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 UNet2DConditionModel, 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
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
from torch.nn.attention import sdpa_kernel, SDPBackend

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

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/ds1234_noanime_704_vae8x16x" #alchemist_704_vae8x16x_imgpool"
ds_path = "/root/sdxs-2b/datasets/ds12345_640_vae_qwen" 
project = "unet"

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 = 4e-5
min_learning_rate = 4e-6

# 0.5 - pretrain (base forms)
# 1 - base train (composition)
# 3 - finetuning (anatomy)
# 5 - small details (faces)
learning_rate_scale = 3
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 = 20
sample_interval_min = 60
cfg_dropout = -0.10
# Время t, bias = -0.5 (Фокус на Деталях ~300) bias = 0.5 (Фокус на структуре) bias = 0 (колокол/ равномерно)
sigmoid_bias = -0.1
max_length = 250
max_snr_gamma = 5.0
use_precomputed_embeddings = False
use_wandb = False
use_comet_ml = True
save_model = True
use_decay = True
fbp = False
torch_compile = False
unet_gradient = True
loss_normalize = False
fixed_seed = False
shuffle = True
optimizer_type = "adafactor" #"adam8bit"
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" # hardcoded for blind run, i don't care about key
comet_ml_workspace = "recoilme" 
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
# MAX_JOBS=4 pip install flash-attn --no-build-isolation
#torch.backends.cuda.enable_flash_sdp(True)
#torch.backends.cuda.enable_mem_efficient_sdp(True)
#torch.backends.cuda.enable_math_sdp(False) # Отключаем медленный вариант
save_barrier = 1.25
warmup_percent = 0.0025
betta2 = 0.997
eps = 1e-6
clip_grad_norm = 1.0
limit = 0
checkpoints_folder = ""
gradient_accumulation_steps = 1
dtype = torch.float32
mixed_precision = "bf16"

# Параметры для диффузии
n_diffusion_steps = 40
samples_to_generate = 12
guidance_scale = 4

# 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)

accelerator = Accelerator(
    mixed_precision=mixed_precision,
    gradient_accumulation_steps=gradient_accumulation_steps
)
device = accelerator.device

print("init")
# Создаём объект ArgumentParser с рассчитанными значениями по умолчанию
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('--lvl', type=float, default=0.0, help='Train level, from 0.5 to 5')

# Парсим аргументы командной строки
args = parser.parse_args()

# Используем значения из аргументов
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.dry_run:
    save_model = 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(device).eval()
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("scheduler")
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=torch.float16
        ).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   # 🔥 фикс
            )
        )

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

    return hidden.to(dtype=dtype), toks.attention_mask.to(dtype=torch.bool)
    
@torch.no_grad()
def encode_texts_fast(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, add_generation_prompt=False, tokenize=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)
    
    last_hidden = outputs.hidden_states[-2]

    return last_hidden.to(dtype=dtype), 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)

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

        # batch на одну GPU
        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))

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

        # --- создаём список всех глобальных батчей ---
        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):
    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)
    
    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 == 5:
            latents = latents.squeeze(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.int64
            )
        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:
    dataset = load_from_disk(ds_path)


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

def collate_fn_simple(batch):
    latents = torch.from_numpy(
        np.array([item["vae"] for item in batch], dtype=np.float16)
    ).to(device, dtype=dtype)
    
    if latents.ndim == 5:
        latents = latents.squeeze(2)
    #elif latents.ndim == 6:
    #    latents = latents.squeeze(2)

    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("The image shows ", "").replace("The image is ", "").replace("This image captures ","").strip()
        for t in raw_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_simple)

if accelerator.is_main_process:
    print("Total samples", len(dataloader))
dataloader = accelerator.prepare(dataloader)

start_epoch = 0
global_step = 0
total_training_steps = (len(dataloader) * num_epochs)
world_size = accelerator.state.num_processes

# Загрузка UNet
latest_checkpoint = os.path.join(checkpoints_folder, project)
if os.path.isdir(latest_checkpoint):
    print("Загружаем UNet из чекпоинта:", latest_checkpoint)
    unet = UNet2DConditionModel.from_pretrained(latest_checkpoint).to(device=device, dtype=dtype)
    if unet_gradient:
        unet.enable_gradient_checkpointing()
    #unet.set_use_memory_efficient_attention_xformers(False)
    print(dir(SDPBackend))
    try:
        unet.set_attn_processor(AttnProcessor2_0())
    except Exception as e:
        print(f"Ошибка при включении SDPA: {e}")
        unet.set_use_memory_efficient_attention_xformers(True)
else:
    raise FileNotFoundError(f"UNet 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=0.001
        )
    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 == "muon_adam8bit":
        return MuonAdamW8bit(
            params, 
            lr=base_learning_rate, 
            betas=(0.9, betta2), 
            eps=eps, 
            weight_decay=0.01,
            muon_lr_mult=muon_lr_scale,
        )
    else:
        raise ValueError(f"Unknown optimizer: {name}")

if fbp:
    trainable_params = list(unet.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)
    unet, optimizer = accelerator.prepare(unet, optimizer_dict)
else:
    unet.requires_grad_(True)
    optimizer = create_optimizer(optimizer_type, unet.parameters())
    # 1. Сначала замораживаем ВСЕ параметры UNet
    #unet.requires_grad_(False)
    
    # 2. Размораживаем только нужные
    #trainable_params_names = ["conv_in.weight", "conv_in.bias", "conv_out.weight", "conv_out.bias"]
    #train_params = []
    
    #for name, param in unet.named_parameters():
    #    if any(target in name for target in trainable_params_names):
    #        param.requires_grad = True
    #        train_params.append(param)
    #        print(f"Обучаемый слой: {name}")
    
    def lr_schedule(step):
        x = step / (total_training_steps * world_size)
        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 UNet... Это займет несколько минут, не прерывайте!")
    unet = torch.compile(unet)
    print("Compiling - ok")

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

# Фиксированные семплы
fixed_samples = get_fixed_samples_by_resolution(dataset)

# --- [UPDATED] Функция для негативного эмбеддинга (возвращает 3 элемента) ---
def get_negative_embedding(neg_prompt="", batch_size=1):
    if not neg_prompt:
        hidden_dim = 1024 
        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.int64, 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
    
# Получаем негативные (пустые) условия для валидации
if use_precomputed_embeddings:
    # 1. грузим encoder ВРЕМЕННО
    load_text_encoder()

    # 2. считаем negative
    uncond_emb, uncond_mask = get_negative_embedding("low quality")

    # 3. уносим на CPU (очень важно)
    uncond_emb = uncond_emb.to("cpu")
    uncond_mask = uncond_mask.to("cpu")

    # 4. выгружаем encoder с GPU
    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):
    # a, b: [B, T, D]
    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)

def pad_mask(a, b):
    Ta, Tb = a.shape[1], b.shape[1]
    T = max(Ta, Tb)

    def pad(x):
        pad_len = T - x.shape[1]
        if pad_len <= 0:
            return x
        return torch.nn.functional.pad(x, (0, pad_len), value=0)

    return pad(a), pad(b)
    
# --- Функция генерации семплов  ---
@torch.compiler.disable()
@torch.no_grad()
def generate_and_save_samples(fixed_samples_cpu, uncond_data, step):
    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(unet, keep_torch_compile=True).eval()
        else:
            original_model = unet.eval()

        vae.to(device=device).eval() 
        
        all_generated_images = []
        all_captions = [] 
        
        # Распаковываем 5 элементов (добавились mask)
        for size, (sample_latents, sample_text_embeddings, sample_mask, sample_text) in fixed_samples_cpu.items():
            width, height = size
            sample_latents = sample_latents.to(dtype=dtype, device=device)
            sample_text_embeddings = sample_text_embeddings.to(dtype=dtype, device=device)
            sample_mask = sample_mask.to(device=device)
        
            latents = torch.randn(
                sample_latents.shape,
                device=device,
                dtype=sample_latents.dtype,
                generator=torch.Generator(device=device).manual_seed(seed)
            )
        
            scheduler.set_timesteps(n_diffusion_steps, device=device)

            for t in scheduler.timesteps:
                if guidance_scale != 1:
                    latent_model_input = torch.cat([latents, latents], dim=0)
                    
                    curr_batch_size = sample_text_embeddings.shape[0]
                    seq_len = sample_text_embeddings.shape[1]
                    hidden_dim = sample_text_embeddings.shape[2]

                    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)
                    neg_mask_batch = uncond_mask[0:1].expand(curr_batch_size, -1)
                    neg_mask_batch, sample_mask = pad_mask(neg_mask_batch, sample_mask)

                    text_embeddings_batch = torch.cat([neg_emb_batch, sample_text_embeddings], dim=0)
                    attention_mask_batch = torch.cat([neg_mask_batch, sample_mask], dim=0)
                    
                    #neg_emb_batch = uncond_emb[0:1].expand(curr_batch_size, -1, -1)
                    #text_embeddings_batch = torch.cat([neg_emb_batch, sample_text_embeddings], dim=0)
                    
                    #neg_mask_batch = uncond_mask[0:1].expand(curr_batch_size, -1)
                    #attention_mask_batch = torch.cat([neg_mask_batch, sample_mask], dim=0)
                    
                else:
                    latent_model_input = latents
                    text_embeddings_batch = sample_text_embeddings
                    attention_mask_batch = sample_mask
            
                # Теперь всё имеет одинаковый batch size
                model_out = original_model(
                    latent_model_input,
                    t,
                    encoder_hidden_states=text_embeddings_batch,
                    encoder_attention_mask=attention_mask_batch,
                )

                flow = getattr(model_out, "sample", model_out)
        
                if guidance_scale != 1:
                    flow_uncond, flow_cond = flow.chunk(2)
                    flow = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
        
                latents = scheduler.step(flow, t, latents).prev_sample
        
            current_latents = latents
            if step==0:
                current_latents = sample_latents

            # VAE Qwen ожидает 5D, добавляем 1 кадр времени
            vae_input = current_latents.unsqueeze(2).to(torch.float32) 
            if latents_mean is not None and latents_std is not None:
                 vae_input = vae_input * latents_std.unsqueeze(2) + latents_mean.unsqueeze(2)
            
            decoded = vae.decode(vae_input).sample
            # После декодирования у Qwen на выходе [B, C, 1, H, W], убираем 1
            if decoded.ndim == 5:
                decoded = decoded.squeeze(2) 
            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)
                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, flow
            del decoded, decoded_fp32
            del sample_latents, sample_text_embeddings, sample_mask # Копии на GPU
            del model_out
        except UnboundLocalError:
            pass
            
        # 3. Синхронизируем CUDA перед очисткой
        torch.cuda.synchronize()
        # 4. Теперь чистим кэш аллокатора и вызываем GC
        torch.cuda.empty_cache()
        gc.collect()
        
# --------------------------- Генерация сэмплов перед обучением ---------------------------
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(unet, variant=""):
    if accelerator.is_main_process:
        model_to_save = None
        if not torch_compile:
            model_to_save = accelerator.unwrap_model(unet)
        else:
            model_to_save = unet

        if variant != "":
            model_to_save.to(dtype=torch.float16).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()

# --------------------------- Тренировочный цикл ---------------------------
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")

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  # 60 минут

for epoch in range(start_epoch, start_epoch + num_epochs):
    batch_losses = []
    batch_grads = []
    batch_sampler.set_epoch(epoch)
    accelerator.wait_for_everyone()
    unet.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(unet):
            with amp_context:
                # шум
                noise = torch.randn_like(latents, dtype=latents.dtype)
    
                t = torch.sigmoid(torch.randn(latents.shape[0], device=latents.device, dtype=latents.dtype) + sigmoid_bias)
                
                # интерполяция между x0 и шумом
                noisy_latents = (1.0 - t.view(-1, 1, 1, 1)) * latents + t.view(-1, 1, 1, 1) * noise
                # делаем integer timesteps для UNet
                timesteps = t.to(torch.float32).mul(999.0)
    
                # --- Вызов UNet с маской  ---
                #with sdpa_kernel([SDPBackend.FLASH_ATTENTION, SDPBackend.EFFICIENT_ATTENTION], SDPBackend.CUDNN_ATTENTION):
                with sdpa_kernel([
                    SDPBackend.FLASH_ATTENTION, 
                    SDPBackend.EFFICIENT_ATTENTION, 
                    SDPBackend.CUDNN_ATTENTION, 
                    SDPBackend.MATH
                ]):
                    model_pred = unet(
                        noisy_latents, 
                        timesteps, 
                        encoder_hidden_states=embeddings,
                        encoder_attention_mask=attention_mask,
                    ).sample
                
                target = noise - latents

                if max_snr_gamma > 0:
                    # 1. Считаем сырой лосс попиксельно (без усреднения по всему батчу)
                    raw_loss = F.mse_loss(model_pred.float(), target.float(), reduction='none')
                    # Усредняем ошибку внутри каждой картинки, чтобы получить вектор лоссов [Batch_Size]
                    loss_per_sample = raw_loss.mean(dim=[1, 2, 3]) 
                    
                    # 2. Считаем SNR (Signal-to-Noise Ratio) для текущих таймстепов батча
                    # Сигнал — это (1 - t), шум — это t. 
                    snr = ((1.0 - t) / (t + 1e-5)) ** 2

                    # Для твоего формата предсказания (v-prediction) формула веса: min(snr, gamma) / (snr + 1)
                    min_snr_weights = torch.clamp(snr, max=max_snr_gamma) / (snr + 1.0)
                    
                    # 4. Применяем веса к лоссам картинок и получаем итоговый скаляр для backward()
                    # (view(-1) гарантирует, что размерности совпадут)
                    mse_loss = (loss_per_sample * min_snr_weights.view(-1)).mean()
                else:
                    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_(unet.parameters(), clip_grad_norm)
                        grad = grad_val.float().item() if torch.is_tensor(grad_val) else float(grad_val)
                        optimizer.step()
                        lr_scheduler.step()
                        optimizer.zero_grad(set_to_none=True)
    
                if accelerator.sync_gradients:
                    global_step += 1
                    progress_bar.update(1)
                    if accelerator.is_main_process:
                        if fbp:
                            current_lr = base_learning_rate
                        else:
                            current_lr = lr_scheduler.get_last_lr()[0]
                        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 == 50:
                            # Передаем tuple (emb, mask) для негатива
                            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}"
                                )
                                # 6. Сохранение и обновление
                                if should_save:
                                    min_loss = max_loss
                                    save_checkpoint(unet)
                            last_sample_time = current_time
                            unet.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}")
        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...")
    #if save_model:
    save_checkpoint(unet,"fp16")
    if use_comet_ml:
        comet_experiment.end()
accelerator.free_memory()
if torch.distributed.is_initialized():
    torch.distributed.destroy_process_group()
    
print("Готово!")