2601

train-Copy1.py

@@ -0,0 +1,897 @@
+#from comet_ml import Experiment
+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 diffusers import UNet2DConditionModel, AutoencoderKL,AutoencoderKLFlux2,AsymmetricAutoencoderKL,FlowMatchEulerDiscreteScheduler
+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
+from transformers import AutoTokenizer, AutoModel
+
+# --------------------------- Параметры ---------------------------
+ds_path = "/workspace/sdxs-08b/datasets/d123_640_sd15"
+project = "unet"
+batch_size = 15
+base_learning_rate = 4e-5
+min_learning_rate = 2e-5
+num_epochs = 10
+sample_interval_share = 5
+cfg_dropout = 0.10
+max_length = 248
+use_wandb = False
+use_comet_ml = True
+save_model = True
+use_decay = True
+fbp = False
+optimizer_type = "adam8bit"
+torch_compile = False
+unet_gradient = True
+loss_normalize = False
+fixed_seed = False
+shuffle = True
+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_mem_efficient_sdp(False)
+dtype = torch.float32
+save_barrier = 1.05
+warmup_percent = 0.03
+percentile_clipping = 95
+betta2 = 0.995
+eps = 1e-7
+clip_grad_norm = 1.0
+limit = 0
+checkpoints_folder = ""
+mixed_precision = "no"
+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 = 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"))
+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")
+
+loss_ratios = {
+    "mse":   1.25,
+    "mae":   0.25,
+}
+median_coeff_steps = 256 
+
+# Нормализация лоссов по медианам: считаем КОЭФФИЦИЕНТЫ
+class MedianLossNormalizer:
+    def __init__(self, desired_ratios: dict, window_steps: int):
+        # нормируем доли на случай, если сумма != 1
+        #s = sum(desired_ratios.values())
+        #self.ratios = {k: (v / s) for k, v in desired_ratios.items()}
+        self.ratios = {k: float(v) for k, v in desired_ratios.items()}
+        self.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
+        self.window = window_steps
+
+    def update_and_total(self, losses: dict):
+        """
+        losses: dict ключ->тензор (значения лоссов)
+        Поведение:
+          - буферим ABS(l) только для активных (ratio>0) лоссов
+          - coeff = ratio / median(abs(loss))
+          - total = sum(coeff * loss) по активным лоссам
+        CHANGED: буферим abs() — чтобы медиана была положительной и не ломала деление.
+        """
+        # буферим только активные лоссы
+        for k, v in losses.items():
+            if k in self.buffers and self.ratios.get(k, 0) > 0:
+                val = v.detach().abs().mean().cpu().item() # .item() лучше float() для тензоров
+                self.buffers[k].append(val)
+                #self.buffers[k].append(float(v.detach().abs().cpu()))
+
+        meds = {k: (np.median(self.buffers[k]) if len(self.buffers[k]) > 0 else 1.0) for k in self.buffers}
+        coeffs = {k: (self.ratios[k] / max(meds[k], 1e-12)) for k in self.ratios}
+
+        # суммируем только по активным (ratio>0)
+        total = sum(coeffs[k] * losses[k] for k in coeffs if self.ratios.get(k, 0) > 0)
+        return total, coeffs, meds
+
+# создаём normalizer после определения loss_ratios
+normalizer = MedianLossNormalizer(loss_ratios, median_coeff_steps)
+
+# --------------------------- 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,
+            "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)
+
+# Включение Flash Attention 2/SDPA
+torch.backends.cuda.enable_flash_sdp(True)
+
+# --------------------------- Загрузка моделей ---------------------------
+#vae = AutoencoderKL.from_pretrained("vae", torch_dtype=dtype).to("cpu").eval()
+#vae = AutoencoderKLFlux2.from_pretrained("black-forest-labs/FLUX.2-dev",subfolder="vae",torch_dtype=dtype).to(device).eval()
+vae = AsymmetricAutoencoderKL.from_pretrained("vae",torch_dtype=dtype).to(device).eval()
+tokenizer = AutoTokenizer.from_pretrained("tokenizer")
+text_model = AutoModel.from_pretrained("text_encoder").to(device).eval()
+scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("scheduler")
+
+def encode_texts(texts, max_length=77): # Для SD 1.5 лучше жестко 77
+    if texts is None:
+        texts = [""]
+    
+    if isinstance(texts, str):
+        texts = [texts]
+
+    with torch.no_grad():
+        # 1. Основная токенизация
+        toks = tokenizer(
+            texts,
+            padding="max_length",
+            max_length=max_length,
+            truncation=True,
+            return_tensors="pt"
+        ).to(device)
+
+        text_input_ids = toks.input_ids
+        
+        # 2. Проверка на обрезку текста (логика для вывода предупреждений)
+        # Получаем айдишники без обрезки по длине
+        # untruncated_ids = tokenizer(texts, padding="longest", return_tensors="pt").input_ids.to(device)
+
+        # Исправляем проверку: сравниваем тензоры, а не объект BatchEncoding
+        #if untruncated_ids.shape[-1] > text_input_ids.shape[-1] and not torch.equal(
+        #    text_input_ids, untruncated_ids[:, :max_length]
+        #):
+            # Заменяем self.tokenizer на tokenizer
+        #    removed_text = tokenizer.batch_decode(
+        #        untruncated_ids[:, max_length - 1 : -1]
+        #    )
+            #print(f"Warning: Text truncated. Removed part: {removed_text}")
+
+        # 3. Маска внимания
+        #if hasattr(text_model.config, "use_attention_mask") and text_model.config.use_attention_mask:
+        #    attention_mask = toks.attention_mask
+        #else:
+            #attention_mask = None
+        attention_mask = toks.attention_mask
+
+        # 4. Прогон через модель
+        # Правильный вызов: передаем конкретные тензоры или распаковываем словарь **toks
+        outputs = text_model(
+            input_ids=text_input_ids, 
+            attention_mask=attention_mask,
+            output_hidden_states=True # Часто нужно для SD 1.5 (слой -2)
+        )
+        
+        layer_index = -2
+        prompt_embeds = outputs.hidden_states[layer_index]
+
+        # 2. ДОБАВЛЯЕМ ФИНАЛЬНУЮ НОРМАЛИЗАЦИЮ
+        # В CLIP после всех блоков стоит слой LayerNorm. 
+        final_layer_norm = text_model.text_model.final_layer_norm
+        prompt_embeds = final_layer_norm(prompt_embeds)
+
+        return prompt_embeds, attention_mask
+        
+# --- [UPDATED] Функция кодирования текста (с маской и пулингом) ---
+def encode_texts22(texts, max_length=max_length):
+    # Если тексты пустые (для unconditional), создаем заглушки
+    if texts is None: 
+        # В случае None возвращаем нули (логика для get_negative_embedding)
+        # Но здесь мы обычно ожидаем список строк.
+        pass
+
+    with torch.no_grad():
+        if isinstance(texts, str):
+            texts = [texts]
+
+        #for i, prompt_item in enumerate(texts):
+        #    messages = [
+        #        {"role": "user", "content": prompt_item},
+        #    ]
+        #    prompt_item = tokenizer.apply_chat_template(
+        #        messages,
+        #        tokenize=False,
+        #        add_generation_prompt=True,
+        #        enable_thinking=True,
+        #    )
+            #print(prompt_item+"\n")
+        #    texts[i] = prompt_item
+            
+        toks = tokenizer(
+            texts,
+            return_tensors="pt",
+            padding="max_length",
+            truncation=True,
+            max_length=max_length
+        ).to(device)
+
+        text_input_ids = toks.input_ids
+        untruncated_ids = tokenizer(texts, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= toks.shape[-1] and not torch.equal(
+            toks, untruncated_ids
+        ):
+            removed_text = tokenizer.batch_decode(
+                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+            )
+            print(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                )
+
+        attention_mask = toks.attention_mask.to(device)
+
+        prompt_embeds = text_model(toks.to(device), attention_mask=attention_mask)
+        prompt_embeds = prompt_embeds[-1]
+        return prompt_embeds, attention_mask
+        #outs = text_model(**toks, output_hidden_states=True, return_dict=True)
+        
+        # Используем last_hidden_state или hidden_states[-1] (если Qwen, лучше last_hidden_state - прим человека: ХУЙ)
+        #hidden = outs.hidden_states[-2] 
+        
+        # 2. Маска внимания
+        #attention_mask = toks["attention_mask"] 
+        
+        # 3. Пулинг-эмбеддинг (Последний токен)
+        #sequence_lengths = attention_mask.sum(dim=1) - 1
+        #batch_size = hidden.shape[0]
+        #pooled = hidden[torch.arange(batch_size, device=hidden.device), sequence_lengths]
+
+        #return hidden, attention_mask
+        # --- НОВАЯ ЛОГИКА: ОБЪЕДИНЕНИЕ ДЛЯ КРОСС-ВНИМАНИЯ ---
+        # 1. Расширяем пулинг-вектор до последовательности [B, 1, emb]
+        #pooled_expanded = pooled.unsqueeze(1) 
+                    
+        # 2. Объединяем последовательность токенов и пулинг-вектор
+        # !!! ИЗМЕНЕНИЕ ЗДЕСЬ !!!: Пулинг идет ПЕРВЫМ
+        # Теперь: [B, 1 + L, emb]. Пулинг стал токеном в НАЧАЛЕ.
+        #new_encoder_hidden_states = torch.cat([pooled_expanded, hidden], dim=1) 
+                    
+        # 3. Обновляем маску внимания для нового токена
+        # Маска внимания: [B, 1 + L]. Добавляем 1 в НАЧАЛО.
+        # torch.ones((batch_size, 1), device=device) создает маску [B, 1] со значениями 1.
+        #new_attention_mask = torch.cat([torch.ones((batch_size, 1), device=device), attention_mask], dim=1)
+        
+        #return new_encoder_hidden_states, new_attention_mask
+
+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
+scaling_factor = 1.0
+
+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
+
+# --- [UPDATED] Функция для фиксированных семплов ---
+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)
+        texts = [item["text"] for item in samples_data]
+        
+        # Кодируем тексты на лету, чтобы получить маски и пулинг
+        embeddings, masks = encode_texts(texts)
+        
+        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)
+
+dataset = dataset.filter(
+    lambda x: [not (path.startswith("/workspace/ds/animesfw") or path.startswith("/workspace/ds/d4/animesfw"))  for path in x["image_path"]],
+    batched=True,
+    batch_size=10000, # обрабатываем по 10к строк за раз
+    num_proc=8
+)
+print(f"Осталось примеров после фильтрации: {len(dataset)}")
+
+# --- [UPDATED] Collate Function ---
+def collate_fn_simple(batch):
+    # 1. Латенты (VAE)
+    latents = torch.tensor(np.array([item["vae"] for item in batch])).to(device, dtype=dtype)
+    
+    # 2. Текст берем сырой из датасета
+    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
+    ]
+    
+    # 3. Кодируем на лету
+    # Возвращает: hidden (B, L, D), mask (B, L)
+    embeddings, attention_mask = encode_texts(texts)
+    
+    # attention_mask от токенизатора уже имеет нужный формат, но на всякий случай приведем к long
+    attention_mask = attention_mask.to(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)
+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)
+    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}")
+
+if lora_name:
+    # ... (Код LoRA без изменений, опущен для краткости, если не используется, иначе раскомментируйте оригинальный блок) ...
+    pass 
+
+# 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, betta2), eps=1e-8, weight_decay=0.01
+        )
+    elif name == "muon":
+        from muon import MuonWithAuxAdam
+        trainable_params = [p for p in params if p.requires_grad]
+        hidden_weights = [p for p in trainable_params if p.ndim >= 2]
+        hidden_gains_biases = [p for p in trainable_params if p.ndim < 2]
+    
+        param_groups = [ 
+            dict(params=hidden_weights, use_muon=True,
+                lr=1e-3, weight_decay=1e-4),
+            dict(params=hidden_gains_biases, use_muon=False,
+                lr=1e-4, betas=(0.9, 0.95), weight_decay=1e-4),
+        ]
+        optimizer = MuonWithAuxAdam(param_groups)
+        from snooc import SnooC
+        return SnooC(optimizer)
+    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:
+    # 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}")
+    
+    # 3. Передаем в оптимизатор ТОЛЬКО обучаемые параметры
+    #optimizer = create_optimizer(optimizer_type, train_params)
+
+    unet.requires_grad_(True)
+    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)
+    unet, optimizer, lr_scheduler = accelerator.prepare(unet, optimizer, lr_scheduler)
+
+if torch_compile:
+    print("compiling")
+    unet = torch.compile(unet)
+    print("compiling - ok")
+
+# 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 = 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.int64, device=device)
+        return empty_emb, empty_mask
+
+    uncond_emb, uncond_mask = encode_texts([neg_prompt])
+    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
+uncond_emb, uncond_mask = get_negative_embedding("low quality")
+
+# --- 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
+    
+    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)
+                
+                    # Подготовка батчей для CFG (Negative + Positive)
+                    # 1. Embeddings
+                    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)
+                    text_embeddings_batch = torch.cat([neg_emb_batch, sample_text_embeddings], dim=0)
+                    
+                    # 2. Masks
+                    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
+
+                # Предсказание с передачей всех условий
+                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
+
+            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):
+                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=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]}
+                )
+    finally:
+        vae.to("cpu")
+        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:
+        if lora_name:
+            save_lora_checkpoint(unet)
+        else:
+            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 = 4.
+
+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):
+        with accelerator.accumulate(unet):
+            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")
+                
+            # шум
+            noise = torch.randn_like(latents, dtype=latents.dtype)
+
+            # 3. Время t (сэмплим, как и раньше, но чуть сжимаем края)
+            u = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
+            t = u * (1 - 2 * 1e-5) + 1e-5  # Теперь t строго в (0.00001 ... 0.99999)
+            # 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)
+            timesteps = timesteps.clamp(0, scheduler.config.num_train_timesteps - 1)
+            
+            # --- Вызов UNet с маской  ---
+            model_pred = unet(
+                noisy_latents, 
+                timesteps, 
+                encoder_hidden_states=embeddings,
+                encoder_attention_mask=attention_mask
+            ).sample
+            
+            target = noise - latents
+
+            mse_loss = F.mse_loss(model_pred.float(), target.float())
+            mae_loss = F.l1_loss(model_pred.float(), target.float())
+                
+            batch_losses.append(mse_loss.detach().item())
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+
+            losses_dict = {}
+            losses_dict["mse"] = mse_loss
+            losses_dict["mae"] = mae_loss
+
+            # === Нормализация всех лоссов ===
+            abs_for_norm = {k: losses_dict.get(k, torch.tensor(0.0, device=device)) for k in normalizer.ratios.keys()}
+            total_loss, coeffs, meds = normalizer.update_and_total(abs_for_norm)
+
+            if (global_step % 100 == 0) or (global_step % sample_interval == 0):
+                accelerator.wait_for_everyone()
+
+            if loss_normalize:    
+                accelerator.backward(total_loss)
+            else:
+                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)
+                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["loss_mae"] = mae_loss.detach().item()
+                    log_data["lr"] = current_lr
+                    log_data["grad"] = grad
+                    log_data["loss_norm"] = float(total_loss.item())
+                    for k, c in coeffs.items():
+                        log_data[f"coeff_{k}"] = float(c)
+                    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)
+
+                    if global_step % sample_interval == 0:
+                        # 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 = sample_interval
+                        
+                        if save_model:
+                            has_losses = len(batch_losses) > 0
+                            avg_sample_loss = np.mean(batch_losses[-sample_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)
+
+    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!")

train.py

@@ -29,10 +29,10 @@ from transformers import AutoTokenizer, AutoModel
 # --------------------------- Параметры ---------------------------
 ds_path = "/workspace/sdxs-08b/datasets/d123_640_sd15"
 project = "unet"
-batch_size = 10
-base_learning_rate = 5e-5
+batch_size = 15
+base_learning_rate = 4e-5
 min_learning_rate = 2e-5
-num_epochs = 20
+num_epochs = 10
 sample_interval_share = 5
 cfg_dropout = 0.10
 max_length = 248
@@ -54,7 +54,7 @@ torch.backends.cudnn.allow_tf32 = True
 torch.backends.cuda.enable_mem_efficient_sdp(False)
 dtype = torch.float32
 save_barrier = 1.05
-warmup_percent = 0.05
+warmup_percent = 0.03
 percentile_clipping = 95
 betta2 = 0.995
 eps = 1e-7
@@ -62,7 +62,7 @@ clip_grad_norm = 1.0
 limit = 0
 checkpoints_folder = ""
 mixed_precision = "no"
-gradient_accumulation_steps = 2
+gradient_accumulation_steps = 1
 
 accelerator = Accelerator(
     mixed_precision=mixed_precision,
@@ -89,59 +89,12 @@ if fixed_seed:
     if torch.cuda.is_available():
         torch.cuda.manual_seed_all(seed)
 
-# --------------------------- Параметры LoRA ---------------------------
-lora_name = "" 
-lora_rank = 32
-lora_alpha = 64
-
 print("init")
 
-loss_ratios = {
-    "mse":   1.25,
-    "mae":   0.25,
-}
-median_coeff_steps = 256 
-
-# Нормализация лоссов по медианам: считаем КОЭФФИЦИЕНТЫ
-class MedianLossNormalizer:
-    def __init__(self, desired_ratios: dict, window_steps: int):
-        # нормируем доли на случай, если сумма != 1
-        #s = sum(desired_ratios.values())
-        #self.ratios = {k: (v / s) for k, v in desired_ratios.items()}
-        self.ratios = {k: float(v) for k, v in desired_ratios.items()}
-        self.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
-        self.window = window_steps
-
-    def update_and_total(self, losses: dict):
-        """
-        losses: dict ключ->тензор (значения лоссов)
-        Поведение:
-          - буферим ABS(l) только для активных (ratio>0) лоссов
-          - coeff = ratio / median(abs(loss))
-          - total = sum(coeff * loss) по активным лоссам
-        CHANGED: буферим abs() — чтобы медиана была положительной и не ломала деление.
-        """
-        # буферим только активные лоссы
-        for k, v in losses.items():
-            if k in self.buffers and self.ratios.get(k, 0) > 0:
-                val = v.detach().abs().mean().cpu().item() # .item() лучше float() для тензоров
-                self.buffers[k].append(val)
-                #self.buffers[k].append(float(v.detach().abs().cpu()))
-
-        meds = {k: (np.median(self.buffers[k]) if len(self.buffers[k]) > 0 else 1.0) for k in self.buffers}
-        coeffs = {k: (self.ratios[k] / max(meds[k], 1e-12)) for k in self.ratios}
-
-        # суммируем только по активным (ratio>0)
-        total = sum(coeffs[k] * losses[k] for k in coeffs if self.ratios.get(k, 0) > 0)
-        return total, coeffs, meds
-
-# создаём normalizer после определения loss_ratios
-normalizer = MedianLossNormalizer(loss_ratios, median_coeff_steps)
-
 # --------------------------- Инициализация WandB ---------------------------
 if accelerator.is_main_process:
     if use_wandb:
-        wandb.init(project=project+lora_name, config={
+        wandb.init(project=project, config={
             "batch_size": batch_size,
             "base_learning_rate": base_learning_rate,
             "num_epochs": num_epochs,
@@ -172,7 +125,7 @@ tokenizer = AutoTokenizer.from_pretrained("tokenizer")
 text_model = AutoModel.from_pretrained("text_encoder").to(device).eval()
 scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("scheduler")
 
-def encode_texts(texts, max_length=77): # Для SD 1.5 лучше жестко 77
+def encode_texts(texts, max_length=max_length): # Для SD 1.5 лучше жестко 77
     if texts is None:
         texts = [""]
     
@@ -190,26 +143,6 @@ def encode_texts(texts, max_length=77): # Для SD 1.5 лучше жестко
         ).to(device)
 
         text_input_ids = toks.input_ids
-        
-        # 2. Проверка на обрезку текста (логика для вывода предупреждений)
-        # Получаем айдишники без обрезки по длине
-        # untruncated_ids = tokenizer(texts, padding="longest", return_tensors="pt").input_ids.to(device)
-
-        # Исправляем проверку: сравниваем тензоры, а не объект BatchEncoding
-        #if untruncated_ids.shape[-1] > text_input_ids.shape[-1] and not torch.equal(
-        #    text_input_ids, untruncated_ids[:, :max_length]
-        #):
-            # Заменяем self.tokenizer на tokenizer
-        #    removed_text = tokenizer.batch_decode(
-        #        untruncated_ids[:, max_length - 1 : -1]
-        #    )
-            #print(f"Warning: Text truncated. Removed part: {removed_text}")
-
-        # 3. Маска внимания
-        #if hasattr(text_model.config, "use_attention_mask") and text_model.config.use_attention_mask:
-        #    attention_mask = toks.attention_mask
-        #else:
-            #attention_mask = None
         attention_mask = toks.attention_mask
 
         # 4. Прогон через модель
@@ -229,93 +162,11 @@ def encode_texts(texts, max_length=77): # Для SD 1.5 лучше жестко
         prompt_embeds = final_layer_norm(prompt_embeds)
 
         return prompt_embeds, attention_mask
-        
-# --- [UPDATED] Функция кодирования текста (с маской и пулингом) ---
-def encode_texts22(texts, max_length=max_length):
-    # Если тексты пустые (для unconditional), создаем заглушки
-    if texts is None: 
-        # В случае None возвращаем нули (логика для get_negative_embedding)
-        # Но здесь мы обычно ожидаем список строк.
-        pass
-
-    with torch.no_grad():
-        if isinstance(texts, str):
-            texts = [texts]
-
-        #for i, prompt_item in enumerate(texts):
-        #    messages = [
-        #        {"role": "user", "content": prompt_item},
-        #    ]
-        #    prompt_item = tokenizer.apply_chat_template(
-        #        messages,
-        #        tokenize=False,
-        #        add_generation_prompt=True,
-        #        enable_thinking=True,
-        #    )
-            #print(prompt_item+"\n")
-        #    texts[i] = prompt_item
-            
-        toks = tokenizer(
-            texts,
-            return_tensors="pt",
-            padding="max_length",
-            truncation=True,
-            max_length=max_length
-        ).to(device)
-
-        text_input_ids = toks.input_ids
-        untruncated_ids = tokenizer(texts, padding="longest", return_tensors="pt").input_ids
-
-        if untruncated_ids.shape[-1] >= toks.shape[-1] and not torch.equal(
-            toks, untruncated_ids
-        ):
-            removed_text = tokenizer.batch_decode(
-                untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
-            )
-            print(
-                "The following part of your input was truncated because CLIP can only handle sequences up to"
-                )
-
-        attention_mask = toks.attention_mask.to(device)
-
-        prompt_embeds = text_model(toks.to(device), attention_mask=attention_mask)
-        prompt_embeds = prompt_embeds[-1]
-        return prompt_embeds, attention_mask
-        #outs = text_model(**toks, output_hidden_states=True, return_dict=True)
-        
-        # Используем last_hidden_state или hidden_states[-1] (если Qwen, лучше last_hidden_state - прим человека: ХУЙ)
-        #hidden = outs.hidden_states[-2] 
-        
-        # 2. Маска внимания
-        #attention_mask = toks["attention_mask"] 
-        
-        # 3. Пулинг-эмбеддинг (Последний токен)
-        #sequence_lengths = attention_mask.sum(dim=1) - 1
-        #batch_size = hidden.shape[0]
-        #pooled = hidden[torch.arange(batch_size, device=hidden.device), sequence_lengths]
-
-        #return hidden, attention_mask
-        # --- НОВАЯ ЛОГИКА: ОБЪЕДИНЕНИЕ ДЛЯ КРОСС-ВНИМАНИЯ ---
-        # 1. Расширяем пулинг-вектор до последовательности [B, 1, emb]
-        #pooled_expanded = pooled.unsqueeze(1) 
-                    
-        # 2. Объединяем последовательность токенов и пулинг-вектор
-        # !!! ИЗМЕНЕНИЕ ЗДЕСЬ !!!: Пулинг идет ПЕРВЫМ
-        # Теперь: [B, 1 + L, emb]. Пулинг стал токеном в НАЧАЛЕ.
-        #new_encoder_hidden_states = torch.cat([pooled_expanded, hidden], dim=1) 
-                    
-        # 3. Обновляем маску внимания для нового токена
-        # Маска внимания: [B, 1 + L]. Добавляем 1 в НАЧАЛО.
-        # torch.ones((batch_size, 1), device=device) создает маску [B, 1] со значениями 1.
-        #new_attention_mask = torch.cat([torch.ones((batch_size, 1), device=device), attention_mask], dim=1)
-        
-        #return new_encoder_hidden_states, new_attention_mask
 
 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
-scaling_factor = 1.0
 
 class DistributedResolutionBatchSampler(Sampler):
     def __init__(self, dataset, batch_size, num_replicas, rank, shuffle=True, drop_last=True):
@@ -427,7 +278,7 @@ dataset = dataset.filter(
 )
 print(f"Осталось примеров после фильтрации: {len(dataset)}")
 
-# --- [UPDATED] Collate Function ---
+# --- Collate Function ---
 def collate_fn_simple(batch):
     # 1. Латенты (VAE)
     latents = torch.tensor(np.array([item["vae"] for item in batch])).to(device, dtype=dtype)
@@ -441,7 +292,6 @@ def collate_fn_simple(batch):
         else t.replace("The image shows ", "").replace("The image is ", "").replace("This image captures ","").strip()
         for t in raw_texts
     ]
-    
     # 3. Кодируем на лету
     # Возвращает: hidden (B, L, D), mask (B, L)
     embeddings, attention_mask = encode_texts(texts)
@@ -484,17 +334,6 @@ if os.path.isdir(latest_checkpoint):
         unet.set_use_memory_efficient_attention_xformers(True)
 else:
     raise FileNotFoundError(f"UNet checkpoint not found at {latest_checkpoint}")
-
-if lora_name:
-    # ... (Код LoRA без изменений, опущен для краткости, если не используется, иначе раскомментируйте оригинальный блок) ...
-    pass 
-
-# 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):
@@ -507,25 +346,11 @@ def create_optimizer(name, params):
         return torch.optim.AdamW(
             params, lr=base_learning_rate, betas=(0.9, betta2), eps=1e-8, weight_decay=0.01
         )
-    elif name == "muon":
-        from muon import MuonWithAuxAdam
-        trainable_params = [p for p in params if p.requires_grad]
-        hidden_weights = [p for p in trainable_params if p.ndim >= 2]
-        hidden_gains_biases = [p for p in trainable_params if p.ndim < 2]
-    
-        param_groups = [ 
-            dict(params=hidden_weights, use_muon=True,
-                lr=1e-3, weight_decay=1e-4),
-            dict(params=hidden_gains_biases, use_muon=False,
-                lr=1e-4, betas=(0.9, 0.95), weight_decay=1e-4),
-        ]
-        optimizer = MuonWithAuxAdam(param_groups)
-        from snooc import SnooC
-        return SnooC(optimizer)
     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()
@@ -546,9 +371,6 @@ else:
     #        param.requires_grad = True
     #        train_params.append(param)
     #        print(f"Обучаемый слой: {name}")
-    
-    # 3. Передаем в оптимизатор ТОЛЬКО обучаемые параметры
-    #optimizer = create_optimizer(optimizer_type, train_params)
 
     unet.requires_grad_(True)
     optimizer = create_optimizer(optimizer_type, unet.parameters())
@@ -709,6 +531,17 @@ def generate_and_save_samples(fixed_samples_cpu, uncond_data, step):
                 )
     finally:
         vae.to("cpu")
+        try:
+            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()
 
@@ -721,23 +554,20 @@ accelerator.wait_for_everyone()
 
 def save_checkpoint(unet, variant=""):
     if accelerator.is_main_process:
-        if lora_name:
-            save_lora_checkpoint(unet)
+        model_to_save = None
+        if not torch_compile:
+            model_to_save = accelerator.unwrap_model(unet)
         else:
-            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}"))
+            model_to_save = unet
 
-            unet = unet.to(dtype=dtype)
+        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:
@@ -785,9 +615,7 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
             
             target = noise - latents
 
-            mse_loss = F.mse_loss(model_pred.float(), target.float())
-            mae_loss = F.l1_loss(model_pred.float(), target.float())
-                
+            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 % sample_interval == 0):
@@ -795,19 +623,11 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
 
             losses_dict = {}
             losses_dict["mse"] = mse_loss
-            losses_dict["mae"] = mae_loss
-
-            # === Нормализация всех лоссов ===
-            abs_for_norm = {k: losses_dict.get(k, torch.tensor(0.0, device=device)) for k in normalizer.ratios.keys()}
-            total_loss, coeffs, meds = normalizer.update_and_total(abs_for_norm)
 
             if (global_step % 100 == 0) or (global_step % sample_interval == 0):
                 accelerator.wait_for_everyone()
 
-            if loss_normalize:    
-                accelerator.backward(total_loss)
-            else:
-                accelerator.backward(mse_loss)
+            accelerator.backward(mse_loss)
 
             if (global_step % 100 == 0) or (global_step % sample_interval == 0):
                 accelerator.wait_for_everyone()
@@ -815,7 +635,6 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
             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()
@@ -834,12 +653,8 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
     
                     log_data = {}
                     log_data["loss_mse"] = mse_loss.detach().item()
-                    log_data["loss_mae"] = mae_loss.detach().item()
                     log_data["lr"] = current_lr
                     log_data["grad"] = grad
-                    log_data["loss_norm"] = float(total_loss.item())
-                    for k, c in coeffs.items():
-                        log_data[f"coeff_{k}"] = float(c)
                     if accelerator.sync_gradients:
                         if use_wandb:
                             wandb.log(log_data, step=global_step)
@@ -868,6 +683,7 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
                             if should_save:
                                 min_loss = max_loss
                                 save_checkpoint(unet)
+                                unet.train()
 
     if accelerator.is_main_process:
         avg_epoch_loss = np.mean(batch_losses) if len(batch_losses) > 0 else 0.0