src/train_no_pooling.py

import os
import math
import torch
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import DataLoader, Sampler
from torch.utils.data.distributed import DistributedSampler
from torch.optim.lr_scheduler import LambdaLR
from collections import defaultdict
from torch.optim.lr_scheduler import LambdaLR
from diffusers import UNet2DConditionModel, AutoencoderKLWan,AutoencoderKL
from accelerate import Accelerator
from datasets import load_from_disk
from tqdm import tqdm
from PIL import Image,ImageOps
import wandb
import random
import gc
from accelerate.state import DistributedType
from torch.distributed import broadcast_object_list
from torch.utils.checkpoint import checkpoint
from diffusers.models.attention_processor import AttnProcessor2_0
from datetime import datetime
import bitsandbytes as bnb
import torch.nn.functional as F
from collections import deque

# --------------------------- Параметры ---------------------------
ds_path = "/workspace/sdxs/datasets/ds1234_640"
project = "unet"
batch_size = 64
base_learning_rate = 6e-5
min_learning_rate = 2.5e-5
num_epochs = 80
# samples/save per epoch
sample_interval_share = 2
use_wandb = True
use_comet_ml = False
save_model = True
use_decay = True
fbp = False # fused backward pass
optimizer_type = "adam8bit"
torch_compile = False
unet_gradient = True
clip_sample = False #Scheduler
fixed_seed = False
shuffle = True
comet_ml_api_key = "Agctp26mbqnoYrrlvQuKSTk6r"  # Добавлен API ключ для Comet ML
comet_ml_workspace = "recoilme"  # Добавлен workspace для Comet ML
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cuda.enable_mem_efficient_sdp(False)
dtype = torch.float32
save_barrier = 1.006
warmup_percent = 0.01
percentile_clipping = 99 # 8bit optim
betta2 = 0.99
eps = 1e-8
clip_grad_norm = 1.0
steps_offset = 0 # Scheduler
limit = 0
checkpoints_folder = ""
mixed_precision = "no" #"fp16"
gradient_accumulation_steps = 1
accelerator = Accelerator(
    mixed_precision=mixed_precision,
    gradient_accumulation_steps=gradient_accumulation_steps
)
device = accelerator.device

# Параметры для диффузии
n_diffusion_steps = 50
samples_to_generate = 12
guidance_scale = 4

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

# Настройка seed для воспроизводимости
current_date = datetime.now()
seed = int(current_date.strftime("%Y%m%d"))
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)

# --------------------------- Параметры LoRA ---------------------------
lora_name = "" 
lora_rank = 32
lora_alpha = 64

print("init")

# --------------------------- Инициализация WandB ---------------------------
if accelerator.is_main_process:
    if use_wandb:
        wandb.init(project=project+lora_name, config={
            "batch_size": batch_size,
            "base_learning_rate": base_learning_rate,
            "num_epochs": num_epochs,
            "fbp": fbp,
            "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
        )
        # Логируем гиперпараметры в Comet ML
        hyper_params = {
            "batch_size": batch_size,
            "base_learning_rate": base_learning_rate,
            "min_learning_rate": min_learning_rate,
            "num_epochs": num_epochs,
            "n_diffusion_steps": n_diffusion_steps,
            "guidance_scale": guidance_scale,
            "optimizer_type": optimizer_type,
            "mixed_precision": mixed_precision,
        }
        comet_experiment.log_parameters(hyper_params)

# Включение Flash Attention 2/SDPA
torch.backends.cuda.enable_flash_sdp(True)
# --------------------------- Инициализация Accelerator --------------------
gen = torch.Generator(device=device)
gen.manual_seed(seed)

# --------------------------- Загрузка моделей ---------------------------
# VAE загружается на CPU для экономии GPU-памяти (как в твоём оригинальном коде)
vae = AutoencoderKL.from_pretrained("AiArtLab/simplevae", subfolder="vae", torch_dtype=dtype).to("cpu").eval()

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
    
latents_mean = getattr(vae.config, "latents_mean", None)
latents_std = getattr(vae.config, "latents_std", None)

from diffusers import FlowMatchEulerDiscreteScheduler

# Подстрой под свои параметры
num_train_timesteps = 1000

scheduler = FlowMatchEulerDiscreteScheduler(
    num_train_timesteps=num_train_timesteps,
    #shift=3.0,               # пример; подбирается при необходимости
    #use_dynamic_shifting=True
)


class DistributedResolutionBatchSampler(Sampler):
    def __init__(self, dataset, batch_size, num_replicas, rank, shuffle=True, drop_last=True):
        self.dataset = dataset
        self.batch_size = max(1, batch_size // num_replicas)
        self.num_replicas = num_replicas
        self.rank = rank
        self.shuffle = shuffle
        self.drop_last = drop_last
        self.epoch = 0
        
        try:
            widths = np.array(dataset["width"])
            heights = np.array(dataset["height"])
        except KeyError:
            widths = np.zeros(len(dataset))
            heights = np.zeros(len(dataset))
        
        self.size_keys = np.unique(np.stack([widths, heights], axis=1), axis=0)
        self.size_groups = {}
        for w, h in self.size_keys:
            mask = (widths == w) & (heights == h)
            self.size_groups[(w, h)] = np.where(mask)[0]
            
        self.group_num_batches = {}
        total_batches = 0
        for size, indices in self.size_groups.items():
            num_full_batches = len(indices) // (self.batch_size * self.num_replicas)
            self.group_num_batches[size] = num_full_batches
            total_batches += num_full_batches
            
        self.num_batches = (total_batches // self.num_replicas) * self.num_replicas
        
    def __iter__(self):
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        all_batches = []
        rng = np.random.RandomState(self.epoch)
        
        for size, indices in self.size_groups.items():
            indices = indices.copy()
            if self.shuffle:
                rng.shuffle(indices)
            num_full_batches = self.group_num_batches[size]
            if num_full_batches == 0:
                continue
            valid_indices = indices[:num_full_batches * self.batch_size * self.num_replicas]
            batches = valid_indices.reshape(-1, self.batch_size * self.num_replicas)
            start_idx = self.rank * self.batch_size
            end_idx = start_idx + self.batch_size
            gpu_batches = batches[:, start_idx:end_idx]
            all_batches.extend(gpu_batches)
        
        if self.shuffle:
            rng.shuffle(all_batches)
        accelerator.wait_for_everyone()
        return iter(all_batches)

    def __len__(self):
        return self.num_batches 

    def set_epoch(self, epoch):
        self.epoch = epoch

# Функция для выборки фиксированных семплов по размерам
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)
        embeddings = torch.tensor(np.array([item["embeddings"] for item in samples_data])).to(device,dtype=dtype)
        texts = [item["text"] for item in samples_data]
        fixed_samples[size] = (latents, embeddings, 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)

def collate_fn_simple(batch):
    latents = torch.tensor(np.array([item["vae"] for item in batch])).to(device,dtype=dtype)
    embeddings = torch.tensor(np.array([item["embeddings"] for item in batch])).to(device,dtype=dtype)
    attention_mask = torch.abs(embeddings).sum(dim=-1) > 1e-6
    attention_mask = attention_mask.to(device, dtype=torch.int64)
    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)
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

# Опция загрузки модели из последнего чекпоинта (если существует)
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)
    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:
    # FIX: если чекпоинта нет — прекращаем с понятной ошибкой (лучше, чем неожиданные NameError дальше)
    raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}. Положи UNet чекпоинт в {latest_checkpoint} или укажи другой путь.")

if lora_name:
    print(f"--- Настройка LoRA через PEFT (Rank={lora_rank}, Alpha={lora_alpha}) ---")
    from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
    from peft.tuners.lora import LoraModel
    import os
    unet.requires_grad_(False)
    print("Параметры базового UNet заморожены.")

    lora_config = LoraConfig(
        r=lora_rank,
        lora_alpha=lora_alpha,
        target_modules=["to_q", "to_k", "to_v", "to_out.0"],
    )
    unet.add_adapter(lora_config)

    from peft import get_peft_model
    peft_unet = get_peft_model(unet, lora_config)
    params_to_optimize = list(p for p in peft_unet.parameters() if p.requires_grad)

    if accelerator.is_main_process:
        lora_params_count = sum(p.numel() for p in params_to_optimize)
        total_params_count = sum(p.numel() for p in unet.parameters())
        print(f"Количество обучаемых параметров (LoRA): {lora_params_count:,}")
        print(f"Общее количество параметров UNet: {total_params_count:,}")

    lora_save_path = os.path.join("lora", lora_name)
    os.makedirs(lora_save_path, exist_ok=True)

    def save_lora_checkpoint(model):
        if accelerator.is_main_process:
            print(f"Сохраняем LoRA адаптеры  в {lora_save_path}")
            from peft.utils.save_and_load import get_peft_model_state_dict
            lora_state_dict = get_peft_model_state_dict(model)
            torch.save(lora_state_dict, os.path.join(lora_save_path, "adapter_model.bin"))
            model.peft_config["default"].save_pretrained(lora_save_path)
            from diffusers import StableDiffusionXLPipeline
            StableDiffusionXLPipeline.save_lora_weights(lora_save_path, lora_state_dict)

# --------------------------- Оптимизатор ---------------------------
if lora_name:
    trainable_params = [p for p in unet.parameters() if p.requires_grad]
else:
    if fbp:
        trainable_params = list(unet.parameters())

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=0.01,
            percentile_clipping=percentile_clipping
        )
    elif name == "adam":
        return torch.optim.AdamW(
            params, lr=base_learning_rate, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.01
        )
    else:
        raise ValueError(f"Unknown optimizer: {name}")

if fbp:
    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:
    optimizer = create_optimizer(optimizer_type, unet.parameters())
    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)

    num_params = sum(p.numel() for p in unet.parameters())
    print(f"[rank {accelerator.process_index}] total params: {num_params}")
    for name, param in unet.named_parameters():
        if torch.isnan(param).any() or torch.isinf(param).any():
            print(f"[rank {accelerator.process_index}] NaN/Inf in {name}")
    unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)

if torch_compile:
    print("compiling")
    torch.set_float32_matmul_precision('high')
    torch.backends.cudnn.allow_tf32 = True
    torch.backends.cuda.matmul.allow_tf32 = True
    unet = torch.compile(unet)#, mode='max-autotune')
    print("compiling - ok")

# --------------------------- Фиксированные семплы для генерации ---------------------------
fixed_samples = get_fixed_samples_by_resolution(dataset)

def get_negative_embedding(neg_prompt="", batch_size=1):
    """
    Возвращает эмбеддинг негативного промпта с батчем.
    Загружает модели, вычисляет эмбеддинг, выгружает модели на CPU.
    """
    import torch
    from transformers import AutoTokenizer, AutoModel

    # Настройки
    dtype = torch.float16
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # Загрузка моделей (если ещё не загружены)
    if not hasattr(get_negative_embedding, "tokenizer"):
        get_negative_embedding.tokenizer = AutoTokenizer.from_pretrained(
            "Qwen/Qwen3-0.6B"
        )
        get_negative_embedding.text_model = AutoModel.from_pretrained(
            "Qwen/Qwen3-0.6B"
        ).to(device).eval()

    # Вычисление эмбеддинга
    def encode_texts(texts, max_length=150):
        with torch.inference_mode():
            toks = get_negative_embedding.tokenizer(
                texts, return_tensors="pt", padding="max_length", truncation=True, max_length=max_length
            ).to(device)

            outs = get_negative_embedding.text_model(**toks, output_hidden_states=True, return_dict=True)
            hidden = outs.hidden_states[-1]  # [B, L, D]
            mask = toks["attention_mask"].unsqueeze(-1)  # (B, L, 1)
            hidden = hidden * mask   
            
            return hidden

    # Возвращаем эмбеддинг
    if not neg_prompt:
        hidden_dim = 1024  # Размерность эмбеддинга Qwen3-Embedding-0.6B
        seq_len = 150
        return torch.zeros((batch_size, seq_len, hidden_dim), dtype=dtype, device=device)

    uncond_emb = encode_texts([neg_prompt]).to(dtype=dtype, device=device)
    uncond_emb = uncond_emb.repeat(batch_size, 1, 1)  # Добавляем батч

    # Выгружаем модели
    if 1:
        if hasattr(get_negative_embedding, "text_model"):
            get_negative_embedding.text_model = get_negative_embedding.text_model.to("cpu")
        if hasattr(get_negative_embedding, "tokenizer"):
            del get_negative_embedding.tokenizer  # Освобождаем память
        torch.cuda.empty_cache()

    return uncond_emb
    
uncond_emb = get_negative_embedding("low quality")

@torch.compiler.disable()
@torch.no_grad()
def generate_and_save_samples(fixed_samples_cpu,empty_embeddings, step):
    original_model = None
    try:
        # безопасный unwrap: если компилировано, unwrap не нужен
        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()  # временно подгружаем VAE на GPU для декодинга

        
        all_generated_images = []
        all_captions = [] 
        
        for size, (sample_latents, sample_text_embeddings, 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)
        
            # начальный шум
            latents = torch.randn(
                sample_latents.shape,
                device=device,
                dtype=sample_latents.dtype,
                generator=torch.Generator(device=device).manual_seed(seed)
            )
        
            # подготовим timesteps через шедулер
            scheduler.set_timesteps(n_diffusion_steps, device=device)
            prompt_mask = torch.abs(sample_text_embeddings).sum(dim=-1) > 1e-6 # (B, Seq)
            prompt_mask = prompt_mask.to(dtype=torch.int64)
            
            # Создаем маску для негатива (empty_embeddings)
            # empty_embeddings у вас [Batch, Seq, Dim], скорее всего там нули кроме первых токенов
            neg_mask = torch.abs(empty_embeddings).sum(dim=-1) > 1e-6
            neg_mask = neg_mask.repeat(sample_text_embeddings.shape[0], 1).to(dtype=torch.int64, device=device)
        
            for t in scheduler.timesteps:
                # guidance: удваиваем батч
                if guidance_scale != 1:
                    latent_model_input = torch.cat([latents, latents], dim=0)
                
                    # empty_embeddings: [1, 1, hidden_dim] → повторяем по seq_len и batch
                    seq_len = sample_text_embeddings.shape[1]
                    hidden_dim = sample_text_embeddings.shape[2]
                    empty_embeddings_exp = empty_embeddings.expand(-1, seq_len, hidden_dim)  # [1, seq_len, hidden_dim]
                    empty_embeddings_exp = empty_embeddings_exp.repeat(sample_text_embeddings.shape[0], 1, 1)  # [batch, seq_len, hidden_dim]
                
                    text_embeddings_batch = torch.cat([empty_embeddings_exp, sample_text_embeddings], dim=0)
                    attention_mask_batch = torch.cat([neg_mask, prompt_mask], dim=0)
                else:
                    latent_model_input = latents
                    text_embeddings_batch = sample_text_embeddings
                    attention_mask_batch = prompt_mask

        
                # предсказание потока (velocity)
                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)
        
                # guidance объединение
                if guidance_scale != 1:
                    flow_uncond, flow_cond = flow.chunk(2)
                    flow = flow_uncond + guidance_scale * (flow_cond - flow_uncond)
        
                # шаг через scheduler
                latents = scheduler.step(flow, t, latents).prev_sample
        
            current_latents = latents


            # Параметры нормализации
            latent_for_vae = current_latents.detach() / scaling_factor + shift_factor
            decoded = vae.decode(latent_for_vae.to(torch.float32)).sample
            
            decoded_fp32 = decoded.to(torch.float32)
            for img_idx, img_tensor in enumerate(decoded_fp32):

                # Форма: [3, H, W] -> преобразуем в [H, W, 3]
                img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy()
                img = img.transpose(1, 2, 0)  # Из [3, H, W] в [H, W, 3]
    
                #img = (img_tensor / 2 + 0.5).clamp(0, 1).cpu().numpy().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][:200] 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=96)
        
        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],
                        "width": img.width,
                        "height": img.height,
                        "global_step": step
                    }
                )
    finally:
        # вернуть VAE на CPU (как было в твоём коде)
        vae.to("cpu")
        for var in list(locals().keys()):
            if isinstance(locals()[var], torch.Tensor):
                del locals()[var]
        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,0)
accelerator.wait_for_everyone()

# Модифицируем функцию сохранения модели для поддержки LoRA
def save_checkpoint(unet, variant=""):
    if accelerator.is_main_process:
        if lora_name:
            save_lora_checkpoint(unet)
        else:
            # безопасный unwrap для компилированной модели
            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}"))

            unet = unet.to(dtype=dtype)

# --------------------------- 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)
sample_interval = max(1, steps_per_epoch // sample_interval_share)
min_loss = 2.

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()
    #print("epoch:",epoch)
    for step, (latents, embeddings, attention_mask) in enumerate(dataloader):
        with accelerator.accumulate(unet):
            if save_model == False and step == 5 :
                used_gb = torch.cuda.max_memory_allocated() / 1024**3
                print(f"Шаг {step}: {used_gb:.2f} GB")
                
            # шум
            noise = torch.randn_like(latents, dtype=latents.dtype)
    
            # берём t из [0, 1]
            t = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
    
            # 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 * scheduler.config.num_train_timesteps).long()
    
            # предсказание потока (Flow)
            model_pred = unet(noisy_latents, timesteps, embeddings, encoder_attention_mask=attention_mask).sample
    
            # таргет — векторное поле (= разность между конечными точками)
            target = noise - latents   # или latents - noise?
    
            # MSE лосс
            mse_loss = F.mse_loss(model_pred.float(), target.float())

            # Сохраняем для логов (мы сохраняем MSE отдельно — как показатель)
            batch_losses.append(mse_loss.detach().item())

            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
                accelerator.wait_for_everyone()
                
            # Backward
            accelerator.backward(mse_loss)

            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
                accelerator.wait_for_everyone()
                
            grad = 0.0
            if not fbp:
                if accelerator.sync_gradients:
                    with torch.amp.autocast('cuda', enabled=False):
                        grad_val = accelerator.clip_grad_norm_(unet.parameters(), clip_grad_norm)
                    grad = 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)
                # Логируем метрики
                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_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)
    
                    # Генерируем сэмплы с заданным интервалом
                    if global_step % sample_interval == 0:
                        generate_and_save_samples(fixed_samples,uncond_emb, global_step)
                        last_n = sample_interval
    
                        if save_model:
                            avg_sample_loss = np.mean(batch_losses[-sample_interval:]) if len(batch_losses) > 0 else 0.0
                            print("saving:", avg_sample_loss < min_loss * save_barrier, "Avg:", avg_sample_loss)
                            if avg_sample_loss is not None and avg_sample_loss < min_loss * save_barrier:
                                min_loss = avg_sample_loss
                                save_checkpoint(unet)


    if accelerator.is_main_process:
        # local averages
        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)

# Завершение обучения - сохраняем финальную модель
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!")