OM_train_3modes.py

import os, sys, contextlib

ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(ROOT_DIR)

import gc
import torch
import torchvision
from torch import nn
from torchvision.utils import save_image
from torch.utils.data import DataLoader

from torch.optim import Adam, SGD
from Diffusion.diffuser import DeformDDPM
from Diffusion.networks import get_net, STN
# from torchvision.transforms import Lambda
import torch.nn.functional as F
import Diffusion.losses as losses
import random
import glob
import numpy as np
import utils
from tqdm import tqdm

# from Dataloader.dataloader0 import get_dataloader
from Dataloader.dataLoader import *

from Dataloader.dataloader_utils import thresh_img
import yaml
import argparse

# XPU support: import Intel Extension for PyTorch and oneCCL bindings if available
try:
    import intel_extension_for_pytorch as ipex
except ImportError:
    ipex = None
try:
    import oneccl_bindings_for_pytorch
except (ImportError, Exception) as e:
    print(f"WARNING: Failed to import oneccl_bindings_for_pytorch: {e}")

####################
import torch.multiprocessing as mp
from torch.utils.data.distributed import DistributedSampler
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.distributed as dist
# from torch.distributed import init_process_group
###############
def _device_available(device_type):
    if device_type == 'xpu':
        return hasattr(torch, 'xpu') and torch.xpu.is_available()
    return torch.cuda.is_available()

def _device_count(device_type):
    if device_type == 'xpu':
        return torch.xpu.device_count() if hasattr(torch, 'xpu') else 0
    return torch.cuda.device_count()

def _set_device(rank, device_type):
    if device_type == 'xpu':
        torch.xpu.set_device(rank)
    else:
        torch.cuda.set_device(rank)

def _empty_cache(device_type):
    if device_type == 'xpu' and hasattr(torch, 'xpu'):
        torch.xpu.empty_cache()
    elif torch.cuda.is_available():
        torch.cuda.empty_cache()

def ddp_setup(rank, world_size):
    """
    Args:
        rank: Unique identifier of each process (local_rank when launched by torchrun)
        world_size: Total number of processes
    """
    backend = "ccl" if DEVICE_TYPE == "xpu" else "nccl"
    if "LOCAL_RANK" in os.environ:
        # Launched by torchrun: MASTER_ADDR, MASTER_PORT, RANK, WORLD_SIZE already set
        dist.init_process_group(backend=backend)
        _set_device(int(os.environ["LOCAL_RANK"]), DEVICE_TYPE)
    else:
        # Single-node mp.spawn
        os.environ["MASTER_ADDR"] = "localhost"
        os.environ["MASTER_PORT"] = "12355"
        dist.init_process_group(backend=backend, rank=rank, world_size=world_size)
        _set_device(rank, DEVICE_TYPE)

EPS = 1e-5
MSK_EPS = 0.01
TEXT_EMBED_PROB = 0.5
AUG_RESAMPLE_PROB = 0.5
LOSS_WEIGHTS_DIFF = [4.0, 2.0, 8.0]  # [ang, dist, reg]
# LOSS_WEIGHTS_REGIST = [9.0, 1.0, 16.0]  # [imgsim, imgmse, ddf]
LOSS_WEIGHTS_REGIST = [1.0, 0.01, 1e2]  # [imgsim, imgmse, ddf]
DIFF_REG_BATCH_RATIO = 2
# LOSS_WEIGHT_CONTRASTIVE = 1e-4
LOSS_WEIGHT_CONTRASTIVE = 1e-1
REGISTRATION_STEP_RATIO = 1
CONTRASTIVE_STEP_RATIO = 1
ACCEPT_THRESH_CONTRASTIVE = 0.1
ACCEPT_THRESH_ANGLE = -0.8
MID_EPOCH_SAVE_STEPS = 1e4  # Save mid-epoch checkpoint every N steps for crash recovery.
                           # XPU autograd leaks ~1.0 GiB/step of device memory (Intel bug).
                           # With gradient checkpointing, training survives ~26 steps from fresh start,
                           # but fewer when carrying leaked memory from previous epoch.
                           # Save every 10 steps to minimize lost work on OOM crash.
EXIT_CODE_RESTART = 42     # Exit code signaling proactive restart (not a crash).

# AUG_PERMUTE_PROB = 0.35

parser = argparse.ArgumentParser()

# config_file_path = 'Config/config_cmr.yaml'
parser.add_argument(
        "--config",
        "-C",
        help="Path for the config file",
        type=str,
        # default="Config/config_cmr.yaml",
        # default="Config/config_lct.yaml",
        default="Config/config_all.yaml",
        required=False,
    )
parser.add_argument("--dummy-samples", type=int, default=0, help="Use dummy random data for testing (0=use real data)")
parser.add_argument("--batchsize", type=int, default=0, help="Override batch size from config (0=use config value)")
parser.add_argument("--max-steps-before-restart", type=int, default=0,
                    help="Proactive restart: exit after N training steps to reset XPU memory leak. "
                         "0=disabled (rely on OOM crash + auto-resubmit). "
                         "Recommended: 20 for XPU (survives ~26 steps max).")
parser.add_argument("--no-save", action="store_true", default=False,
                    help="Disable all checkpoint saving (for diagnostic/validation runs)")
parser.add_argument("--reset-optimizer", action="store_true",
                    help="Skip optimizer state loading from checkpoint (use when architecture changed)")
parser.add_argument("--eval-only", action="store_true",
                    help="Forward pass only: compute and print losses without backward/optimizer (no memory leak)")
args = parser.parse_args()

# Read config early to determine device type for DDP setup
with open(args.config, 'r') as _f:
    _cfg = yaml.safe_load(_f)
DEVICE_TYPE = _cfg.get('device', 'cuda')  # 'cuda' or 'xpu'

# Auto-detect: use DDP only when multiple devices are available
use_distributed = _device_available(DEVICE_TYPE) and _device_count(DEVICE_TYPE) > 1
# use_distributed = True
# use_distributed = False
#=======================================================================================================================

class _DummyIndiv(torch.utils.data.Dataset):
    def __init__(self, n, sz, embd_dim=1024):
        self.n, self.sz, self.embd_dim = n, sz, embd_dim
    def __len__(self): return self.n
    def __getitem__(self, i):
        return np.random.rand(1, self.sz, self.sz, self.sz).astype(np.float64), np.random.randn(self.embd_dim).astype(np.float32)

class _DummyPair(torch.utils.data.Dataset):
    def __init__(self, n, sz, embd_dim=1024):
        self.n, self.sz, self.embd_dim = n, sz, embd_dim
    def __len__(self): return self.n
    def __getitem__(self, i):
        return (np.random.rand(1, self.sz, self.sz, self.sz).astype(np.float64),
                np.random.rand(1, self.sz, self.sz, self.sz).astype(np.float64),
                np.random.randn(self.embd_dim).astype(np.float32),
                np.random.randn(self.embd_dim).astype(np.float32))


def main_train(rank=0,world_size=1,train_mode_ratio=1,thresh_imgsim=0.01):
    if use_distributed:
        ddp_setup(rank,world_size)

        if torch.distributed.is_initialized() and rank == 0:
            print(f"World size: {torch.distributed.get_world_size()}")
            print(f"Communication backend: {torch.distributed.get_backend()}")
            print(f"PYTORCH_ALLOC_CONF: {os.environ.get('PYTORCH_ALLOC_CONF', 'not set')}")
            if DEVICE_TYPE == 'xpu' and hasattr(torch, 'xpu'):
                props = torch.xpu.get_device_properties(0)
                print(f"XPU device: {props.name}, total memory: {props.total_memory / 1024**3:.2f} GiB")
    # gpu_id = global rank (for save/print guards); rank = local device index
    if "RANK" in os.environ:
        gpu_id = int(os.environ["RANK"])
        rank = int(os.environ["LOCAL_RANK"])
    else:
        gpu_id = rank
    
    # Load the YAML file into a dictionary
    with open(args.config, 'r') as file:
        hyp_parameters = yaml.safe_load(file)
    if args.batchsize > 0:
        hyp_parameters['batchsize'] = args.batchsize
    if gpu_id == 0:
        print(hyp_parameters)

    # epoch_per_save=10
    epoch_per_save=hyp_parameters['epoch_per_save']

    data_name=hyp_parameters['data_name']
    net_name = hyp_parameters['net_name']

    Net=get_net(net_name)

    suffix_pth=f'_{data_name}_{net_name}.pth'
    model_save_path = os.path.join('Models',f'{data_name}_{net_name}/')
    model_dir=model_save_path
    # transformer=utils.get_transformer(img_sz=hyp_parameters["ndims"]*[hyp_parameters['img_size']])
    
    # Data_Loader=get_dataloader(data_name=hyp_parameters['data_name'], mode='train')

    # tsfm = torchvision.transforms.Compose([
    #             torchvision.transforms.ToTensor(),
    #             ])

    # dataset = Data_Loader(target_res = [hyp_parameters["img_size"]]*hyp_parameters["ndims"], transforms=None, noise_scale=hyp_parameters['noise_scale'])
    # train_loader = DataLoader(
    #     dataset,
    #     batch_size=hyp_parameters['batchsize'],
    #     # shuffle=False,
    #     shuffle=True,
    #     drop_last=True,
    # )

    if args.dummy_samples > 0:
        dataset = _DummyIndiv(args.dummy_samples, hyp_parameters['img_size'])
        datasetp = _DummyPair(args.dummy_samples, hyp_parameters['img_size'])
    else:
        # dataset = OminiDataset_v1(transform=None)
        dataset = OMDataset_indiv(transform=None)
        # datasetp = OminiDataset_paired(transform=None)
        datasetp = OMDataset_pair(transform=None)

    if use_distributed:
        sampler = DistributedSampler(dataset, shuffle=True)
        sampler_p = DistributedSampler(datasetp, shuffle=True)
    else:
        sampler = None
        sampler_p = None

    train_loader = DataLoader(
        dataset,
        batch_size=hyp_parameters['batchsize'],
        shuffle=(sampler is None),
        drop_last=True,
        sampler=sampler,
    )
    train_loader_p = DataLoader(
        datasetp,
        batch_size=max(1, hyp_parameters['batchsize']//DIFF_REG_BATCH_RATIO),
        shuffle=(sampler_p is None),
        drop_last=True,
        sampler=sampler_p,
    )



    network = Net(
        n_steps=hyp_parameters["timesteps"],
        ndims=hyp_parameters["ndims"],
        num_input_chn = hyp_parameters["num_input_chn"],
        res = hyp_parameters['img_size']
    )
    # Enable gradient checkpointing on XPU to reduce peak activation memory.
    # XPU autograd leaks ~1.0 GiB/step; lower peak buys more steps before OOM.
    if DEVICE_TYPE == 'xpu' and hasattr(network, 'use_checkpoint'):
        network.use_checkpoint = True
        if gpu_id == 0:
            print("  [init] Gradient checkpointing enabled for XPU", flush=True)

    Deformddpm = DeformDDPM(
        network=network,
        n_steps=hyp_parameters["timesteps"],
        image_chw=[1] + [hyp_parameters["img_size"]]*hyp_parameters["ndims"],
        device=hyp_parameters["device"],
        batch_size=hyp_parameters["batchsize"],
        img_pad_mode=hyp_parameters["img_pad_mode"],
        v_scale=hyp_parameters["v_scale"],
    )


    ddf_stn = STN(
        img_sz=hyp_parameters["img_size"],
        ndims=hyp_parameters["ndims"],
        # padding_mode="zeros",
        padding_mode=hyp_parameters["padding_mode"],
        device=hyp_parameters["device"],
    )


    if use_distributed:
        device = f"{DEVICE_TYPE}:{rank}"
        # NO pre-allocation. CCL/oneDNN accumulate ~1.4 GiB/step of device memory outside
        # PyTorch's caching allocator. Pre-allocating steals from that budget:
        #   92% pre-alloc → crash at step 3, 78% → step 10, none (70% cap) → step 14.
        # Instead, use empty_cache() between training phases to release unused cached memory
        # back to the device for CCL/oneDNN.
        if gpu_id == 0 and DEVICE_TYPE == 'xpu' and hasattr(torch, 'xpu'):
            total_mem = torch.xpu.get_device_properties(rank).total_memory
            print(f"  [init] XPU device memory: {total_mem/1024**3:.1f} GiB, no pre-allocation (relying on empty_cache between phases)", flush=True)
        Deformddpm.to(device)
        Deformddpm = DDP(Deformddpm, device_ids=[rank], find_unused_parameters=True)
        ddf_stn.to(device)
    else:
        Deformddpm.to(hyp_parameters["device"])
        ddf_stn.to(hyp_parameters["device"])
    # ddf_stn = DDP(ddf_stn, device_ids=[rank])


    # mse = nn.MSELoss()
    # loss_reg = losses.Grad(penalty=['l1', 'negdetj'], ndims=hyp_parameters["ndims"])
    # loss_reg = losses.Grad(penalty=['l1', 'negdetj', 'range'], ndims=hyp_parameters["ndims"])
    loss_reg = losses.Grad(penalty=['l1', 'negdetj', 'range'], ndims=hyp_parameters["ndims"],outrange_thresh=0.2,outrange_weight=1e3)
    loss_reg1 = losses.Grad(penalty=['l1', 'negdetj', 'range'], ndims=hyp_parameters["ndims"],outrange_thresh=0.6,outrange_weight=1e3)

    loss_dist = losses.MRSE(img_sz=hyp_parameters["img_size"])
    # loss_ang = losses.MRSE(img_sz=hyp_parameters["img_size"])
    loss_ang = losses.NCC(img_sz=hyp_parameters["img_size"])
    loss_imgsim = losses.MSLNCC()
    loss_imgmse = losses.LMSE()

    optimizer = Adam(Deformddpm.parameters(), lr=hyp_parameters["lr"])
    # hyp_parameters["lr"]=0.00000001
    # optimizer_regist = Adam(Deformddpm.parameters(), lr=hyp_parameters["lr"]*0.01)
    # optimizer_regist = SGD(Deformddpm.parameters(), lr=hyp_parameters["lr"]*0.01, momentum=0.98)
    # optimizer = SGD(Deformddpm.parameters(), lr=hyp_parameters["lr"], momentum=0.9)

    # # LR scheduler ----- YHM
    # scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer, hyp_parameters["lr"], hyp_parameters["lr"]*10, step_size_up=500, step_size_down=500, mode='triangular', gamma=1.0, scale_fn=None, scale_mode='cycle', cycle_momentum=True, base_momentum=0.8, max_momentum=0.9, last_epoch=-1)

    # Deformddpm.network.load_state_dict(torch.load('/home/data/jzheng/Adaptive_Motion_Generator-master/models/1000.pth'))

    # check for existing models
    if not os.path.exists(model_dir):
        os.makedirs(model_dir, exist_ok=True)
    # Check for checkpoints: first check tmp/ for mid-epoch, then main dir for epoch-level
    tmp_dir = os.path.join(model_dir, "tmp")
    tmp_files = sorted(glob.glob(os.path.join(tmp_dir, "*.pth")))
    model_files = sorted(glob.glob(os.path.join(model_dir, "*.pth")))
    initial_step = 0

    # Epoch stats and RNG states to restore when resuming from mid-epoch checkpoint
    _resume_epoch_stats = None
    _resume_rng = None

    if tmp_files and not args.eval_only and args.max_steps_before_restart > 0:
        # Mid-epoch checkpoint: only use when proactive restart is enabled
        latest = tmp_files[-1]
        if gpu_id == 0:
            print(f"  [resume] Found mid-epoch checkpoint: {latest}")
        initial_epoch, Deformddpm, optimizer = ddp_load_dict(gpu_id, Deformddpm, optimizer, latest, use_distributed=use_distributed)
        basename = os.path.basename(latest)
        initial_step = int(basename.split('_step')[1].split('_')[0].split('.')[0])
        _ckpt = torch.load(latest, map_location='cpu', weights_only=False)
        _resume_epoch_stats = _ckpt.get('epoch_stats', None)
        del _ckpt
        if gpu_id == 0:
            print(f"  [resume] Resuming epoch {initial_epoch} from step {initial_step}"
                  f"{' (with epoch_stats)' if _resume_epoch_stats else ''}", flush=True)
    elif model_files:
        if gpu_id == 0:
            print(model_files)
        latest = model_files[-1]
        initial_epoch, Deformddpm, optimizer = ddp_load_dict(gpu_id, Deformddpm, optimizer, latest, use_distributed=use_distributed)
    else:
        initial_epoch = 0

    if gpu_id == 0:
        print('len_train_data: ',len(dataset))

    # Proactive restart: track steps since process start to exit before OOM.
    max_steps_restart = args.max_steps_before_restart
    steps_since_start = 0
    loss_contra_gate = 0.0

    # Training loop
    for epoch in range(initial_epoch,hyp_parameters["epoch"]):
        if use_distributed and sampler is not None:
            sampler.set_epoch(epoch)
            sampler_p.set_epoch(epoch)

        epoch_loss_tot = 0.0
        epoch_loss_gen_d = 0.0
        epoch_loss_gen_a = 0.0
        epoch_loss_reg = 0.0
        epoch_loss_regist = 0.0
        epoch_loss_imgsim = 0.0
        epoch_loss_imgmse = 0.0
        epoch_loss_ddfreg = 0.0
        epoch_loss_contrastive = 0.0
        total_contra = 0
        total_reg_restored = None
        total_contra_restored = None

        # Restore epoch accumulators from mid-epoch checkpoint (only for the resumed epoch)
        if _resume_epoch_stats is not None and epoch == initial_epoch:
            epoch_loss_tot = _resume_epoch_stats.get('epoch_loss_tot', 0.0)
            epoch_loss_gen_d = _resume_epoch_stats.get('epoch_loss_gen_d', 0.0)
            epoch_loss_gen_a = _resume_epoch_stats.get('epoch_loss_gen_a', 0.0)
            epoch_loss_reg = _resume_epoch_stats.get('epoch_loss_reg', 0.0)
            epoch_loss_regist = _resume_epoch_stats.get('epoch_loss_regist', 0.0)
            epoch_loss_imgsim = _resume_epoch_stats.get('epoch_loss_imgsim', 0.0)
            epoch_loss_imgmse = _resume_epoch_stats.get('epoch_loss_imgmse', 0.0)
            epoch_loss_ddfreg = _resume_epoch_stats.get('epoch_loss_ddfreg', 0.0)
            epoch_loss_contrastive = _resume_epoch_stats.get('epoch_loss_contrastive', 0.0)
            total_reg_restored = _resume_epoch_stats.get('total_reg', None)
            total_contra_restored = _resume_epoch_stats.get('total_contra', None)
            loss_nan_step = _resume_epoch_stats.get('loss_nan_step', 0)
            # RNG states are restored INSIDE the skip loop (at the last skipped step)
            # to avoid DataLoader __getitem__ calls corrupting the restored state.
            _resume_rng = {k: _resume_epoch_stats[k] for k in
                           ('rng_torch', 'rng_numpy', 'rng_python', 'rng_xpu', 'rng_cuda')
                           if k in _resume_epoch_stats}
            if gpu_id == 0:
                print(f"  [resume] Restored epoch stats from checkpoint (loss_tot={epoch_loss_tot:.4f})", flush=True)
            _resume_epoch_stats = None  # Only restore once
        else:
            loss_nan_step = 0  # only reset when NOT resuming mid-epoch

        # Set model inside to train model
        Deformddpm.train()

        total = min(len(train_loader), len(train_loader_p))
        total_reg = total // REGISTRATION_STEP_RATIO
        # Restore total_reg and total_contra from checkpoint if available (mid-epoch resume)
        if total_reg_restored is not None:
            total_reg = total_reg_restored
            total_reg_restored = None
        if total_contra_restored is not None:
            total_contra = total_contra_restored
            total_contra_restored = None
        # for step, batch in tqdm(enumerate(train_loader)):
        # for step, batch in tqdm(enumerate(train_loader)):
        # for step, batch in enumerate(train_loader_omni):
        for step, (batch, batch_p) in tqdm(enumerate(zip(train_loader, train_loader_p)), total=total):

            # Skip steps already completed (mid-epoch resume).
            # Checkpoint at step N is saved AFTER step N's training completes,
            # so step N itself must also be skipped (use <=, not <).
            if epoch == initial_epoch and initial_step > 0 and step <= initial_step:
                # Restore RNG at the last skipped step, AFTER DataLoader __getitem__
                # has consumed RNG for all skipped batches. This way the first
                # non-skipped step starts with exactly the saved RNG state.
                if step == initial_step and _resume_rng is not None:
                    # Restore rank 0's RNG as base state, then re-seed per-rank
                    # so each rank has independent RNG (matching continuous run's
                    # divergent-per-rank behavior). Without this, all ranks would
                    # share rank 0's RNG → correlated augmentation/dropout decisions.
                    if 'rng_torch' in _resume_rng:
                        torch.set_rng_state(_resume_rng['rng_torch'])
                    if 'rng_numpy' in _resume_rng:
                        np.random.set_state(_resume_rng['rng_numpy'])
                    if 'rng_python' in _resume_rng:
                        random.setstate(_resume_rng['rng_python'])
                    if 'rng_xpu' in _resume_rng and DEVICE_TYPE == 'xpu':
                        torch.xpu.set_rng_state(_resume_rng['rng_xpu'])
                    elif 'rng_cuda' in _resume_rng and torch.cuda.is_available():
                        torch.cuda.set_rng_state(_resume_rng['rng_cuda'])
                    # Per-rank re-seed: checkpoint only has rank 0's RNG state.
                    # Advance each rank's RNG by a deterministic offset so they
                    # diverge (as they would in a continuous run).
                    if gpu_id > 0:
                        rank_seed = gpu_id * 100003 + initial_step * 31
                        torch.manual_seed(torch.initial_seed() + rank_seed)
                        np.random.seed((np.random.get_state()[1][0] + rank_seed) % (2**31))
                        random.seed(random.getrandbits(32) + rank_seed)
                        if DEVICE_TYPE == 'xpu' and hasattr(torch, 'xpu'):
                            torch.xpu.manual_seed(torch.initial_seed() + rank_seed)
                        elif torch.cuda.is_available():
                            torch.cuda.manual_seed(torch.initial_seed() + rank_seed)
                    _resume_rng = None
                    if gpu_id == 0:
                        print(f"  [resume] RNG states restored at step {step} (per-rank re-seeded)", flush=True)
                continue

            # Free registration tensors from previous step
            x1 = y1 = ddf_comp = img_rec = img_diff = None
            ddf_rand = y1_proc = msk_tgt = img_save = None
            loss_regist = loss_sim = loss_mse = loss_ddf1 = None

            # Memory diagnostic (one per node via local rank 0) — only warn when abnormal
            # Normal at step start: ~16 GiB reserved, ~48 GiB free (of 64 GiB total)
            if rank == 0 and DEVICE_TYPE == 'xpu' and hasattr(torch, 'xpu'):
                torch.xpu.reset_peak_memory_stats(rank)
                free_mem, total_mem_dev = torch.xpu.mem_get_info(rank)
                used_gib = (total_mem_dev - free_mem) / 1024**3
                if used_gib > 24:  # Normal is ~16 GiB at step start; warn if accumulating
                    alloc = torch.xpu.memory_allocated() / 1024**3
                    reserved = torch.xpu.memory_reserved() / 1024**3
                    free_gib = free_mem / 1024**3
                    print(f"  [mem WARNING] gpu_id={gpu_id} epoch {epoch} step {step}: "
                          f"{used_gib:.1f} GiB used ({alloc:.1f} alloc / {reserved:.1f} reserved), "
                          f"{free_gib:.1f} GiB free", flush=True)
            
            # ==========================================================================
            # diffusion train on single image

            # x0 = batch # for omni dataset
            [x0,embd] = batch # for om dataset
            x0 = x0.to(hyp_parameters["device"]).type(torch.float32)
            # print('embd:', embd.shape)
            embd_dev = embd.to(hyp_parameters["device"]).type(torch.float32)
            if np.random.uniform(0,1)<TEXT_EMBED_PROB:
                embd_in = embd_dev
            else:
                embd_in = None

            n = x0.size()[0]  # batch_size -> n
            x0 = x0.to(hyp_parameters["device"])
            
            blind_mask = utils.get_random_deformed_mask(x0.shape[2:],apply_possibility=0.6).to(hyp_parameters["device"])

            # random deformation + rotation
            if hyp_parameters["ndims"]>2:
                if np.random.uniform(0,1)<AUG_RESAMPLE_PROB:
                    x0 = utils.random_resample(x0, deform_scale=0)
                # elif np.random.uniform(0,1)<AUG_RESAMPLE_PROB+AUG_PERMUTE_PROB:
                else:
                    [x0] = utils.random_permute([x0], select_dims=[-1,-2,-3])
            # x0 = transformer(x0)
            if hyp_parameters['noise_scale']>0:
                if np.random.uniform(0,1)<AUG_RESAMPLE_PROB:
                    x0 = thresh_img(x0, [0, 2*hyp_parameters['noise_scale']])
                x0 = x0 * (np.random.normal(1, hyp_parameters['noise_scale'] * 1)) + np.random.normal(0, hyp_parameters['noise_scale'] * 1)

            # Picking some noise for each of the images in the batch, a timestep and the respective alpha_bars
            t = torch.randint(0, hyp_parameters["timesteps"], (n,)).to(
                hyp_parameters["device"]
            )  # pick up a seq of rand number from 0 to 'timestep'

            # proc_type = random.choice(['adding', 'independ', 'downsample', 'slice', 'project', 'none', 'uncon', 'uncon', 'uncon'])
            proc_type = random.choice(['adding', 'downsample', 'slice', 'slice1', 'none', 'uncon', 'uncon', 'uncon'])
            # print('proc_type:', proc_type)
            ddpm = Deformddpm.module if use_distributed else Deformddpm
            cond_img, _, cond_ratio = ddpm.proc_cond_img(x0,proc_type=proc_type)

            if loss_contra_gate < ACCEPT_THRESH_CONTRASTIVE:
            
                pre_dvf_I,dvf_I = Deformddpm(img_org=x0, t=t, cond_imgs=cond_img, mask=blind_mask,proc_type=[],text=embd_in)  # forward diffusion process

                loss_tot=0

                loss_ddf = loss_reg(pre_dvf_I,img=x0)
                trm_pred = ddf_stn(pre_dvf_I, dvf_I)
                loss_gen_d = loss_dist(pred=trm_pred,inv_lab=dvf_I,ddf_stn=None,mask=blind_mask)
                loss_gen_a = loss_ang(pred=trm_pred,inv_lab=dvf_I,ddf_stn=None,mask=blind_mask)

                loss_tot += LOSS_WEIGHTS_DIFF[0] * loss_gen_a + LOSS_WEIGHTS_DIFF[1] * loss_gen_d
                loss_tot += LOSS_WEIGHTS_DIFF[2] * loss_ddf
                loss_tot = torch.sqrt(1.+MSK_EPS-cond_ratio) * loss_tot

                # >> JZ: print nan in x0
                if torch.isnan(x0).any():
                    print(f"*** Encountered NaN in input image x0 at epoch {epoch}, step {step}.")
                # >> JZ: print loss of ddf
                if loss_ddf>0.001:
                    print(f"*** High diffusion DDF loss at epoch {epoch}, step {step}: {loss_ddf.item()}.")
                # yu: check if loss_tot==nan or inf
                # Synchronize NaN skip across all DDP ranks to avoid collective desync
                # Use broadcast from rank 0 instead of all_reduce to avoid CCL hang on single-node XPU
                is_nan = torch.isnan(loss_tot) or torch.isinf(loss_tot)
                if use_distributed:
                    nan_flag = torch.tensor([1.0 if is_nan else 0.0], device=f"{DEVICE_TYPE}:{rank}")
                    dist.broadcast(nan_flag, src=0)
                    is_nan = nan_flag.item() > 0
                if is_nan:
                    if gpu_id == 0:
                        print(f"*** Encountered NaN or Inf loss at epoch {epoch}, step {step}. Skipping this batch.")
                    loss_nan_step += 1
                    continue
                if loss_nan_step > 5:
                    print(f"*** Too many NaN or Inf losses ({loss_nan_step} times) at epoch {epoch}, step {step}. Stopping training.")
                    raise ValueError("Too many NaN losses detected in loss_tot. Code terminated.")

                # ==========================================================================
                # Diffusion backward (no gradient clipping — diffusion dominates training)
                # print(loss_contra_gate)
                if (not args.eval_only):  # Skip backward when contrastive loss is high to avoid destabilizing diffusion training (especially early on)
                    optimizer.zero_grad()
                    loss_tot.backward()
                    optimizer.step()
                    
                epoch_loss_tot += loss_tot.item() / total
                epoch_loss_gen_d += loss_gen_d.item() / total
                epoch_loss_gen_a += loss_gen_a.item() / total
                epoch_loss_reg += loss_ddf.item() / total

                # Print running average every 20 steps in eval-only mode
                if args.eval_only and gpu_id == 0 and (step + 1) % 20 == 0:
                    n = step + 1
                    print(f"  [eval] step {step}: running_avg ang={epoch_loss_gen_a*total/n:.4f} "
                        f"dist={epoch_loss_gen_d*total/n:.4f} regul={epoch_loss_reg*total/n:.6f}", flush=True)

                # Free diffusion intermediates and aggressively release all memory to device.
                # XPU runtime leaks ~1.3 GiB/step outside the caching allocator.
                # gc.collect() + synchronize() + empty_cache() attempts to reclaim deferred/lazy allocations.
                loss_gen_a_val = loss_gen_a.item()
                
                # del pre_dvf_I, dvf_I, trm_pred, loss_tot, loss_gen_a, loss_gen_d, loss_ddf
                gc.collect()
                if DEVICE_TYPE == 'xpu':
                    torch.xpu.synchronize()
                    _empty_cache(DEVICE_TYPE)

                # Sync loss_gen_a across DDP ranks for contrastive and registration gating
                if use_distributed:
                    loss_gen_a_sync = torch.tensor([loss_gen_a_val], device=f"{DEVICE_TYPE}:{rank}")
                    dist.broadcast(loss_gen_a_sync, src=0)
                    loss_gen_a_gate = loss_gen_a_sync.item()
                else:
                    loss_gen_a_gate = loss_gen_a_val
                
                LOSS_WEIGHT_CONTRASTIVE=1e-4
            else:
                LOSS_WEIGHT_CONTRASTIVE=1e-1
                if gpu_id == 0:
                    print(f"  [train] step {step}: Skipping backward (contra_gate={loss_contra_gate:.4f})", flush=True)


            # ==========================================================================
            # Contrastive train on single image (text-image alignment)
            # Separate backward with gradient clipping to prevent destabilizing diffusion.
            loss_contra_val = None
            if step % CONTRASTIVE_STEP_RATIO == 0:
                n_contra = x0.size()[0]
                t_contra = torch.randint(0, hyp_parameters["timesteps"], (n_contra,)).to(hyp_parameters["device"])
                # Route through DDP wrapper and return img_embd directly so DDP
                # traces the correct subgraph (encoder + mid + attn + img2txt).
                img_embd = Deformddpm(img_org=(x0 * blind_mask).detach(), cond_imgs=cond_img.detach(), T=t_contra, output_embedding=True, text=None)  # [B, 1024]
                loss_contra_preweight = F.relu(1 - F.cosine_similarity(img_embd, embd_dev, dim=-1)-0.25).mean()
                loss_contra = LOSS_WEIGHT_CONTRASTIVE * loss_contra_preweight

                if not args.eval_only:
                    optimizer.zero_grad()
                    loss_contra.backward()
                    torch.nn.utils.clip_grad_norm_(Deformddpm.parameters(), max_norm=LOSS_WEIGHT_CONTRASTIVE*1)
                    optimizer.step()
                loss_contra_val = loss_contra.item()
                epoch_loss_contrastive += loss_contra_val / total * CONTRASTIVE_STEP_RATIO

                # else:
                #     if gpu_id == 0:
                #         print(f"*** Warning: Network does not have img_embd attribute for contrastive loss at epoch {epoch}, step {step}.")

            # Free remaining intermediates and aggressively release memory before registration
            if cond_img is not None:
                del cond_img
            if blind_mask is not None:
                del blind_mask
            gc.collect()
            if DEVICE_TYPE == 'xpu':
                torch.xpu.synchronize()
                _empty_cache(DEVICE_TYPE)
            
            # Sync loss_gen_a across DDP ranks for contrastive and registration gating
            if use_distributed:
                loss_contra_sync = torch.tensor([loss_contra_preweight], device=f"{DEVICE_TYPE}:{rank}")
                dist.broadcast(loss_contra_sync, src=0)
                loss_contra_gate = loss_contra_sync.item()
            else:
                loss_contra_gate = loss_contra_preweight

            # ==========================================================================
            # registration train on paired images
            # loss_gen_a_gate already synced across DDP ranks above
            do_regist = step % REGISTRATION_STEP_RATIO == 0 and (loss_contra_gate < ACCEPT_THRESH_CONTRASTIVE) and loss_gen_a_gate < ACCEPT_THRESH_ANGLE
            if do_regist:
                [x1, y1, _, embd_y] = batch_p
                if np.random.uniform(0,1)<TEXT_EMBED_PROB:
                    embd_y = embd_y.to(hyp_parameters["device"]).type(torch.float32)
                else:
                    embd_y = None

                x1 = x1.to(hyp_parameters["device"]).type(torch.float32)
                y1 = y1.to(hyp_parameters["device"]).type(torch.float32)
                n = x1.size()[0]  # batch_size -> n
                [x1, y1] = utils.random_permute([x1, y1], select_dims=[-1,-2,-3])
                if hyp_parameters['noise_scale']>0:
                    [x1, y1] = thresh_img([x1, y1], [0, 2*hyp_parameters['noise_scale']])
                    random_scale = np.random.normal(1, hyp_parameters['noise_scale'] * 1)
                    random_shift = np.random.normal(0, hyp_parameters['noise_scale'] * 1)
                    x1 = x1 * random_scale + random_shift
                    y1 = y1 * random_scale + random_shift

                scale_regist = np.random.uniform(0.0,0.5)
                select_timestep = np.random.randint(12, 32)  # select a random number of timesteps to sample, between 8 and 16
                T_regist = sorted(random.sample(range(int(hyp_parameters["timesteps"] * scale_regist),hyp_parameters["timesteps"]), select_timestep), reverse=True)

                T_regist = [[t for _ in range(max(1, hyp_parameters["batchsize"]//2))] for t in T_regist]

                proc_type = random.choice(['downsample', 'slice', 'slice1', 'none', 'none'])
                ddpm_inner = Deformddpm.module if use_distributed else Deformddpm
                y1_proc, msk_tgt, cond_ratio = ddpm_inner.proc_cond_img(y1,proc_type=proc_type)
                msk_tgt = msk_tgt+MSK_EPS
                [ddf_comp,ddf_rand],[img_rec,img_diff,img_save],_ = Deformddpm(img_org=x1, cond_imgs=y1_proc, T=[None, T_regist], proc_type=[],text=embd_y)  # forward diffusion process
                loss_sim = loss_imgsim(img_rec, y1, label=msk_tgt*(y1>thresh_imgsim))  # calculate loss for the registration process
                loss_mse = loss_imgmse(img_rec, y1, label=msk_tgt*(y1>=0.0))  # calculate loss for the registration process
                loss_ddf1 = loss_reg1(ddf_comp, img=y1)  # calculate loss for the registration process

                loss_regist = 0
                loss_regist += LOSS_WEIGHTS_REGIST[0] * loss_sim
                loss_regist += LOSS_WEIGHTS_REGIST[1] * loss_mse
                loss_regist += LOSS_WEIGHTS_REGIST[2] * loss_ddf1

                # >> JZ: print nan in x0
                if torch.isnan(x0).any():
                    print(f"*** Encountered NaN in input image x0 at epoch {epoch}, step {step}.")
                # >> JZ: print loss of ddf
                if loss_ddf1>0.002:
                    print(f"*** High registration DDF loss at epoch {epoch}, step {step}: {loss_ddf1.item()}.")

                loss_regist = torch.sqrt(cond_ratio+MSK_EPS) *loss_regist
                if not args.eval_only:
                    optimizer.zero_grad()
                    loss_regist.backward()

                    torch.nn.utils.clip_grad_norm_(Deformddpm.parameters(), max_norm=0.02)
                    optimizer.step()

                epoch_loss_regist += loss_regist.item()
                epoch_loss_imgsim += loss_sim.item()
                epoch_loss_imgmse += loss_mse.item()
                epoch_loss_ddfreg += loss_ddf1.item()
            else:
                loss_sim = torch.tensor(0.0)
                loss_mse = torch.tensor(0.0)
                loss_ddf1 = torch.tensor(0.0)
                loss_regist = torch.tensor(0.0)
                if step % REGISTRATION_STEP_RATIO==0:
                    total_reg = total_reg-1

            # print for checking
            if step % 10 == 0:
                print('step:',step,':', loss_tot.item(),'=',loss_gen_a.item(),'+', loss_gen_d.item(),'+',loss_ddf.item())
                print(f'-     loss_regist: {loss_regist} = {loss_sim} (imgsim) + {loss_mse} (imgmse) + {loss_ddf1} (ddf)')
                print(f'-     loss_contra: {loss_contra}')
            
            # Mid-epoch checkpoint and proactive restart (only when --max-steps-before-restart > 0)
            if max_steps_restart > 0 and step > 0 and step % MID_EPOCH_SAVE_STEPS == 0 and gpu_id == 0 and not args.no_save:
                _epoch_stats = {
                    'epoch_loss_tot': epoch_loss_tot,
                    'epoch_loss_gen_d': epoch_loss_gen_d,
                    'epoch_loss_gen_a': epoch_loss_gen_a,
                    'epoch_loss_reg': epoch_loss_reg,
                    'epoch_loss_regist': epoch_loss_regist,
                    'epoch_loss_imgsim': epoch_loss_imgsim,
                    'epoch_loss_imgmse': epoch_loss_imgmse,
                    'epoch_loss_ddfreg': epoch_loss_ddfreg,
                    'epoch_loss_contrastive': epoch_loss_contrastive,
                    'total_reg': total_reg,
                    'total_contra': total_contra,
                    'loss_nan_step': loss_nan_step,
                    'rng_torch': torch.get_rng_state(),
                    'rng_numpy': np.random.get_state(),
                    'rng_python': random.getstate(),
                    **(({'rng_xpu': torch.xpu.get_rng_state()} if DEVICE_TYPE == 'xpu' and hasattr(torch, 'xpu') else
                        {'rng_cuda': torch.cuda.get_rng_state()} if torch.cuda.is_available() else {})),
                }
                tmp_dir = os.path.join(model_save_path, "tmp")
                os.makedirs(tmp_dir, exist_ok=True)
                for old_f in glob.glob(os.path.join(tmp_dir, "*.pth")):
                    os.remove(old_f)
                mid_save = os.path.join(tmp_dir, f"{epoch:06d}_step{step:04d}{suffix_pth}")
                state = Deformddpm.module.state_dict() if use_distributed else Deformddpm.state_dict()
                torch.save({
                    'model_state_dict': state,
                    'optimizer_state_dict': optimizer.state_dict(),
                    'epoch': epoch,
                    'step': step,
                    'epoch_stats': _epoch_stats,
                }, mid_save)
                print(f"  [mid-epoch] Saved checkpoint at epoch {epoch} step {step}: {mid_save}", flush=True)

            # Proactive restart: exit cleanly after N steps to reset XPU memory leak.
            # The bash wrapper will re-launch srun within the same SLURM allocation.
            steps_since_start += 1
            if max_steps_restart > 0 and steps_since_start >= max_steps_restart:
                # Save checkpoint at current position (if not just saved above)
                if not (step > 0 and step % MID_EPOCH_SAVE_STEPS == 0) and gpu_id == 0 and not args.no_save:
                    _epoch_stats = {
                        'epoch_loss_tot': epoch_loss_tot, 'epoch_loss_gen_d': epoch_loss_gen_d,
                        'epoch_loss_gen_a': epoch_loss_gen_a, 'epoch_loss_reg': epoch_loss_reg,
                        'epoch_loss_regist': epoch_loss_regist, 'epoch_loss_imgsim': epoch_loss_imgsim,
                        'epoch_loss_imgmse': epoch_loss_imgmse, 'epoch_loss_ddfreg': epoch_loss_ddfreg,
                        'epoch_loss_contrastive': epoch_loss_contrastive, 'total_reg': total_reg, 'total_contra': total_contra,
                        'loss_nan_step': loss_nan_step,
                        'rng_torch': torch.get_rng_state(), 'rng_numpy': np.random.get_state(),
                        'rng_python': random.getstate(),
                        **(({'rng_xpu': torch.xpu.get_rng_state()} if DEVICE_TYPE == 'xpu' and hasattr(torch, 'xpu') else
                            {'rng_cuda': torch.cuda.get_rng_state()} if torch.cuda.is_available() else {})),
                    }
                    tmp_dir = os.path.join(model_save_path, "tmp")
                    os.makedirs(tmp_dir, exist_ok=True)
                    for old_f in glob.glob(os.path.join(tmp_dir, "*.pth")):
                        os.remove(old_f)
                    mid_save = os.path.join(tmp_dir, f"{epoch:06d}_step{step:04d}{suffix_pth}")
                    state = Deformddpm.module.state_dict() if use_distributed else Deformddpm.state_dict()
                    torch.save({
                        'model_state_dict': state,
                        'optimizer_state_dict': optimizer.state_dict(),
                        'epoch': epoch,
                        'step': step,
                        'epoch_stats': _epoch_stats,
                    }, mid_save)
                    print(f"  [restart] Saved checkpoint at epoch {epoch} step {step}: {mid_save}", flush=True)
                if gpu_id == 0:
                    print(f"  [restart] Proactive restart after {steps_since_start} steps "
                          f"(limit {max_steps_restart}). Exiting with code {EXIT_CODE_RESTART}.", flush=True)
                # Clean shutdown
                _empty_cache(DEVICE_TYPE)
                gc.collect()
                if use_distributed and dist.is_initialized():
                    dist.barrier()
                    dist.destroy_process_group()
                sys.exit(EXIT_CODE_RESTART)

        if gpu_id == 0:
            print('==================')
            print(epoch,':', epoch_loss_tot,'=',epoch_loss_gen_a,'+', epoch_loss_gen_d,'+',epoch_loss_reg, ' (ang+dist+regul)')
            print(f'     loss_contrastive: {epoch_loss_contrastive}')
            total_reg_safe = max(total_reg, 1)
            print(f'     loss_regist: {epoch_loss_regist/total_reg_safe} = {epoch_loss_imgsim/total_reg_safe} (imgsim) + {epoch_loss_imgmse/total_reg_safe} (imgmse) + {epoch_loss_ddfreg/total_reg_safe} (ddf)')
            print('==================')


        if 0 == epoch % epoch_per_save and not args.no_save:
            save_dir=model_save_path + str(epoch).rjust(6, '0') + suffix_pth    
            os.makedirs(os.path.dirname(model_save_path), exist_ok=True)
            # break   # FOR TESTING
            if not use_distributed:
                print(f"saved in {save_dir}")
                # torch.save(Deformddpm.state_dict(), save_dir)
                torch.save({
                    'model_state_dict': Deformddpm.state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                    'epoch': epoch
                }, save_dir)
            elif gpu_id == 0:
                print(f"saved in {save_dir}")
                # torch.save(Deformddpm.module.state_dict(), save_dir)
                torch.save({
                    'model_state_dict': Deformddpm.module.state_dict(),
                    'optimizer_state_dict': optimizer.state_dict(),
                    'epoch': epoch
                }, save_dir)
        # Clean up tmp/ mid-epoch checkpoints after completed epoch
        if gpu_id == 0 and not args.no_save:
            tmp_dir = os.path.join(model_dir, "tmp")
            tmp_pths = glob.glob(os.path.join(tmp_dir, "*.pth"))
            if tmp_pths:
                for f in tmp_pths:
                    os.remove(f)
                print(f"  [cleanup] Cleared {len(tmp_pths)} tmp/ mid-epoch checkpoints", flush=True)
        # Reset initial_step after first epoch completes (no more skipping)
        initial_step = 0

        # XPU CCL workaround: restart after each epoch to avoid CCL hang on 2nd epoch.
        # CCL's Level Zero IPC handles accumulate and cause deadlock after ~200+ collectives.
        # A fresh process resets the L0 context. The bash loop catches exit code 42 and restarts.
        if DEVICE_TYPE == 'xpu' and use_distributed:
            if gpu_id == 0:
                print(f"  [xpu-restart] Epoch {epoch} done. Restarting to reset CCL state.", flush=True)
            _empty_cache(DEVICE_TYPE)
            gc.collect()
            if dist.is_initialized():
                dist.barrier()
                dist.destroy_process_group()
            sys.exit(EXIT_CODE_RESTART)

    # Resource cleanup at the end of training
    _empty_cache(DEVICE_TYPE)
    gc.collect()
    if use_distributed and dist.is_initialized():
        dist.destroy_process_group()

def ddp_load_dict(gpu_id, Deformddpm, optimizer, model_file,use_distributed=True, load_strict=False):

    # All ranks load checkpoint so optimizer state is consistent across DDP processes.
    # (Optimizer state includes per-parameter Adam momentum/variance which are NOT
    # broadcast — only model weights are broadcast. Without this, non-rank-0 processes
    # would have fresh Adam state after restart.)
    gc.collect()
    _empty_cache(DEVICE_TYPE)
    if gpu_id == 0:
        utils.print_memory_usage("Before Loading Model")
    # checkpoint = torch.load(model_file, map_location='cpu', weights_only=False)
    checkpoint = torch.load(model_file, map_location='cpu')
    if use_distributed:
        Deformddpm.module.load_state_dict(checkpoint['model_state_dict'], strict=load_strict)
    else:
        Deformddpm.load_state_dict(checkpoint['model_state_dict'], strict=load_strict)
    # Restore optimizer state when available (needed for mid-epoch resume).
    # Selective loading: load states for parameters with matching shapes, skip mismatched ones
    # (e.g., UpsampleConv replaced ConvTranspose3d — different kernel shapes).
    # After one epoch, the saved checkpoint will have correct state for ALL parameters.
    if 'optimizer_state_dict' in checkpoint and not args.reset_optimizer:
        saved_opt = checkpoint['optimizer_state_dict']
        saved_state = saved_opt.get('state', {})
        param_list = [p for group in optimizer.param_groups for p in group['params']]

        # Check if all shapes match (fast path: full load)
        all_match = True
        skipped = 0
        for idx, s in saved_state.items():
            if int(idx) < len(param_list):
                p = param_list[int(idx)]
                for k, v in s.items():
                    if isinstance(v, torch.Tensor) and v.dim() > 0 and v.shape != p.shape:
                        all_match = False
                        break
                if not all_match:
                    break

        if all_match:
            optimizer.load_state_dict(saved_opt)
        else:
            # Selective load: restore param_groups settings (lr, betas, etc.)
            for saved_g, group in zip(saved_opt['param_groups'], optimizer.param_groups):
                for k, v in saved_g.items():
                    if k != 'params':
                        group[k] = v
            # Restore per-parameter state only where shapes match
            for idx, s in saved_state.items():
                idx_int = int(idx)
                if idx_int < len(param_list):
                    p = param_list[idx_int]
                    shapes_ok = all(
                        v.shape == p.shape for k, v in s.items()
                        if isinstance(v, torch.Tensor) and v.dim() > 0
                    )
                    if shapes_ok:
                        # Cast state tensors to match parameter dtype/device
                        new_state = {}
                        for k, v in s.items():
                            if isinstance(v, torch.Tensor):
                                new_state[k] = v.to(dtype=p.dtype, device=p.device) if v.dim() > 0 else v
                            else:
                                new_state[k] = v
                        optimizer.state[p] = new_state
                    else:
                        skipped += 1
            if gpu_id == 0:
                loaded = len(saved_state) - skipped
                print(f"  [checkpoint] Selective optimizer load: {loaded} params restored, "
                      f"{skipped} skipped (shape mismatch, fresh Adam for those)", flush=True)
    elif args.reset_optimizer and gpu_id == 0:
        print("  [checkpoint] --reset-optimizer: skipping optimizer state, starting fresh Adam", flush=True)
    del checkpoint
    if gpu_id == 0:
        utils.print_memory_usage("After Loading Checkpoint on GPU")

    if use_distributed:
        # Broadcast model weights from rank 0 to ensure exact consistency
        dist.barrier()
        for param in Deformddpm.parameters():
            dist.broadcast(param.data, src=0)
        
    # get the epoch number from the filename
    basename = os.path.basename(model_file)
    epoch_from_file = int(basename[:6])
    if '_step' in basename:
        # Mid-epoch checkpoint: resume at same epoch (don't +1)
        initial_epoch = epoch_from_file
    else:
        # End-of-epoch checkpoint: start next epoch
        initial_epoch = epoch_from_file + 1

    return initial_epoch, Deformddpm, optimizer

            

if __name__ == "__main__":
    if "LOCAL_RANK" in os.environ:
        # Multi-node: launched by torchrun / srun
        use_distributed = True
        local_rank = int(os.environ["LOCAL_RANK"])
        world_size = int(os.environ["WORLD_SIZE"])
        print(f"torchrun launch: LOCAL_RANK={local_rank}, RANK={os.environ.get('RANK')}, WORLD_SIZE={world_size}")
        try:
            main_train(local_rank, world_size)
        except Exception as e:
            import traceback
            print(f"\n{'='*60}\nRANK {os.environ.get('RANK')} FAILED:\n{'='*60}", flush=True)
            traceback.print_exc()
            raise
    elif use_distributed:
        # Single-node multi-GPU: use mp.spawn
        world_size = _device_count(DEVICE_TYPE)
        print(f"Distributed {DEVICE_TYPE.upper()} device number = {world_size}")
        mp.spawn(main_train,args = (world_size,),nprocs = world_size)
    else:
        main_train(0,1)