)
        hidden = outputs.last_hidden_state

        # Gather word-level representations using pre-computed positions
        word_starts_exp = word_starts.unsqueeze(-1).expand(-1, -1, self.hidden_size)
        word_starts_exp = word_starts_exp.clamp(0, hidden.size(1) - 1)
        word_hidden = torch.gather(hidden, 1, word_starts_exp)

        return word_hidden, word_mask

    def forward(self, word_hidden, word_mask):
        """Compute arc and relation scores from word representations."""
        word_hidden = self.dropout(word_hidden)

        # Biaffine scoring
        arc_dep = self.mlp_arc_dep(word_hidden)
        arc_head = self.mlp_arc_head(word_hidden)
        rel_dep = self.mlp_rel_dep(word_hidden)
        rel_head = self.mlp_rel_head(word_hidden)

        arc_scores = self.arc_attn(arc_dep, arc_head)
        rel_scores = self.rel_attn(rel_dep, rel_head)

        return arc_scores, rel_scores

    def loss(self, arc_scores, rel_scores, heads, rels, mask):
        """Compute cross-entropy loss."""
        batch_size, seq_len = mask.shape

        # Arc loss
        arc_scores = arc_scores.masked_fill(~mask.unsqueeze(2), float('-inf'))
        arc_loss = F.cross_entropy(
            arc_scores[mask].view(-1, seq_len),
            heads[mask],
            reduction='mean'
        )

        # Rel loss
        rel_scores_gold = rel_scores[torch.arange(batch_size, device=mask.device).unsqueeze(1),
                                      torch.arange(seq_len, device=mask.device), heads]
        rel_loss = F.cross_entropy(
            rel_scores_gold[mask],
            rels[mask],
            reduction='mean'
        )

        return arc_loss + rel_loss

    def decode(self, arc_scores, rel_scores, mask):
        """Greedy decoding."""
        arc_preds = arc_scores.argmax(dim=-1)

        batch_size, seq_len = mask.shape
        rel_scores_pred = rel_scores[torch.arange(batch_size, device=mask.device).unsqueeze(1),
                                      torch.arange(seq_len, device=mask.device), arc_preds]
        rel_preds = rel_scores_pred.argmax(dim=-1)

        return arc_preds, rel_preds


class TransformerDataset(Dataset):
    """Pre-tokenized dataset - tokenizes once at creation for fast training."""

    def __init__(self, sentences: List[Sentence], vocab, tokenizer):
        self.data = []
        for sent in tqdm(sentences, desc="Pre-tokenizing", leave=False):
            input_ids = [tokenizer.cls_token_id]
            word_starts = []
            for word in sent.words:
                word_starts.append(len(input_ids))
                tokens = tokenizer.encode(word, add_special_tokens=False)
                input_ids.extend(tokens if tokens else [tokenizer.unk_token_id])
            input_ids.append(tokenizer.sep_token_id)
            self.data.append((
                torch.tensor(input_ids, dtype=torch.long),
                torch.tensor(word_starts, dtype=torch.long),
                torch.tensor(sent.heads, dtype=torch.long),
                torch.tensor([vocab.encode_rel(r) for r in sent.rels], dtype=torch.long),
                len(sent.words),
            ))

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        return self.data[idx]


def transformer_collate_fn(batch):
    """Collate for pre-tokenized transformer data."""
    input_ids, word_starts, heads, rels, n_words = zip(*batch)

    batch_size = len(batch)
    max_subwords = max(ids.size(0) for ids in input_ids)
    max_words = max(n_words)

    padded_ids = torch.zeros(batch_size, max_subwords, dtype=torch.long)
    attention_mask = torch.zeros(batch_size, max_subwords, dtype=torch.long)
    padded_starts = torch.zeros(batch_size, max_words, dtype=torch.long)
    padded_heads = torch.zeros(batch_size, max_words, dtype=torch.long)
    padded_rels = torch.zeros(batch_size, max_words, dtype=torch.long)
    word_mask = torch.zeros(batch_size, max_words, dtype=torch.bool)

    for i, (ids, starts, h, r, nw) in enumerate(zip(input_ids, word_starts, heads, rels, n_words)):
        padded_ids[i, :ids.size(0)] = ids
        attention_mask[i, :ids.size(0)] = 1
        padded_starts[i, :starts.size(0)] = starts
        padded_heads[i, :nw] = h
        padded_rels[i, :nw] = r
        word_mask[i, :nw] = True

    return padded_ids, attention_mask, padded_starts, padded_heads, padded_rels, word_mask


def evaluate_transformer(model, dataloader, device):
    """Evaluate transformer-based model."""
    model.eval()

    total_arcs = 0
    correct_arcs = 0
    correct_rels = 0

    with torch.no_grad():
        for batch in dataloader:
            input_ids, attn_mask, word_starts, heads, rels, mask = [x.to(device) for x in batch]

            word_hidden, word_mask = model.encode_pretokenized(input_ids, attn_mask, word_starts, mask)
            arc_scores, rel_scores = model(word_hidden, word_mask)
            arc_preds, rel_preds = model.decode(arc_scores, rel_scores, word_mask)

            arc_correct = (arc_preds == heads) & mask
            rel_correct = (rel_preds == rels) & mask & arc_correct

            total_arcs += mask.sum().item()
            correct_arcs += arc_correct.sum().item()
            correct_rels += rel_correct.sum().item()

    uas = correct_arcs / total_arcs * 100
    las = correct_rels / total_arcs * 100

    return uas, las


# ============================================================================
# Training
# ============================================================================

def evaluate(model, dataloader, device):
    """Evaluate model and return UAS/LAS."""
    model.eval()

    total_arcs = 0
    correct_arcs = 0
    correct_rels = 0

    with torch.no_grad():
        for batch in dataloader:
            words, chars, heads, rels, mask, lengths = [x.to(device) for x in batch]

            arc_scores, rel_scores = model(words, chars, mask)
            arc_preds, rel_preds = model.decode(arc_scores, rel_scores, mask)

            # Count correct
            arc_correct = (arc_preds == heads) & mask
            rel_correct = (rel_preds == rels) & mask & arc_correct

            total_arcs += mask.sum().item()
            correct_arcs += arc_correct.sum().item()
            correct_rels += rel_correct.sum().item()

    uas = correct_arcs / total_arcs * 100
    las = correct_rels / total_arcs * 100

    return uas, las


@click.command()
@click.option('--method', type=click.Choice(['baseline', 'trankit']), default='baseline',
              help='Parser method: baseline (BiLSTM) or trankit (XLM-RoBERTa)')
@click.option('--dataset', type=click.Choice(['udd1', 'ud-vtb', 'vndt']), default='udd1',
              help='Dataset: udd1 (UDD-1), ud-vtb (UD Vietnamese VTB), or vndt (VnDT v1.1)')
@click.option('--encoder', default='xlm-roberta-base',
              help='Transformer encoder for trankit method')
@click.option('--output', '-o', default='models/bamboo-1', help='Output directory')
@click.option('--epochs', default=100, type=int, help='Number of epochs')
@click.option('--batch-size', default=64, type=int, help='Batch size')
@click.option('--lr', default=2e-3, type=float, help='Learning rate for baseline')
@click.option('--bert-lr', default=2e-5, type=float, help='Encoder learning rate for trankit')
@click.option('--head-lr', default=2e-4, type=float, help='Head learning rate for trankit')
@click.option('--warmup-steps', default=200, type=int, help='Warmup steps for trankit')
@click.option('--lstm-hidden', default=400, type=int, help='LSTM hidden size (baseline)')
@click.option('--lstm-layers', default=3, type=int, help='LSTM layers (baseline)')
@click.option('--patience', default=5, type=int, help='Early stopping patience')
@click.option('--force-download', is_flag=True, help='Force re-download dataset')
@click.option('--data-dir', default=None, help='Custom data directory')
@click.option('--gpu-type', default='RTX_A4000', help='GPU type for cost estimation')
@click.option('--cost-interval', default=300, type=int, help='Cost report interval in seconds')
@click.option('--wandb', 'use_wandb', is_flag=True, help='Enable W&B logging')
@click.option('--wandb-project', default='bamboo-1', help='W&B project name')
@click.option('--max-time', default=0, type=int, help='Max training time in minutes (0=unlimited)')
@click.option('--sample', default=0, type=int, help='Sample N sentences from each split (0=all)')
@click.option('--eval-every', default=2, type=int, help='Evaluate every N epochs')
@click.option('--fp16', is_flag=True, default=True, help='Use mixed precision training')
def train(method, dataset, encoder, output, epochs, batch_size, lr, bert_lr, head_lr, warmup_steps,
          lstm_hidden, lstm_layers, patience, force_download, data_dir, gpu_type, cost_interval,
          use_wandb, wandb_project, max_time, sample, eval_every, fp16):
    """Train Bamboo-1 Vietnamese Dependency Parser."""

    # Detect hardware
    hardware = detect_hardware()
    detected_gpu_type = hardware.get_gpu_type()

    if gpu_type == "RTX_A4000":
        gpu_type = detected_gpu_type

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    click.echo(f"Using device: {device}")
    click.echo(f"Hardware: {hardware}")

    # CUDA optimizations
    if torch.cuda.is_available():
        torch.backends.cudnn.benchmark = True
        torch.backends.cuda.matmul.allow_tf32 = True
        torch.backends.cudnn.allow_tf32 = True

    # Mixed precision
    use_amp = fp16 and torch.cuda.is_available()
    scaler = torch.amp.GradScaler('cuda') if use_amp else None
    if use_amp:
        click.echo("Mixed precision (FP16): enabled")

    # Initialize wandb
    if use_wandb:
        import wandb
        wandb.init(
            project=wandb_project,
            config={
                "method": method,
                "dataset": dataset,
                "encoder": encoder if method == "trankit" else "bilstm",
                "epochs": epochs,
                "batch_size": batch_size,
                "lr": lr if method == "baseline" else bert_lr,
                "head_lr": head_lr if method == "trankit" else None,
                "lstm_hidden": lstm_hidden if method == "baseline" else None,
                "lstm_layers": lstm_layers if method == "baseline" else None,
                "patience": patience,
                "gpu_type": gpu_type,
                "hardware": hardware.to_dict(),
            }
        )
        click.echo(f"W&B logging enabled: {wandb.run.url}")

    click.echo("=" * 60)
    click.echo(f"Bamboo-1: Vietnamese Dependency Parser ({method.upper()})")
    click.echo("=" * 60)

    # Load corpus
    click.echo(f"\nLoading {dataset.upper()} corpus...")
    if dataset == 'udd1':
        corpus = UDD1Corpus(data_dir=data_dir, force_download=force_download)
    elif dataset == 'ud-vtb':
        corpus = UDVietnameseVTB(data_dir=data_dir, force_download=force_download)
    else:  # vndt
        corpus = VnDTCorpus(data_dir=data_dir, force_download=force_download)

    train_sents = read_conllu(corpus.train)
    dev_sents = read_conllu(corpus.dev)
    test_sents = read_conllu(corpus.test)

    # Sample subset if requested
    if sample > 0:
        train_sents = train_sents[:sample]
        dev_sents = dev_sents[:min(sample // 2, len(dev_sents))]
        test_sents = test_sents[:min(sample // 2, len(test_sents))]
        click.echo(f"  Sampling {sample} sentences...")

    click.echo(f"  Train: {len(train_sents)} sentences")
    click.echo(f"  Dev: {len(dev_sents)} sentences")
    click.echo(f"  Test: {len(test_sents)} sentences")

    # Build vocabulary
    click.echo("\nBuilding vocabulary...")
    vocab = Vocabulary(min_freq=2)
    vocab.build(train_sents)
    if method == "baseline":
        click.echo(f"  Words: {vocab.n_words}")
        click.echo(f"  Chars: {vocab.n_chars}")
    click.echo(f"  Relations: {vocab.n_rels}")

    # Create datasets and model based on method
    if method == "trankit":
        # Trankit method: XLM-RoBERTa + Biaffine
        click.echo(f"\nInitializing model with {encoder}...")
        model = TransformerDependencyParser(
            n_rels=vocab.n_rels,
            encoder=encoder,
        ).to(device)

        n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
        encoder_params = sum(p.numel() for p in model.encoder.parameters())
        head_params = n_params - encoder_params
        click.echo(f"  Total parameters: {n_params:,}")
        click.echo(f"  Encoder parameters: {encoder_params:,}")
        click.echo(f"  Head parameters: {head_params:,}")

        # Pre-tokenize datasets (tokenize once, not every epoch)
        click.echo("\nPre-tokenizing datasets...")
        use_pin = torch.cuda.is_available()
        train_dataset = TransformerDataset(train_sents, vocab, model.tokenizer)
        dev_dataset = TransformerDataset(dev_sents, vocab, model.tokenizer)
        test_dataset = TransformerDataset(test_sents, vocab, model.tokenizer)

        train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True,
                                  collate_fn=transformer_collate_fn, pin_memory=use_pin)
        dev_loader = DataLoader(dev_dataset, batch_size=batch_size,
                                collate_fn=transformer_collate_fn, pin_memory=use_pin)
        test_loader = DataLoader(test_dataset, batch_size=batch_size,
                                 collate_fn=transformer_collate_fn, pin_memory=use_pin)

        # Differential learning rates
        encoder_params_list = list(model.encoder.parameters())
        head_params_list = [p for n, p in model.named_parameters() if 'encoder' not in n]
        optimizer = AdamW([
            {'params': encoder_params_list, 'lr': bert_lr},
            {'params': head_params_list, 'lr': head_lr},
        ], weight_decay=0.01)

        # Learning rate scheduler with warmup
        total_steps = len(train_loader) * epochs
        def lr_lambda(step):
            if step < warmup_steps:
                return step / warmup_steps
            return max(0.0, (total_steps - step) / (total_steps - warmup_steps))
        scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)

        eval_fn = evaluate_transformer
    else:
        # Baseline method: BiLSTM + Biaffine
        train_dataset = DependencyDataset(train_sents, vocab)
        dev_dataset = DependencyDataset(dev_sents, vocab)
        test_dataset = DependencyDataset(test_sents, vocab)

        train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn)
        dev_loader = DataLoader(dev_dataset, batch_size=batch_size, collate_fn=collate_fn)
        test_loader = DataLoader(test_dataset, batch_size=batch_size, collate_fn=collate_fn)

        click.echo("\nInitializing BiLSTM model...")
        model = BiaffineDependencyParser(
            n_words=vocab.n_words,
            n_chars=vocab.n_chars,
            n_rels=vocab.n_rels,
            lstm_hidden=lstm_hidden,
            lstm_layers=lstm_layers,
        ).to(device)

        n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
        click.echo(f"  Parameters: {n_params:,}")

        optimizer = Adam(model.parameters(), lr=lr, betas=(0.9, 0.9))
        scheduler = ExponentialLR(optimizer, gamma=0.75 ** (1 / 5000))

        eval_fn = evaluate

    # Training
    click.echo(f"\nTraining for {epochs} epochs...")
    if max_time > 0:
        click.echo(f"Time limit: {max_time} minutes")
    output_path = Path(output)
    output_path.mkdir(parents=True, exist_ok=True)

    # Cost tracking
    cost_tracker = CostTracker(gpu_type=gpu_type)
    cost_tracker.report_interval = cost_interval
    cost_tracker.start()
    click.echo(f"Cost tracking: {gpu_type} @ ${cost_tracker.hourly_rate}/hr")

    best_las = -1
    no_improve = 0
    time_limit_seconds = max_time * 60 if max_time > 0 else float('inf')

    for epoch in range(1, epochs + 1):
        # Check time limit
        if cost_tracker.elapsed_seconds() >= time_limit_seconds:
            click.echo(f"\nTime limit reached ({max_time} minutes)")
            break
        model.train()
        total_loss = 0

        pbar = tqdm(train_loader, desc=f"Epoch {epoch:3d}", leave=False)
        for batch in pbar:
            optimizer.zero_grad()

            if method == "trankit":
                input_ids, attn_mask, word_starts, heads, rels, mask = [x.to(device) for x in batch]

                with torch.amp.autocast('cuda', enabled=use_amp):
                    word_hidden, word_mask = model.encode_pretokenized(input_ids, attn_mask, word_starts, mask)
                    arc_scores, rel_scores = model(word_hidden, word_mask)
                    loss = model.loss(arc_scores, rel_scores, heads, rels, mask)
            else:
                words, chars, heads, rels, mask, lengths = [x.to(device) for x in batch]
                arc_scores, rel_scores = model(words, chars, mask)
                loss = model.loss(arc_scores, rel_scores, heads, rels, mask)

            if use_amp and scaler:
                scaler.scale(loss).backward()
                scaler.unscale_(optimizer)
                nn.utils.clip_grad_norm_(model.parameters(), 5.0)
                scaler.step(optimizer)
                scaler.update()
            else:
                loss.backward()
                nn.utils.clip_grad_norm_(model.parameters(), 5.0)
                optimizer.step()

            scheduler.step()
            total_loss += loss.item()
            pbar.set_postfix({'loss': f'{loss.item():.4f}'})

        # Evaluate (skip if not eval epoch, unless last epoch)
        if epoch % eval_every != 0 and epoch != epochs:
            avg_loss = total_loss / len(train_loader)
            current_lr = optimizer.param_groups[0]['lr']
            click.echo(f"Epoch {epoch:3d} | Loss: {avg_loss:.4f} | LR: {current_lr:.2e}")
            continue

        dev_uas, dev_las = eval_fn(model, dev_loader, device)

        # Cost update
        progress = epoch / epochs
        current_cost = cost_tracker.current_cost()
        estimated_total_cost = cost_tracker.estimate_total_cost(progress)
        elapsed_minutes = cost_tracker.elapsed_seconds() / 60

        cost_status = cost_tracker.update(epoch, epochs)
        if cost_status:
            click.echo(f"  [{cost_status}]")

        avg_loss = total_loss / len(train_loader)
        click.echo(f"Epoch {epoch:3d} | Loss: {avg_loss:.4f} | "
                   f"Dev UAS: {dev_uas:.2f}% | Dev LAS: {dev_las:.2f}%")

        # Log to wandb
        if use_wandb:
            wandb.log({
                "epoch": epoch,
                "train/loss": avg_loss,
                "dev/uas": dev_uas,
                "dev/las": dev_las,
                "cost/current_usd": current_cost,
                "cost/estimated_total_usd": estimated_total_cost,
                "cost/elapsed_minutes": elapsed_minutes,
            })

        # Save best model
        if dev_las >= best_las:
            best_las = dev_las
            no_improve = 0
            if method == "trankit":
                config = {
                    'method': 'trankit',
                    'encoder': encoder,
                    'n_rels': vocab.n_rels,
                }
            else:
                config = {
                    'method': 'baseline',
                    'n_words': vocab.n_words,
                    'n_chars': vocab.n_chars,
                    'n_rels': vocab.n_rels,
                    'lstm_hidden': lstm_hidden,
                    'lstm_layers': lstm_layers,
                }
            # Save to local tmp first to avoid network filesystem issues
            import tempfile, shutil
            with tempfile.NamedTemporaryFile(suffix='.pt', delete=False) as tmp:
                torch.save({
                    'model': model.state_dict(),
                    'vocab': vocab,
                    'config': config,
                }, tmp.name)
                shutil.move(tmp.name, output_path / 'model.pt')
            click.echo(f"  -> Saved best model (LAS: {best_las:.2f}%)")
        else:
            no_improve += 1
            if no_improve >= patience:
                click.echo(f"\nEarly stopping after {patience} epochs without improvement")
                break

    # Final evaluation
    click.echo("\nLoading best model for final evaluation...")
    checkpoint = torch.load(output_path / 'model.pt', weights_only=False)
    model.load_state_dict(checkpoint['model'])

    test_uas, test_las = eval_fn(model, test_loader, device)
    click.echo(f"\nTest Results:")
    click.echo(f"  UAS: {test_uas:.2f}%")
    click.echo(f"  LAS: {test_las:.2f}%")

    click.echo(f"\nModel saved to: {output_path}")

    # Final cost summary
    final_cost = cost_tracker.current_cost()
    click.echo(f"\n{cost_tracker.summary(epoch, epochs)}")

    # Log final metrics to wandb
    if use_wandb:
        wandb.log({
            "test/uas": test_uas,
            "test/las": test_las,
            "cost/final_usd": final_cost,
        })
        wandb.finish()


if __name__ == '__main__':
    train()