train_transformer.py

"""
train_transformer.py
====================
训练层级 3DGS split 生成 Transformer。
支持单卡和多卡（DDP）自动切换。

单卡启动：
    python train_transformer.py --seq_paths sequences/*.pkl --codebook_dir ./codebooks

四卡启动：
    torchrun --nproc_per_node=4 train_transformer.py --seq_paths sequences/*.pkl --codebook_dir ./codebooks

修复的三个 NaN 地雷：
  1. NaN * 0 = NaN：_reg_loss 改用 torch.where 屏蔽 PAD，彻底切断 NaN 污染
  2. 缺少 Final LayerNorm：TransformerEncoder 加 norm 参数，约束残差流方差
  3. Softmax -inf → NaN：forward 里对 transformer 输出做 nan_to_num 保底清理

role 编码：
  0 = parent   1 = uncle   2 = child   3 = EOS   4 = PAD
"""

import os
import math
import argparse
import pickle
import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import Dataset, DataLoader, DistributedSampler

# ─────────────────────────────────────────────
# 常量
# ─────────────────────────────────────────────

ROLE_PARENT  = 0
ROLE_UNCLE   = 1
ROLE_CHILD   = 2
ROLE_EOS     = 3
ROLE_PAD     = 4

MAX_CHILDREN = 32
MAX_UNCLES   = 4
MAX_SEQ_LEN  = 1 + MAX_UNCLES + MAX_CHILDREN + 1   # = 38

N_SCALE  = 16384
N_ROT    = 16384
N_DC     = 4096
N_SH     = 4096
N_ROLE   = 4

CB_DIM = {
    'scale': 3,
    'rot':   4,
    'dc':    3,
    'sh':    45,
}

D_CB = 64

TOKEN_DTYPE = np.dtype([
    ('dx',         np.float32),
    ('dy',         np.float32),
    ('dz',         np.float32),
    ('scale_idx',  np.int32),
    ('rot_idx',    np.int32),
    ('dc_idx',     np.int32),
    ('sh_idx',     np.int32),
    ('opacity',    np.float32),
    ('role',       np.uint8),
])

LOSS_WEIGHTS = {
    'role':    0.5,
    'xyz':     1.0,
    'opacity': 2.0,
    'scale':   1.0,
    'rot':     1.0,
    'dc':      1.0,
    'sh':      1.0,
}

OPACITY_LOGIT_CLIP = 20.0
OPACITY_NORM_CLIP = OPACITY_LOGIT_CLIP / 10.0


def normalize_quaternions_np(rotations: np.ndarray) -> np.ndarray:
    rotations = rotations.astype(np.float32, copy=True)
    norms = np.linalg.norm(rotations, axis=1, keepdims=True)
    valid = np.isfinite(norms) & (norms > 1e-8)
    rotations = np.where(valid, rotations / np.maximum(norms, 1e-8), rotations)
    bad = ~valid.squeeze(1)
    if bad.any():
        rotations[bad] = np.array([1.0, 0.0, 0.0, 0.0], dtype=np.float32)
    return rotations


# ─────────────────────────────────────────────
# 分布式工具
# ─────────────────────────────────────────────

def is_dist() -> bool:
    return dist.is_available() and dist.is_initialized()

def get_rank() -> int:
    return dist.get_rank() if is_dist() else 0

def get_world_size() -> int:
    return dist.get_world_size() if is_dist() else 1

def is_main() -> bool:
    return get_rank() == 0

def setup_dist() -> bool:
    if 'RANK' not in os.environ:
        return False
    dist.init_process_group(backend='nccl')
    torch.cuda.set_device(int(os.environ['LOCAL_RANK']))
    return True

def cleanup_dist():
    if is_dist():
        dist.destroy_process_group()

def reduce_mean(tensor: torch.Tensor) -> float:
    if not is_dist():
        return tensor.item()
    rt = tensor.clone()
    dist.all_reduce(rt, op=dist.ReduceOp.SUM)
    return (rt / get_world_size()).item()


# ─────────────────────────────────────────────
# 1. Dataset
# ─────────────────────────────────────────────

class SplitSequenceDataset(Dataset):

    def __init__(self, seq_pkl_paths: list):
        self.sequences = []
        for path in seq_pkl_paths:
            with open(path, 'rb') as f:
                seqs = pickle.load(f)
            self.sequences.extend(seqs)
            if is_main():
                print(f"  加载 {os.path.basename(path)}：{len(seqs)} 条")
        if is_main():
            print(f"[Dataset] 共 {len(self.sequences)} 条序列，"
                  f"固定长度 {MAX_SEQ_LEN}")

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

    def __getitem__(self, idx):
        seq  = self.sequences[idx]
        role = seq['role'].astype(np.int64)

        attn_mask      = (role != ROLE_PAD)
        loss_mask_feat = (role == ROLE_CHILD)
        loss_mask_role = (role != ROLE_PAD)

        xyz = np.stack([seq['dx'], seq['dy'], seq['dz']], axis=1)

        # opacity 归一化到 [-1, 1] 附近
        opacity = np.nan_to_num(
            seq['opacity'].astype(np.float32),
            nan=0.0,
            posinf=OPACITY_LOGIT_CLIP,
            neginf=-OPACITY_LOGIT_CLIP,
        )
        opacity = np.clip(opacity, -OPACITY_LOGIT_CLIP, OPACITY_LOGIT_CLIP)
        opacity_norm = opacity / 10.0

        return {
            'xyz':             torch.from_numpy(xyz).float(),
            'scale':           torch.from_numpy(seq['scale_idx'].astype(np.int64)),
            'rot':             torch.from_numpy(seq['rot_idx'].astype(np.int64)),
            'dc':              torch.from_numpy(seq['dc_idx'].astype(np.int64)),
            'sh':              torch.from_numpy(seq['sh_idx'].astype(np.int64)),
            'opacity':         torch.from_numpy(opacity_norm),
            'role':            torch.from_numpy(role),
            'attn_mask':       torch.from_numpy(attn_mask),
            'loss_mask_feat':  torch.from_numpy(loss_mask_feat),
            'loss_mask_role':  torch.from_numpy(loss_mask_role),
        }


def collate_fn(batch):
    keys = ['xyz', 'scale', 'rot', 'dc', 'sh', 'opacity',
            'role', 'attn_mask', 'loss_mask_feat', 'loss_mask_role']
    return {k: torch.stack([b[k] for b in batch], dim=0) for k in keys}


# ─────────────────────────────────────────────
# 2. Token Embedding
# ─────────────────────────────────────────────

class TokenEmbedding(nn.Module):

    def __init__(self, d_model: int):
        super().__init__()
        d = d_model // 8

        # 输入侧：codebook 向量查表 + 小 Linear，不用大 Embedding table
        self.inp_proj_scale = nn.Linear(CB_DIM['scale'], d)
        self.inp_proj_rot   = nn.Linear(CB_DIM['rot'],   d)
        self.inp_proj_dc    = nn.Linear(CB_DIM['dc'],    d)
        self.inp_proj_sh    = nn.Linear(CB_DIM['sh'],    d)

        # role 只有 5 个值，小 Embedding 完全没问题
        self.emb_role = nn.Embedding(5, d, padding_idx=ROLE_PAD)

        self.xyz_norm  = nn.LayerNorm(3)
        self.proj_xyz  = nn.Linear(3, d * 2)
        self.proj_opa  = nn.Linear(1, d)

        self.proj = nn.Linear(d * 8, d_model)

    def forward(self,
                batch: dict,
                cb_scale: torch.Tensor,
                cb_rot:   torch.Tensor,
                cb_dc:    torch.Tensor,
                cb_sh:    torch.Tensor) -> torch.Tensor:

        with torch.no_grad():
            s_vec = cb_scale[batch['scale'].clamp(0, cb_scale.shape[0] - 1)]
            r_vec = cb_rot[  batch['rot'].clamp(0,   cb_rot.shape[0]   - 1)]
            d_vec = cb_dc[   batch['dc'].clamp(0,    cb_dc.shape[0]    - 1)]
            h_vec = cb_sh[   batch['sh'].clamp(0,    cb_sh.shape[0]    - 1)]

        # 【地雷修复】F.normalize 加 eps，防止零向量导致除以零
        e_s = self.inp_proj_scale(F.normalize(s_vec, dim=-1, eps=1e-8))
        e_r = self.inp_proj_rot(  F.normalize(r_vec, dim=-1, eps=1e-8))
        e_d = self.inp_proj_dc(   F.normalize(d_vec, dim=-1, eps=1e-8))
        e_h = self.inp_proj_sh(   F.normalize(h_vec, dim=-1, eps=1e-8))

        e_role = self.emb_role(batch['role'].clamp(0, 4))

        e_xyz  = self.proj_xyz(self.xyz_norm(batch['xyz'].float()))
        e_opa  = self.proj_opa(batch['opacity'].unsqueeze(-1).float())

        cat = torch.cat([e_xyz, e_s, e_r, e_d, e_h, e_opa, e_role], dim=-1)
        return self.proj(cat)


# ─────────────────────────────────────────────
# 3. Transformer Model
# ─────────────────────────────────────────────

class SplitTransformer(nn.Module):

    def __init__(
        self,
        d_model:      int   = 512,
        n_heads:      int   = 8,
        n_layers:     int   = 6,
        d_ff:         int   = 2048,
        max_seq_len:  int   = MAX_SEQ_LEN,
        dropout:      float = 0.1,
        codebook_dir: str   = None,
        d_cb:         int   = D_CB,
    ):
        super().__init__()
        self.d_model     = d_model
        self.max_seq_len = max_seq_len
        self.d_cb        = d_cb

        self.token_emb = TokenEmbedding(d_model)
        self.pos_emb   = nn.Embedding(max_seq_len, d_model)

        layer = nn.TransformerEncoderLayer(
            d_model=d_model,
            nhead=n_heads,
            dim_feedforward=d_ff,
            dropout=dropout,
            batch_first=True,
            norm_first=True,
        )

        # 【地雷二修复】加 Final LayerNorm，约束 Pre-LN 残差流方差
        final_norm = nn.LayerNorm(d_model)
        self.transformer = nn.TransformerEncoder(
            layer, num_layers=n_layers, norm=final_norm
        )

        self.register_buffer(
            'causal_mask',
            torch.triu(torch.ones(max_seq_len, max_seq_len), diagonal=1).bool()
        )

        # 输出头
        self.head_role    = nn.Linear(d_model, N_ROLE)
        self.head_xyz     = nn.Linear(d_model, 3)
        self.head_opacity = nn.Linear(d_model, 1)
        self.head_scale_emb = nn.Linear(d_model, d_cb)
        self.head_rot_emb   = nn.Linear(d_model, d_cb)
        self.head_dc_emb    = nn.Linear(d_model, d_cb)
        self.head_sh_emb    = nn.Linear(d_model, d_cb)

        # 输出侧 codebook 投影（冻结）
        self.cb_proj_scale = nn.Linear(CB_DIM['scale'], d_cb)
        self.cb_proj_rot   = nn.Linear(CB_DIM['rot'],   d_cb)
        self.cb_proj_dc    = nn.Linear(CB_DIM['dc'],    d_cb)
        self.cb_proj_sh    = nn.Linear(CB_DIM['sh'],    d_cb)

        if codebook_dir is not None:
            self._load_codebooks(codebook_dir)
        else:
            self.register_buffer('cb_scale', torch.zeros(1, CB_DIM['scale']))
            self.register_buffer('cb_rot',   torch.zeros(1, CB_DIM['rot']))
            self.register_buffer('cb_dc',    torch.zeros(1, CB_DIM['dc']))
            self.register_buffer('cb_sh',    torch.zeros(1, CB_DIM['sh']))

        self._init_weights()

        # 冻结 cb_proj
        for name in ['cb_proj_scale', 'cb_proj_rot', 'cb_proj_dc', 'cb_proj_sh']:
            for param in getattr(self, name).parameters():
                param.requires_grad_(False)

    def _load_codebooks(self, codebook_dir: str):
        name_map = {
            'scale':    'cb_scale',
            'rotation': 'cb_rot',
            'dc':       'cb_dc',
            'sh':       'cb_sh',
        }
        for file_name, buf_name in name_map.items():
            path = os.path.join(codebook_dir, f"{file_name}_codebook.npz")
            if not os.path.exists(path):
                raise FileNotFoundError(f"找不到 codebook：{path}")
            cb = np.load(path)['codebook'].astype(np.float32)
            if file_name == 'rotation':
                cb = normalize_quaternions_np(cb)
            self.register_buffer(buf_name, torch.from_numpy(cb))
            if is_main():
                print(f"  [codebook] {file_name}: {cb.shape}")

    def _init_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Linear):
                nn.init.xavier_uniform_(m.weight)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)
            elif isinstance(m, nn.Embedding):
                nn.init.normal_(m.weight, std=0.02)
                if m.padding_idx is not None:
                    nn.init.zeros_(m.weight[m.padding_idx])

        for head in [self.head_role, self.head_xyz, self.head_opacity,
                     self.head_scale_emb, self.head_rot_emb,
                     self.head_dc_emb, self.head_sh_emb]:
            nn.init.normal_(head.weight, std=0.02)
            nn.init.zeros_(head.bias)

    def forward(self, batch: dict) -> dict:
        B, L = batch['scale'].shape

        tok_emb = self.token_emb(
            batch,
            cb_scale=self.cb_scale,
            cb_rot=self.cb_rot,
            cb_dc=self.cb_dc,
            cb_sh=self.cb_sh,
        )

        pos = torch.arange(L, device=tok_emb.device)
        x   = tok_emb + self.pos_emb(pos).unsqueeze(0)

        pad_mask = ~batch['attn_mask']
        causal   = self.causal_mask[:L, :L]

        out = self.transformer(
            src=x,
            mask=causal,
            src_key_padding_mask=pad_mask,
        )

        # 【地雷三修复】清理 PAD 位置 softmax(-inf) 产生的 NaN
        # 只对废弃的 PAD 位置做保底，不影响有效位置的梯度
        out = torch.nan_to_num(out, nan=0.0)

        return {
            'role':      self.head_role(out),
            'xyz':       self.head_xyz(out),
            'opacity':   self.head_opacity(out),
            'scale_emb': self.head_scale_emb(out),
            'rot_emb':   self.head_rot_emb(out),
            'dc_emb':    self.head_dc_emb(out),
            'sh_emb':    self.head_sh_emb(out),
        }

    def get_cb_emb(self, name: str) -> torch.Tensor:
        cb   = getattr(self, f'cb_{name}')
        proj = getattr(self, f'cb_proj_{name}')
        with torch.no_grad():
            return proj(cb)

    def nearest_codebook_idx(self, pred_emb: torch.Tensor, name: str) -> int:
        cb_emb = self.get_cb_emb(name)
        dist2  = ((cb_emb - pred_emb.unsqueeze(0)) ** 2).sum(dim=-1)
        return int(dist2.argmin().item())


# ─────────────────────────────────────────────
# 4. Loss
# ─────────────────────────────────────────────

def compute_loss(pred: dict, batch: dict,
                 model: nn.Module,
                 weights: dict = None) -> tuple:
    if weights is None:
        weights = LOSS_WEIGHTS

    feat_mask = batch['loss_mask_feat'][:, 1:]
    role_mask = batch['loss_mask_role'][:, 1:]

    raw_model = model.module if hasattr(model, 'module') else model

    # 【地雷一修复】用 torch.where 代替乘法屏蔽，彻底切断 NaN * 0 = NaN
    def _reg_loss(pred_key, tgt_key, mask, squeeze=False, scale=1.0):
        p = pred[pred_key][:, :-1]
        t = batch[tgt_key][:, 1:]
        if squeeze:
            p = p.squeeze(-1)
        if not mask.any():
            return torch.tensor(0.0, device=p.device)

        p = torch.nan_to_num(p, nan=0.0, posinf=1e4, neginf=-1e4)
        t = torch.nan_to_num(t.float(), nan=0.0, posinf=1e4, neginf=-1e4)
        if p.dim() == 3:
            valid = mask & torch.isfinite(p).all(dim=-1) & torch.isfinite(t).all(dim=-1)
        else:
            valid = mask & torch.isfinite(p) & torch.isfinite(t)
        if not valid.any():
            return torch.tensor(0.0, device=p.device)

        mse = F.mse_loss(p / scale, t / scale, reduction='none')
        if mse.dim() == 3:
            mse = mse.mean(-1)

        # torch.where：mask=True 的位置保留 mse，mask=False 填 0.0
        # 彻底切断 PAD 位置 NaN 的污染（NaN * 0 = NaN，但 where 选 0.0 安全）
        masked_mse = torch.where(valid, mse, torch.zeros_like(mse))
        return masked_mse.sum() / valid.sum().clamp(min=1)

    def _opacity_loss(mask):
        p = pred['opacity'][:, :-1].squeeze(-1)
        t = batch['opacity'][:, 1:].float()
        p = torch.nan_to_num(
            p,
            nan=0.0,
            posinf=OPACITY_NORM_CLIP,
            neginf=-OPACITY_NORM_CLIP,
        )
        t = torch.nan_to_num(
            t,
            nan=0.0,
            posinf=OPACITY_NORM_CLIP,
            neginf=-OPACITY_NORM_CLIP,
        ).clamp(-OPACITY_NORM_CLIP, OPACITY_NORM_CLIP)
        valid = mask & torch.isfinite(p) & torch.isfinite(t)
        if not valid.any():
            return torch.tensor(0.0, device=p.device)
        loss = F.smooth_l1_loss(p, t, reduction='none', beta=0.25)
        loss = torch.where(valid, loss, torch.zeros_like(loss))
        return loss.sum() / valid.sum().clamp(min=1)

    def _cls_loss_role(mask):
        p = pred['role'][:, :-1]
        t = batch['role'][:, 1:]
        if not mask.any():
            return torch.tensor(0.0, device=p.device)
        # p[mask] 直接丢弃 PAD 位置，天然安全
        p_m = p[mask]
        t_m = t[mask]
        valid = (t_m >= 0) & (t_m < N_ROLE)
        if not valid.all():
            p_m, t_m = p_m[valid], t_m[valid]
        if p_m.numel() == 0:
            return torch.tensor(0.0, device=p.device)
        return F.cross_entropy(p_m, t_m, label_smoothing=0.1)

    def _emb_loss(pred_emb_key, tgt_idx_key, mask, cb_name):
        p     = pred[pred_emb_key][:, :-1]
        t_idx = batch[tgt_idx_key][:, 1:]
        if not mask.any():
            return torch.tensor(0.0, device=p.device)

        p_m     = p[mask]
        t_idx_m = t_idx[mask]

        cb      = getattr(raw_model, f'cb_{cb_name}')
        cb_proj = getattr(raw_model, f'cb_proj_{cb_name}')

        valid = (t_idx_m >= 0) & (t_idx_m < cb.shape[0])
        if not valid.all():
            p_m, t_idx_m = p_m[valid], t_idx_m[valid]
        if p_m.numel() == 0:
            return torch.tensor(0.0, device=p.device)

        with torch.no_grad():
            t_emb = cb_proj(cb[t_idx_m])

        # 两边 normalize 后算 MSE，梯度有界
        p_norm = F.normalize(p_m,   dim=-1, eps=1e-8)
        t_norm = F.normalize(t_emb, dim=-1, eps=1e-8)
        return F.mse_loss(p_norm, t_norm)

    loss_role  = _cls_loss_role(role_mask)
    loss_xyz   = _reg_loss('xyz',     'xyz',     feat_mask, scale=5.0)
    loss_opa   = _opacity_loss(feat_mask)
    loss_scale = _emb_loss('scale_emb', 'scale', feat_mask, 'scale')
    loss_rot   = _emb_loss('rot_emb',   'rot',   feat_mask, 'rot')
    loss_dc    = _emb_loss('dc_emb',    'dc',    feat_mask, 'dc')
    loss_sh    = _emb_loss('sh_emb',    'sh',    feat_mask, 'sh')

    total = (
        weights['role']    * loss_role  +
        weights['xyz']     * loss_xyz   +
        weights['opacity'] * loss_opa   +
        weights['scale']   * loss_scale +
        weights['rot']     * loss_rot   +
        weights['dc']      * loss_dc    +
        weights['sh']      * loss_sh
    )

    if not torch.isfinite(total):
        bad = {k: v.item() for k, v in {
            'role': loss_role, 'xyz': loss_xyz, 'opa': loss_opa,
            'scale': loss_scale, 'rot': loss_rot,
            'dc': loss_dc, 'sh': loss_sh,
        }.items() if not torch.isfinite(v)}
        if is_main():
            print(f"[NaN警告] 非有限 loss 来自：{bad}")
        total = torch.tensor(0.0, requires_grad=True, device=loss_role.device)

    return total, {
        'role':    loss_role.item(),
        'xyz':     loss_xyz.item(),
        'opacity': loss_opa.item(),
        'scale':   loss_scale.item(),
        'rot':     loss_rot.item(),
        'dc':      loss_dc.item(),
        'sh':      loss_sh.item(),
        'total':   total.item(),
    }


# ─────────────────────────────────────────────
# 5. 诊断（第一个 batch）
# ─────────────────────────────────────────────

def diagnose_first_batch(model, batch, loss_weights=None):
    if not is_main():
        return
    print("\n========== 第一个 batch 诊断 ==========")

    for key, val in batch.items():
        if not isinstance(val, torch.Tensor):
            continue
        if val.dtype == torch.float32:
            finite = torch.isfinite(val)
            val_finite = val[finite]
            if val_finite.numel() == 0:
                min_val, max_val = float('nan'), float('nan')
            else:
                min_val, max_val = val_finite.min().item(), val_finite.max().item()
            print(f"  batch[{key:16s}]: shape={str(val.shape):25s} "
                  f"nan={torch.isnan(val).sum().item()} "
                  f"inf={torch.isinf(val).sum().item()} "
                  f"min={min_val:10.4f} max={max_val:10.4f}")
        else:
            print(f"  batch[{key:16s}]: shape={str(val.shape):25s} "
                  f"dtype={val.dtype} "
                  f"min={val.min().item()} max={val.max().item()}")

    raw_model = model.module if hasattr(model, 'module') else model
    with torch.no_grad():
        pred_check = raw_model(batch)

    print()
    for key, val in pred_check.items():
        print(f"  pred[{key:12s}]: "
              f"nan={torch.isnan(val).sum().item()} "
              f"min={val.min().item():9.4f} "
              f"max={val.max().item():9.4f} "
              f"std={val.std().item():.4f}")

    _, loss_dict_check = compute_loss(pred_check, batch, model, weights=loss_weights)
    print()
    for key, val in loss_dict_check.items():
        print(f"  loss_{key:8s} = {val:.6f}")

    print("========================================\n")


# ─────────────────────────────────────────────
# 6. 训练主循环
# ─────────────────────────────────────────────

def train(
    seq_pkl_paths: list,
    codebook_dir:  str,
    save_dir:      str,
    d_model:       int   = 512,
    n_heads:       int   = 8,
    n_layers:      int   = 6,
    d_ff:          int   = 2048,
    d_cb:          int   = D_CB,
    dropout:       float = 0.1,
    batch_size:    int   = 64,
    lr:            float = 1e-4,
    epochs:        int   = 50,
    warmup_steps:  int   = 2000,
    grad_clip:     float = 1.0,
    val_ratio:     float = 0.05,
    save_every:    int   = 5,
    opacity_weight: float = LOSS_WEIGHTS['opacity'],
    num_workers:   int   = 4,
    val_num_workers: int = 2,
):
    use_ddp = setup_dist()

    if use_ddp:
        local_rank = int(os.environ['LOCAL_RANK'])
        device     = f'cuda:{local_rank}'
    elif torch.cuda.is_available():
        device = 'cuda'
    else:
        device = 'cpu'

    if is_main():
        print(f"[train] device={device}  "
              f"world_size={get_world_size()}  "
              f"DDP={'开启' if use_ddp else '关闭'}")
        print(f"[train] opacity_loss_weight={opacity_weight}")
        print(f"[train] dataloader_workers train={num_workers} val={val_num_workers}")
        os.makedirs(save_dir, exist_ok=True)

    loss_weights = dict(LOSS_WEIGHTS)
    loss_weights['opacity'] = opacity_weight

    # ── 数据集 ───────────────────────────────
    full_dataset = SplitSequenceDataset(seq_pkl_paths)
    n_val        = max(1, int(len(full_dataset) * val_ratio))
    n_train      = len(full_dataset) - n_val
    train_set, val_set = torch.utils.data.random_split(
        full_dataset, [n_train, n_val],
        generator=torch.Generator().manual_seed(42)
    )

    if use_ddp:
        train_sampler = DistributedSampler(train_set, shuffle=True)
        val_sampler   = DistributedSampler(val_set,   shuffle=False)
        train_loader  = DataLoader(
            train_set, batch_size=batch_size, sampler=train_sampler,
            collate_fn=collate_fn, num_workers=num_workers, pin_memory=True,
            persistent_workers=(num_workers > 0),
        )
        val_loader = DataLoader(
            val_set, batch_size=batch_size, sampler=val_sampler,
            collate_fn=collate_fn, num_workers=val_num_workers, pin_memory=True,
            persistent_workers=(val_num_workers > 0),
        )
    else:
        train_loader = DataLoader(
            train_set, batch_size=batch_size, shuffle=True,
            collate_fn=collate_fn, num_workers=num_workers, pin_memory=True,
            persistent_workers=(num_workers > 0),
        )
        val_loader = DataLoader(
            val_set, batch_size=batch_size, shuffle=False,
            collate_fn=collate_fn, num_workers=val_num_workers,
            pin_memory=True,
            persistent_workers=(val_num_workers > 0),
        )

    # ── 模型 ─────────────────────────────────
    model = SplitTransformer(
        d_model=d_model, n_heads=n_heads, n_layers=n_layers,
        d_ff=d_ff, max_seq_len=MAX_SEQ_LEN, dropout=dropout,
        codebook_dir=codebook_dir, d_cb=d_cb,
    ).to(device)

    if use_ddp:
        model = DDP(
            model,
            device_ids=[local_rank],
            output_device=local_rank,
            broadcast_buffers=False,
        )

    if is_main():
        raw = model.module if use_ddp else model
        n_params = sum(p.numel() for p in raw.parameters() if p.requires_grad)
        print(f"[train] 参数量：{n_params / 1e6:.2f}M")

    # ── 优化器（只更新未冻结参数）────────────
    optimizer = torch.optim.AdamW(
        filter(lambda p: p.requires_grad, model.parameters()),
        lr=lr, weight_decay=1e-2, eps=1e-8,
    )

    total_steps = epochs * len(train_loader)

    def lr_lambda(step):
        if step < warmup_steps:
            return step / max(1, warmup_steps)
        progress = (step - warmup_steps) / max(1, total_steps - warmup_steps)
        return max(0.1, 0.5 * (1 + math.cos(math.pi * progress)))

    scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)

    # ── 训练循环 ─────────────────────────────
    best_val_loss = float('inf')
    global_step   = 0

    for epoch in range(1, epochs + 1):
        if use_ddp:
            train_sampler.set_epoch(epoch)

        model.train()
        epoch_loss_sum = 0.0
        epoch_steps    = 0

        for batch in train_loader:
            batch = {k: v.to(device) if isinstance(v, torch.Tensor) else v
                     for k, v in batch.items()}

            if global_step == 0:
                diagnose_first_batch(model, batch, loss_weights=loss_weights)

            pred = model(batch)
            loss, loss_dict = compute_loss(pred, batch, model, weights=loss_weights)

            # NaN batch 保底跳过
            if not torch.isfinite(loss):
                if is_main():
                    print(f"  [step {global_step}] 跳过 NaN batch")
                optimizer.zero_grad()
                global_step += 1
                continue

            optimizer.zero_grad()
            loss.backward()

            # 梯度监控（前 20 步）
            if is_main() and global_step < 20:
                total_norm = 0.0
                for p in model.parameters():
                    if p.grad is not None:
                        total_norm += p.grad.data.norm(2).item() ** 2
                total_norm = total_norm ** 0.5
                print(f"  [step {global_step:03d}] "
                      f"loss={loss_dict['total']:.4f}  "
                      f"grad_norm={total_norm:.4f}")

            nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
            optimizer.step()
            scheduler.step()

            epoch_loss_sum += loss_dict['total']
            epoch_steps    += 1
            global_step    += 1

        train_loss = reduce_mean(torch.tensor(
            epoch_loss_sum / max(epoch_steps, 1), device=device
        ))

        # ── 验证 ─────────────────────────────
        model.eval()
        val_loss_sum = 0.0
        val_steps    = 0
        with torch.no_grad():
            for batch in val_loader:
                batch = {k: v.to(device) if isinstance(v, torch.Tensor) else v
                         for k, v in batch.items()}
                pred = model(batch)
                _, ld = compute_loss(pred, batch, model, weights=loss_weights)
                val_loss_sum += ld['total']
                val_steps    += 1

        val_loss = reduce_mean(torch.tensor(
            val_loss_sum / max(val_steps, 1), device=device
        ))

        if is_main():
            print(f"[epoch {epoch:03d}/{epochs}] "
                  f"train={train_loss:.4f}  val={val_loss:.4f}  "
                  f"lr={scheduler.get_last_lr()[0]:.2e}")

            raw_model = model.module if use_ddp else model

            if epoch % save_every == 0:
                ckpt_path = os.path.join(save_dir, f"ckpt_epoch{epoch:03d}.pt")
                torch.save({
                    'epoch':           epoch,
                    'model_state':     raw_model.state_dict(),
                    'optimizer_state': optimizer.state_dict(),
                    'val_loss':        val_loss,
                    'loss_weights':    loss_weights,
                    'config': dict(
                        d_model=d_model, n_heads=n_heads, n_layers=n_layers,
                        d_ff=d_ff, max_seq_len=MAX_SEQ_LEN, dropout=dropout,
                        d_cb=d_cb, codebook_dir=codebook_dir,
                    ),
                }, ckpt_path)
                print(f"            checkpoint → {ckpt_path}")

            if val_loss < best_val_loss:
                best_val_loss = val_loss
                best_path = os.path.join(save_dir, 'best_model.pt')
                torch.save({
                    'model_state': raw_model.state_dict(),
                    'loss_weights': loss_weights,
                    'config': dict(
                        d_model=d_model, n_heads=n_heads, n_layers=n_layers,
                        d_ff=d_ff, max_seq_len=MAX_SEQ_LEN, dropout=dropout,
                        d_cb=d_cb, codebook_dir=codebook_dir,
                    ),
                }, best_path)

    if is_main():
        print(f"\n[train] 训练完成！最优 val_loss={best_val_loss:.4f}")
        print(f"        最优模型 → {best_path}")

    cleanup_dist()


# ─────────────────────────────────────────────
# 7. CLI
# ─────────────────────────────────────────────

def parse_args():
    p = argparse.ArgumentParser(description="训练 3DGS split 生成 Transformer")
    p.add_argument('--seq_paths',    nargs='+', required=True)
    p.add_argument('--codebook_dir', required=True)
    p.add_argument('--save_dir',     default='./checkpoints')
    p.add_argument('--d_model',      type=int,   default=512)
    p.add_argument('--n_heads',      type=int,   default=8)
    p.add_argument('--n_layers',     type=int,   default=6)
    p.add_argument('--d_ff',         type=int,   default=2048)
    p.add_argument('--d_cb',         type=int,   default=D_CB)
    p.add_argument('--batch_size',   type=int,   default=64,
                   help='每张卡的 batch size')
    p.add_argument('--lr',           type=float, default=1e-4)
    p.add_argument('--epochs',       type=int,   default=50)
    p.add_argument('--warmup',       type=int,   default=2000)
    p.add_argument('--val_ratio',    type=float, default=0.05)
    p.add_argument('--save_every',   type=int,   default=5)
    p.add_argument('--dropout',      type=float, default=0.1)
    p.add_argument('--grad_clip',    type=float, default=1.0)
    p.add_argument('--opacity_weight', type=float, default=LOSS_WEIGHTS['opacity'],
                   help='Opacity reconstruction loss weight. Increase if inferred points become too opaque.')
    p.add_argument('--num_workers', type=int, default=4,
                   help='DataLoader workers for training.')
    p.add_argument('--val_num_workers', type=int, default=2,
                   help='DataLoader workers for validation.')
    return p.parse_args()


if __name__ == '__main__':
    args = parse_args()
    train(
        seq_pkl_paths=args.seq_paths,
        codebook_dir=args.codebook_dir,
        save_dir=args.save_dir,
        d_model=args.d_model,
        n_heads=args.n_heads,
        n_layers=args.n_layers,
        d_ff=args.d_ff,
        d_cb=args.d_cb,
        dropout=args.dropout,
        batch_size=args.batch_size,
        lr=args.lr,
        epochs=args.epochs,
        warmup_steps=args.warmup,
        val_ratio=args.val_ratio,
        save_every=args.save_every,
        grad_clip=args.grad_clip,
        opacity_weight=args.opacity_weight,
        num_workers=args.num_workers,
        val_num_workers=args.val_num_workers,
    )