train.py

import os
import math
import torch
import numpy as np
import matplotlib.pyplot as plt
import wandb, comet_ml
import random, time
import gc
import bitsandbytes as bnb
import torch.nn.functional as F
import argparse

from datetime import datetime
from diffusers import CosmosTransformer3DModel, AutoencoderKLQwenImage, FlowMatchEulerDiscreteScheduler
from transformers import Qwen3_5Tokenizer, Qwen3_5ForConditionalGeneration
from torch.utils.data import DataLoader, Sampler
from torch.optim.lr_scheduler import LambdaLR
from collections import defaultdict
from accelerate import Accelerator
from datasets import load_from_disk,concatenate_datasets
from tqdm import tqdm
from PIL import Image, ImageOps
from torch.utils.checkpoint import checkpoint
from diffusers.models.attention_processor import AttnProcessor2_0
from contextlib import nullcontext
from transformers.optimization import Adafactor

# Импортируем наш новый оптимизатор
from babka_sotona import BabkaSotona
from train_monitor import TrainMonitor

# Muon not tested! pip install git+https://github.com/recoilme/muon_adamw8bit.git
from muon_adamw8bit import MuonAdamW8bit

#os.environ['MASTER_ADDR'] = '127.0.0.1'
#os.environ["NCCL_P2P_DISABLE"] = "1"
#os.environ["NCCL_IB_DISABLE"] = "1" # comment this on H100!
#os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

# --------------------------- Параметры ---------------------------
ds_path = "datasets"
project = "transformer"

gpu_mem_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
local_bs = max(1, int((gpu_mem_gb / 32) * 7))
num_gpus = torch.cuda.device_count()
batch_size = local_bs * num_gpus

base_learning_rate = 1e-4
min_learning_rate = 5e-5

learning_rate_scale = 8
base_learning_rate = base_learning_rate / learning_rate_scale
min_learning_rate = min_learning_rate / learning_rate_scale
print(f"Calculated params max-lr:{base_learning_rate} min-lr:{min_learning_rate} GPUs: {num_gpus}, Global BS: {batch_size}")

num_epochs = num_gpus
sink_interval_share = 60
sample_interval_min = 60
cfg_dropout = 0.05
world_dropout = 0.01
# Время t, bias = -0.5 (Фокус на Деталях ~300) bias = 0.5 (Фокус на структуре) bias = 0 (колокол/ равномерно)
sigmoid_bias = 0.0
max_length = 250
use_precomputed_embeddings = False
use_wandb = True
use_comet_ml = False
save_model = True
disable_samples = False
use_decay = False
fbp = False
torch_compile = False
transformer_gradient = True 
loss_normalize = False
fixed_seed = False
shuffle = True
crossattn = False
optimizer_type = "adam"

if optimizer_type == "muon_adam8bit":
    batch_size = num_gpus * max(1, int((gpu_mem_gb / 32) * 3))
    muon_lr_scale = 500
    
comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r" 
comet_ml_workspace = "recoilme" 
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_math_sdp(True) 
save_barrier = 1.4
warmup_percent = 0.0025
betta2 = 0.997
eps = 1e-6
clip_grad_norm = 1.0
limit = 0#2000
checkpoints_folder = ""
gradient_accumulation_steps = 1

dtype = torch.bfloat16
mixed_precision = "bf16"

# Параметры для диффузии
n_diffusion_steps = 40
samples_to_generate = 12
guidance_scale = 4.0 

# Папки для сохранения результатов
generated_folder = "samples"
os.makedirs(generated_folder, exist_ok=True)

# Настройка seed
current_date = datetime.now()
seed = int(current_date.strftime("%Y%m%d")) + 42
if fixed_seed:
    torch.manual_seed(seed)
    np.random.seed(seed)
    random.seed(seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(seed)

# 5. DDP kwargs
from accelerate import DistributedDataParallelKwargs
ddp_kwargs = DistributedDataParallelKwargs(bucket_cap_mb=10, find_unused_parameters=False)
accelerator = Accelerator(
    mixed_precision=mixed_precision,
    gradient_accumulation_steps=gradient_accumulation_steps,
    kwargs_handlers=[ddp_kwargs],
)
device = accelerator.device

print("init")
parser = argparse.ArgumentParser(description='Train a model on a dataset.')
parser.add_argument('--ds-path', type=str, default=ds_path, help='Path to the dataset')
parser.add_argument('--ep', type=int, default=num_epochs, help='Number of epochs to train the model')
parser.add_argument('--batch', type=int, default=batch_size, help='Total batch size')
parser.add_argument('--min-lr', type=float, default=min_learning_rate, help='Minimum learning rate')
parser.add_argument('--max-lr', type=float, default=base_learning_rate, help='Maximum learning rate')
parser.add_argument('--dry-run', action='store_true',default=False, help='Dry run train without saving/sampling')
parser.add_argument('--wandb', action='store_true', default=False, help='Enable wandb logging')
parser.add_argument('--lvl', type=float, default=0.0, help='Train level, from 0.5 to 5')

args = parser.parse_args()

batch_size = args.batch
ds_path = args.ds_path
base_learning_rate = args.max_lr
min_learning_rate = args.min_lr
num_epochs = args.ep
lvl = args.lvl
if args.wandb:
    use_wandb = True
if args.dry_run:
    save_model = False
    use_wandb = False
if lvl >= 0.1:
    base_learning_rate = base_learning_rate / lvl
    min_learning_rate = min_learning_rate / lvl
    print(f"max-lr:{base_learning_rate} min-lr:{min_learning_rate}")
    
# --------------------------- Инициализация WandB ---------------------------
if accelerator.is_main_process:
    if use_wandb:
        wandb.init(project=project, config={
            "batch_size": batch_size,
            "base_learning_rate": base_learning_rate,
            "num_epochs": num_epochs,
            "optimizer_type": optimizer_type,
        })
    if use_comet_ml:
        from comet_ml import Experiment
        comet_experiment = Experiment(
            api_key=comet_ml_api_key,
            project_name=project,
            workspace=comet_ml_workspace
        )
        hyper_params = {
            "batch_size": batch_size,
            "base_learning_rate": base_learning_rate,
            "num_epochs": num_epochs,
        }
        comet_experiment.log_parameters(hyper_params)

# --------------------------- Загрузка моделей ---------------------------
vae = AutoencoderKLQwenImage.from_pretrained("vae", torch_dtype=dtype).to("cpu").to(dtype=dtype).eval()
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("scheduler")
if scheduler is not None:
    scheduler.register_to_config(
        sigma_max=getattr(scheduler.config, "sigma_max", 80.0),
        sigma_min=getattr(scheduler.config, "sigma_min", 0.002),
        sigma_data=getattr(scheduler.config, "sigma_data", 1.0),
        final_sigmas_type=getattr(scheduler.config, "final_sigmas_type", "sigma_min"),
    )
tokenizer = None
text_encoder = None

def load_text_encoder():
    global tokenizer, text_encoder
    if tokenizer is None:
        tokenizer = Qwen3_5Tokenizer.from_pretrained("tokenizer")
    if text_encoder is None:
        text_encoder = Qwen3_5ForConditionalGeneration.from_pretrained(
            "text_encoder",
            torch_dtype=dtype
        ).to(device).eval()
        
load_text_encoder()

@torch.no_grad()
def encode_texts(text, max_length=max_length):
    if text is None:
        text = ""
    if isinstance(text, str):
        text = [text]

    formatted_prompts = []
    for t in text:
        messages = [{"role": "user", "content": [{"type": "text", "text": t}]}]
        formatted_prompts.append(
            tokenizer.apply_chat_template(
                messages,
                tokenize=False,
                add_generation_prompt=False  
            )
        )

    toks = tokenizer(
        formatted_prompts,
        padding="max_length",
        max_length=max_length,
        truncation=True,
        return_tensors="pt"
    ).to(device)

    outputs = text_encoder(
        input_ids=toks.input_ids,
        attention_mask=toks.attention_mask,
        output_hidden_states=True
    )

    hidden = outputs.hidden_states[-2].to(dtype=dtype)
    
    lengths = toks.attention_mask.sum(dim=1)
    for i, length in enumerate(lengths):
        hidden[i, length:] = 0

    return hidden, toks.attention_mask.to(dtype=torch.bool)
    
        
shift_factor = getattr(vae.config, "shift_factor", 0.0)
if shift_factor is None:
    shift_factor = 0.0

scaling_factor = getattr(vae.config, "scaling_factor", 1.0)
if scaling_factor is None:
    scaling_factor = 1.0
    
mean = getattr(vae.config, "latents_mean", None)
std = getattr(vae.config, "latents_std", None)
if mean is not None and std is not None:
    latents_std = torch.tensor(std, device=device, dtype=dtype).view(1, len(std), 1, 1)
    latents_mean = torch.tensor(mean, device=device, dtype=dtype).view(1, len(mean), 1, 1)
else:
    latents_std = None
    latents_mean = None


import numpy as np
from torch.utils.data import Sampler

class DistributedResolutionBatchSampler(Sampler):
    def __init__(self, dataset, batch_size, num_replicas, rank, drop_last=True, shuffle=True):
        self.dataset = dataset
        self.num_replicas = num_replicas
        self.rank = rank
        self.shuffle = shuffle
        self.drop_last = drop_last
        self.epoch = 0

        self.batch_size = max(1, batch_size // num_replicas)
        self.global_batch = self.batch_size * num_replicas

        try:
            widths = np.asarray(dataset["width"])
            heights = np.asarray(dataset["height"])
        except KeyError:
            widths = np.zeros(len(dataset))
            heights = np.zeros(len(dataset))

        groups = {}
        for i, (w, h) in enumerate(zip(widths, heights)):
            groups.setdefault((w, h), []).append(i)

        all_batches = []
        for indices in groups.values():
            idx = np.asarray(indices, dtype=np.int64)
            num_batches = len(idx) // self.global_batch
            if num_batches == 0:
                continue
            idx = idx[: num_batches * self.global_batch]
            batches = idx.reshape(num_batches, self.global_batch)
            all_batches.append(batches)

        if len(all_batches) > 0:
            self.global_batches = np.concatenate(all_batches, axis=0)
        else:
            self.global_batches = np.empty((0, self.global_batch), dtype=np.int64)

        self.num_batches = len(self.global_batches)

    def __iter__(self):
        rng = np.random.RandomState(self.epoch)
        order = np.arange(self.num_batches)

        if self.shuffle:
            rng.shuffle(order)

        start = self.rank * self.batch_size
        end = start + self.batch_size

        for i in order:
            yield self.global_batches[i][start:end]

    def __len__(self):
        return self.num_batches

    def set_epoch(self, epoch):
        self.epoch = epoch
        
def get_fixed_samples_by_resolution(dataset, samples_per_group=1, max_sample_groups=6):
    if disable_samples:
        return
    size_groups = defaultdict(list)
    try:
        widths = dataset["width"]
        heights = dataset["height"]
    except KeyError:
        widths = [0] * len(dataset)
        heights = [0] * len(dataset)
    for i, (w, h) in enumerate(zip(widths, heights)):
        size = (w, h)
        size_groups[size].append(i)

    # Не больше max_sample_groups разрешений
    if len(size_groups) > max_sample_groups:
        size_groups = dict(list(size_groups.items())[:max_sample_groups])
    
    fixed_samples = {}
    for size, indices in size_groups.items():
        n_samples = min(samples_per_group, len(indices))
        if len(size_groups)==1:
            n_samples = samples_to_generate
        if n_samples == 0:
            continue
        sample_indices = random.sample(indices, n_samples)
        samples_data = [dataset[idx] for idx in sample_indices]
        
        latents = torch.tensor(np.array([item["vae"] for item in samples_data])).to(device=device, dtype=dtype)
        
        if latents.ndim == 4:
            latents = latents.unsqueeze(2)
        elif latents.ndim == 6:
            latents = latents.squeeze(2)
            
        texts = [item["text"] for item in samples_data]
        
        if use_precomputed_embeddings:
            embeddings = torch.tensor(
                np.array([item["embeddings"] for item in samples_data]),
                device=device,
                dtype=dtype
            )
            masks = torch.tensor(
                np.array([item["attention_mask"] for item in samples_data]),
                device=device,
                dtype=torch.bool
            )
        else:
            embeddings, masks = encode_texts(texts,max_length)
        
        fixed_samples[size] = (latents, embeddings, masks, texts)
    
    print(f"Создано {len(fixed_samples)} групп фиксированных семплов по разрешениям")
    return fixed_samples

if limit > 0:
    dataset = load_from_disk(ds_path).select(range(limit))
else:
    print(">>> Поиск чанков датасета...")
    chunks = []
    
    # os.walk рекурсивно обходит абсолютно все подпапки
    for root, dirs, files in os.walk(ds_path):
        if "dataset_info.json" in files or "state.json" in files:
            chunks.append(root)
            dirs.clear() 

    if not chunks:
        print("❌ Чанки не найдены!")
        exit()

    print(f">>> Найдено чанков: {len(chunks)}. Начинаю загрузку и объединение...")

    ds_list = []
    for c in chunks:
        try:
            ds_list.append(load_from_disk(c))
        except Exception as e:
            print(f"⚠️ Ошибка загрузки чанка {c}: {e}")

    if ds_list:
        dataset = concatenate_datasets(ds_list)
        print("✅ Успешно объединено!")
    else:
        print("❌ Ни один чанк не удалось загрузить. Объединение невозможно.")

    # 3. Конкатенация (создает виртуальный объединенный датасет)
    #dataset = concatenate_datasets(ds_list)

print(f"images: {len(dataset)}")

def collate_fn(batch):
    latents = torch.from_numpy(
        np.array([item["vae"] for item in batch], dtype=np.float16)
    )
    
    if latents.ndim == 4:
        latents = latents.unsqueeze(2)
    elif latents.ndim == 6:
        latents = latents.squeeze(2)

    latents = latents.to(device, non_blocking=True)

    if use_precomputed_embeddings:
        embeddings = torch.from_numpy(
            np.array([item["embeddings"] for item in batch], dtype=np.float16)
        ).to(device, dtype=dtype)

        attention_mask = torch.from_numpy(
            np.array([item["attention_mask"] for item in batch], dtype=np.int64)
        ).to(device)

        return latents, embeddings, attention_mask

    raw_texts = [item["text"] for item in batch]

    texts = [
        "" if t.lower().startswith("zero") 
        else "" if random.random() < cfg_dropout
        else t[1:].lstrip() if t.startswith(".")
        else t.replace("*natcap* ", "").replace("*tags* ", "").replace("This image captures ","").strip()
        for t in raw_texts
    ]

    texts = [
        " ".join([w for w in t.split() if random.random() > world_dropout]) or t if t else t
        for t in texts
    ]

    embeddings, attention_mask = encode_texts(texts,max_length)
    attention_mask = attention_mask.to(dtype=torch.bool)

    return latents, embeddings, attention_mask

batch_sampler = DistributedResolutionBatchSampler(
        dataset=dataset,
        batch_size=batch_size,
        num_replicas=accelerator.num_processes,
        rank=accelerator.process_index,
        shuffle = shuffle
    )

dataloader = DataLoader(dataset, batch_sampler=batch_sampler, collate_fn=collate_fn)

if accelerator.is_main_process:
    print("Total samples", len(dataloader))
# НЕ передаём dataloader в accelerator.prepare:
# DistributedResolutionBatchSampler уже сам шардит глобальные батчи по ранкам.
# prepare добавлял бы второе шардирование (round-robin пропуск батчей), и каждый
# глобальный батч обрабатывался бы один раз срезом одного ранка -> за эпоху модель
# видела бы лишь 1/num_gpus датасета. Перенос тензоров на device — в collate_fn.

start_epoch = 0
global_step = 0
total_training_steps = (len(dataloader) * num_epochs)

latest_checkpoint = os.path.join(checkpoints_folder, project)
if os.path.isdir(latest_checkpoint):
    print("Загружаем Transformer из чекпоинта:", latest_checkpoint)
    transformer = CosmosTransformer3DModel.from_pretrained(latest_checkpoint).to(device=device, dtype=dtype)
    if transformer_gradient:
        transformer.enable_gradient_checkpointing()
else:
    raise FileNotFoundError(f"Transformer checkpoint not found at {latest_checkpoint}")

def create_optimizer(name, params):
    if name == "adam8bit":
        return bnb.optim.AdamW8bit(
            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.001
        )
    elif name == "adam":
        return torch.optim.AdamW(
            params, lr=base_learning_rate, betas=(0.9, betta2), eps=eps, weight_decay=0.01
        )
    elif name == "adafactor":
        return Adafactor(
            params,
            lr=base_learning_rate,
            eps=(1e-30, 1e-3),
            clip_threshold=1.0,
            decay_rate=-0.8,
            beta1=None,
            weight_decay=0.001,
            relative_step=False,
            scale_parameter=False,
            warmup_init=False
        )
    elif name == "babka_sotona":
        return BabkaSotona(
            params=params,
            lr=base_learning_rate,
            aurora_weight_decay=0.01,
            adafactor_weight_decay=0.001,
            adafactor_lr_scale=10.0,
            tail_fraction=0.3# При fbp=True мы скармливаем параметры поштучно, поэтому деление на хвост отключаем!
        )
    else:
        raise ValueError(f"Unknown optimizer: {name}")

if fbp:
    trainable_params = list(transformer.parameters())
    optimizer_dict = {p: create_optimizer(optimizer_type, [p]) for p in trainable_params}
    def optimizer_hook(param):
        optimizer_dict[param].step()
        optimizer_dict[param].zero_grad(set_to_none=True)
    for param in trainable_params:
        param.register_post_accumulate_grad_hook(optimizer_hook)
    transformer, optimizer = accelerator.prepare(transformer, optimizer_dict)
else:
    if not crossattn:
        transformer.requires_grad_(True)
    else:
        transformer.requires_grad_(False)
        trainable_params_names = ["attn2"]
        trainable_params = []
    
        print("--- РАЗМОРОЖЕННЫЕ СЛОИ ---")
        for name, param in transformer.named_parameters():
            if any(target in name for target in trainable_params_names):
                param.requires_grad_(True) 
                trainable_params.append(param)
                print(f"Обучаемый слой: {name}")
        print("--------------------------")
    
        if len(trainable_params) == 0:
            raise ValueError("Ошибка: ни один слой не был разморожен! Проверь ключи.")

    # Собираем параметры для оптимизатора
    params_to_opt = [p for p in transformer.parameters() if p.requires_grad]

    # --- Инициализация нового оптимизатора BabkaSotona ---
    if optimizer_type == "babka_sotona":
        optimizer = BabkaSotona(
            params=params_to_opt,
            lr=base_learning_rate,
            aurora_weight_decay=0.01,
            adafactor_weight_decay=0.001,
            adafactor_lr_scale=10.0,
            tail_fraction=0.3
        )
    else:
        optimizer = create_optimizer(optimizer_type, params_to_opt)
    
    def lr_schedule(step):
        x = step / total_training_steps
        warmup = warmup_percent
        if not use_decay:
            return base_learning_rate
        if x < warmup:
            return min_learning_rate + (base_learning_rate - min_learning_rate) * (x / warmup)
        decay_ratio = (x - warmup) / (1 - warmup)
        return min_learning_rate + 0.5 * (base_learning_rate - min_learning_rate) * \
               (1 + math.cos(math.pi * decay_ratio))

    lr_scheduler = LambdaLR(optimizer, lambda step: lr_schedule(step) / base_learning_rate)

if torch_compile:
    print("Compiling Transformer... Это займет несколько минут, не прерывайте!")
    transformer = torch.compile(transformer)
    print("Compiling - ok")

if not fbp:
    transformer, optimizer, lr_scheduler = accelerator.prepare(transformer, optimizer, lr_scheduler)

# Фиксированные семплы
fixed_samples = get_fixed_samples_by_resolution(dataset)

def get_negative_embedding(neg_prompt="", batch_size=1):
    if not neg_prompt:
        hidden_dim = 2048 
        seq_len = max_length
        empty_emb = torch.zeros((batch_size, seq_len, hidden_dim), dtype=dtype, device=device)
        empty_mask = torch.ones((batch_size, seq_len), dtype=torch.bool, device=device)
        return empty_emb, empty_mask

    uncond_emb, uncond_mask  = encode_texts([neg_prompt],max_length)
    uncond_emb = uncond_emb.to(dtype=dtype, device=device).repeat(batch_size, 1, 1)
    uncond_mask = uncond_mask.to(device=device).repeat(batch_size, 1)

    return uncond_emb, uncond_mask
    
if use_precomputed_embeddings:
    load_text_encoder()
    uncond_emb, uncond_mask = get_negative_embedding("low quality")
    uncond_emb = uncond_emb.to("cpu")
    uncond_mask = uncond_mask.to("cpu")
    del text_encoder
    torch.cuda.empty_cache()
    gc.collect()
    text_encoder = None
else:
    uncond_emb, uncond_mask = get_negative_embedding("low quality")

def pad_to_match(a, b, pad_value=0):
    Ta, Tb = a.shape[1], b.shape[1]
    if Ta == Tb:
        return a, b
    T = max(Ta, Tb)
    def pad(x, T_target):
        pad_len = T_target - x.shape[1]
        if pad_len <= 0:
            return x
        return torch.nn.functional.pad(x, (0, 0, 0, pad_len), value=pad_value)
    return pad(a, T), pad(b, T)


@torch.compiler.disable()
@torch.no_grad()
def generate_and_save_samples(fixed_samples_cpu, uncond_data, step):
    if disable_samples:
        return
    uncond_emb, uncond_mask = uncond_data
    uncond_emb = uncond_emb.to(device)
    uncond_mask = uncond_mask.to(device)
    
    original_model = None
    try:
        if not torch_compile:
            original_model = accelerator.unwrap_model(transformer, keep_torch_compile=True).eval()
        else:
            if optimizer_type == "babka_sotona":
                transformer.zero_grad(set_to_none=True)
            original_model = transformer.eval()

        vae.to(device=device).eval() 
        
        all_generated_images = []
        all_captions = [] 
        
        for size, (sample_latents, sample_text_embeddings, sample_mask, sample_text) in fixed_samples_cpu.items():
            width, height = size
            
            curr_batch_size = sample_latents.shape[0]
            in_channels = original_model.config.in_channels
            
            sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)

            # Используем scheduler (как в pipeline_sdxs) — t-сетка с shift из конфига
            scheduler.set_timesteps(n_diffusion_steps, device=device)
            timesteps = scheduler.timesteps

            latents = torch.randn(
                (curr_batch_size, in_channels, 1, sample_latents.shape[3], sample_latents.shape[4]),
                device=device,
                dtype=dtype,
                generator=torch.Generator(device=device).manual_seed(seed)
            )

            padding_mask = torch.zeros((1, 1, sample_latents.shape[3], sample_latents.shape[4]), device=device, dtype=dtype)

            if guidance_scale != 1:
                neg_emb_batch = uncond_emb[0:1].expand(curr_batch_size, -1, -1)
                neg_emb_batch, sample_text_embeddings = pad_to_match(neg_emb_batch, sample_text_embeddings)
                text_batch = torch.cat([neg_emb_batch, sample_text_embeddings], dim=0)
            else:
                text_batch = sample_text_embeddings

            for i in range(n_diffusion_steps):
                current_t = scheduler.sigmas[i]
                t_step = timesteps[i]
                t_val = float(current_t.item())

                timestep_tensor = torch.tensor([t_val], device=device, dtype=dtype).expand(curr_batch_size)
                latent_input = torch.cat([latents, latents], dim=0) if guidance_scale != 1 else latents
                t_input = torch.cat([timestep_tensor, timestep_tensor], dim=0) if guidance_scale != 1 else timestep_tensor

                model_output = original_model(
                    hidden_states=latent_input,
                    timestep=t_input,
                    encoder_hidden_states=text_batch,
                    padding_mask=padding_mask,
                    return_dict=False
                )[0]

                if guidance_scale != 1:
                    v_uncond, v_cond = model_output.chunk(2)
                    velocity = v_uncond.to(dtype) + guidance_scale * (v_cond.to(dtype) - v_uncond.to(dtype))
                else:
                    velocity = model_output.to(dtype)

                # Шедулер делает Euler: latents += (t_next - t) * velocity
                latents = scheduler.step(velocity, t_step, latents, return_dict=False)[0]
            
            current_latents = latents
            if step == 0:
                current_latents = sample_latents

            if latents_mean is not None and latents_std is not None:
                sigma_data = getattr(scheduler.config, "sigma_data", 1.0)
                l_mean = torch.tensor(vae.config.latents_mean).view(1, -1, 1, 1, 1).to(device, torch.float32)
                l_std = torch.tensor(vae.config.latents_std).view(1, -1, 1, 1, 1).to(device, torch.float32)
                
                latents_for_decode = (current_latents.float() * l_std) / sigma_data + l_mean
                latents_for_decode = latents_for_decode.to(dtype) 
            else:
                latents_for_decode = current_latents.to(dtype) 
                
            with torch.backends.cuda.sdp_kernel(enable_math=True, enable_flash=False, enable_mem_efficient=False):
                decoded = vae.decode(latents_for_decode).sample
                
            if decoded.ndim == 5:
                decoded = decoded[:, :, 0, :, :]
                
            decoded_fp32 = decoded.to(torch.float32)
            
            for img_idx, img_tensor in enumerate(decoded_fp32):
                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy()
                img = img.transpose(1, 2, 0)
    
                if np.isnan(img).any():
                    print("NaNs found, saving stopped! Step:", step)
                    img = np.nan_to_num(img, nan=0.0)
                pil_img = Image.fromarray((img * 255).astype("uint8"))
                
                max_w_overall = max(s[0] for s in fixed_samples_cpu.keys())
                max_h_overall = max(s[1] for s in fixed_samples_cpu.keys())
                max_w_overall = max(255, max_w_overall)
                max_h_overall = max(255, max_h_overall)
            
                padded_img = ImageOps.pad(pil_img, (max_w_overall, max_h_overall), color='white')
                all_generated_images.append(padded_img)

                caption_text = sample_text[img_idx][:300] if img_idx < len(sample_text) else ""
                all_captions.append(caption_text)
                
                sample_path = f"{generated_folder}/{project}_{width}x{height}_{img_idx}.jpg"
                pil_img.save(sample_path, "JPEG", quality=95)
        
        if use_wandb and accelerator.is_main_process:
            wandb_images = [
                wandb.Image(img, caption=f"{all_captions[i]}")
                for i, img in enumerate(all_generated_images)
            ]
            wandb.log({"generated_images": wandb_images})
        if use_comet_ml and accelerator.is_main_process:
            for i, img in enumerate(all_generated_images):
                comet_experiment.log_image(
                    image_data=img,
                    name=f"step_{step}_img_{i}",
                    step=step,
                    metadata={"caption": all_captions[i]}
                )
    finally:
        vae.to("cpu")
        uncond_emb = uncond_emb.to("cpu")
        uncond_mask = uncond_mask.to("cpu")
        try:
            all_generated_images.clear()
            all_captions.clear()
            del all_generated_images, all_captions
            del latents, current_latents, latent_model_input
            del decoded, decoded_fp32
            del sample_latents, sample_text_embeddings, sample_mask
            del model_pred_cond, model_pred_uncond
        except UnboundLocalError:
            pass
            
        torch.cuda.synchronize()
        torch.cuda.empty_cache()
        gc.collect()

if accelerator.is_main_process:
    if save_model:
        print("Генерация сэмплов до старта обучения...")
        generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), 0)
accelerator.wait_for_everyone()

def save_checkpoint(model_net, variant=""):
    if accelerator.is_main_process:
        model_to_save = None
        if not torch_compile:
            model_to_save = accelerator.unwrap_model(model_net)
        else:
            model_to_save = model_net

        if variant != "":
            model_to_save.to(dtype=torch.bfloat16).save_pretrained(
                os.path.join(checkpoints_folder, f"{project}"), variant=variant
            )
        else:
            model_to_save.save_pretrained(os.path.join(checkpoints_folder, f"{project}"))

        torch.cuda.synchronize()
        torch.cuda.empty_cache()
        gc.collect()

if accelerator.is_main_process:
    print(f"Total steps per GPU: {total_training_steps}")

epoch_loss_points = []
progress_bar = tqdm(total=total_training_steps, disable=not accelerator.is_local_main_process, desc="Training", unit="step")
monitor = TrainMonitor(log_every=25, csv_path="monitor_log.csv", warmup_steps=0)

steps_per_epoch = len(dataloader)
sink_interval = max(1, steps_per_epoch // sink_interval_share)
min_loss = 4.
last_sample_time = time.time() 
sample_interval_seconds = sample_interval_min * 60 

for epoch in range(start_epoch, start_epoch + num_epochs):
    batch_losses = []
    batch_grads = []
    batch_sampler.set_epoch(epoch)
    accelerator.wait_for_everyone()
    transformer.train()
    
    for step, (latents, embeddings, attention_mask) in enumerate(dataloader):
        
        if save_model == False and epoch == 0 and step == 5 :
            used_gb = torch.cuda.max_memory_allocated() / 1024**3
            print(f"Шаг {step}: {used_gb:.2f} GB")

        amp_context = accelerator.autocast() if torch_compile else nullcontext()
        with accelerator.accumulate(transformer):
            with amp_context:
                noise = torch.randn_like(latents, dtype=latents.dtype)
    
                t = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
                
                noisy_latents_5d = (1.0 - t.view(-1, 1, 1, 1, 1)) * latents + t.view(-1, 1, 1, 1, 1) * noise
                
                target = noise - latents
                
                padding_mask = torch.zeros((1, 1, latents.shape[3], latents.shape[4]), device=device, dtype=dtype)
                
                timestep_tensor = t.flatten().to(dtype)
    
                model_pred = transformer(
                    hidden_states= noisy_latents_5d, 
                    timestep=timestep_tensor, 
                    encoder_hidden_states=embeddings,
                    padding_mask=padding_mask,
                    return_dict=False
                )[0]
                
                mse_loss = F.mse_loss(model_pred.float(), target.float())
                    
                batch_losses.append(mse_loss.detach().item())
    
                if (global_step % 100 == 0) or (global_step % sink_interval == 0):
                    accelerator.wait_for_everyone()
    
                losses_dict = {}
                losses_dict["mse"] = mse_loss
    
                if (global_step % 100 == 0) or (global_step % sink_interval == 0):
                    accelerator.wait_for_everyone()
    
                accelerator.backward(mse_loss)
    
                if (global_step % 100 == 0) or (global_step % sink_interval == 0):
                    accelerator.wait_for_everyone()
                    
                grad = 0.0
                if not fbp:
                    if accelerator.sync_gradients:
                        grad_val = accelerator.clip_grad_norm_(transformer.parameters(), clip_grad_norm)
                        grad = grad_val.float().item() if torch.is_tensor(grad_val) else float(grad_val)
                        
                        # --- МАКСИМАЛЬНО ЧИСТЫЙ ШАГ ОПТИМИЗАТОРА ---
                        optimizer.step()
                        lr_scheduler.step()
                        
                        if save_model==False and accelerator.is_main_process:
                            monitor.step(mse_loss, accelerator.unwrap_model(transformer), global_step)
                        
                        optimizer.zero_grad(set_to_none=True)
    
                if accelerator.sync_gradients:
                    global_step += 1
                    progress_bar.update(1)
                    if accelerator.is_main_process:
                        current_lr = lr_scheduler.get_last_lr()[0] if not fbp else base_learning_rate
                        batch_grads.append(grad)
        
                        log_data = {}
                        log_data["loss_mse"] = mse_loss.detach().item()
                        log_data["lr"] = current_lr
                        log_data["grad"] = grad
                        if accelerator.sync_gradients:
                            if use_wandb:
                                wandb.log(log_data, step=global_step)
                            if use_comet_ml:
                                comet_experiment.log_metrics(log_data, step=global_step)

                        current_time = time.time()
                        is_time_to_sample = (current_time - last_sample_time) >= sample_interval_seconds
                        if is_time_to_sample or global_step == 5:
                            if accelerator.is_main_process:
                                if save_model:
                                    generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), global_step)
                                elif epoch % 10 == 0:
                                    generate_and_save_samples(fixed_samples, (uncond_emb, uncond_mask), global_step)
                                    
                                last_n = sink_interval
                                if save_model:
                                    has_losses = len(batch_losses) > 0
                                    avg_sample_loss = np.mean(batch_losses[-sink_interval:]) if has_losses else 0.0
                                    last_loss = batch_losses[-1] if has_losses else 0.0
                                    max_loss = max(avg_sample_loss, last_loss)  
                                    should_save = max_loss < min_loss * save_barrier
                                    print(
                                        f"Saving: {should_save} | Max: {max_loss:.4f} | "
                                        f"Last: {last_loss:.4f} | Avg: {avg_sample_loss:.4f}"
                                    )
                                    if should_save:
                                        min_loss = max_loss
                                        save_checkpoint(transformer)
                                last_sample_time = current_time

                            transformer.train()

    if accelerator.is_main_process:
        avg_epoch_loss = np.mean(batch_losses) if len(batch_losses) > 0 else 0.0
        avg_epoch_grad = np.mean(batch_grads) if len(batch_grads) > 0 else 0.0

        print(f"\nЭпоха {epoch} завершена. Средний лосс: {avg_epoch_loss:.6f}")
        monitor.end_epoch(epoch, global_step)
        log_data_ep = {
                        "epoch_loss": avg_epoch_loss,
                        "epoch_grad": avg_epoch_grad,
                        "epoch": epoch + 1,
                    }
        if use_wandb:
            wandb.log(log_data_ep)
        if use_comet_ml:
            comet_experiment.log_metrics(log_data_ep)

if accelerator.is_main_process:
    print("Обучение завершено! Сохраняем финальную модель...")
    save_checkpoint(transformer,"bf16")
    if use_comet_ml:
        comet_experiment.end()
if accelerator.is_main_process:
    monitor.summary()

accelerator.free_memory()
if torch.distributed.is_initialized():
    torch.distributed.destroy_process_group()
    
print("Готово!")