Upload folder using huggingface_hub

.ipynb_checkpoints/config-checkpoint.json

@@ -0,0 +1,72 @@
+{
+  "_class_name": "UNet2DConditionModel",
+  "_diffusers_version": "0.36.0",
+  "act_fn": "silu",
+  "addition_embed_type": null,
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": null,
+  "attention_head_dim": [
+    10,
+    20,
+    20
+  ],
+  "attention_type": "default",
+  "block_out_channels": [
+    320,
+    640,
+    1280
+  ],
+  "center_input_sample": false,
+  "class_embed_type": null,
+  "class_embeddings_concat": false,
+  "conv_in_kernel": 3,
+  "conv_out_kernel": 3,
+  "cross_attention_dim": 1024,
+  "cross_attention_norm": null,
+  "down_block_types": [
+    "DownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "dropout": 0.0,
+  "dual_cross_attention": false,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "in_channels": 32,
+  "layers_per_block": 2,
+  "mid_block_only_cross_attention": null,
+  "mid_block_scale_factor": 1.0,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "out_channels": 32,
+  "projection_class_embeddings_input_dim": null,
+  "resnet_out_scale_factor": 1.0,
+  "resnet_skip_time_act": false,
+  "resnet_time_scale_shift": "default",
+  "reverse_transformer_layers_per_block": null,
+  "sample_size": null,
+  "time_cond_proj_dim": null,
+  "time_embedding_act_fn": null,
+  "time_embedding_dim": null,
+  "time_embedding_type": "positional",
+  "timestep_post_act": null,
+  "transformer_layers_per_block": [
+    1,
+    2,
+    4
+  ],
+  "up_block_types": [
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D",
+    "UpBlock2D"
+  ],
+  "upcast_attention": false,
+  "use_linear_projection": true
+}

.ipynb_checkpoints/dataset_flux-checkpoint.py

@@ -0,0 +1,413 @@
+# pip install flash-attn --no-build-isolation
+import torch
+import os
+import gc
+import numpy as np
+import random
+import json
+import shutil
+import time
+
+from datasets import Dataset, load_from_disk, concatenate_datasets
+from diffusers import AutoencoderKL,AutoencoderKLWan,AutoencoderKLFlux2
+from torchvision.transforms import Resize, ToTensor, Normalize, Compose, InterpolationMode, Lambda
+from transformers import AutoModel, AutoImageProcessor, AutoTokenizer, AutoModelForCausalLM
+from typing import Dict, List, Tuple, Optional, Any
+from PIL import Image
+from tqdm import tqdm
+from datetime import timedelta
+
+# ---------------- 1️⃣ Настройки ----------------
+dtype = torch.float32
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+batch_size = 5
+min_size = 192 #384 #320 #192 #256 #192
+max_size = 320 #768 #640 #384 #256 #384
+step = 64 #64
+empty_share = 0.0
+limit = 0
+# Основная процедура обработки
+folder_path = "/workspace/mjnj" #alchemist"
+save_path = "/workspace/sdxs/datasets/mjnj" #"alchemist"
+os.makedirs(save_path, exist_ok=True)
+
+# Функция для очистки CUDA памяти
+def clear_cuda_memory():
+    if torch.cuda.is_available():
+        used_gb = torch.cuda.max_memory_allocated() / 1024**3
+        print(f"used_gb: {used_gb:.2f} GB")
+        torch.cuda.empty_cache()
+        gc.collect()
+
+# ---------------- 2️⃣ Загрузка моделей ----------------
+def load_models():
+    print("Загрузка моделей...")
+    #vae = AutoencoderKL.from_pretrained("AiArtLab/sdxs",subfolder="vae1x",torch_dtype=dtype).to(device).eval()
+    vae = AutoencoderKLFlux2.from_pretrained("black-forest-labs/FLUX.2-dev",subfolder="vae",torch_dtype=dtype).to(device).eval()
+
+    #model_name = "Qwen/Qwen3-0.6B"
+    #tokenizer = AutoTokenizer.from_pretrained(model_name)
+    #model = AutoModelForCausalLM.from_pretrained(
+    #    model_name,
+    #    torch_dtype=dtype,
+    #    device_map=device
+    #).eval()
+    #tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen3-Embedding-0.6B', padding_side='left')
+    #model = AutoModel.from_pretrained('Qwen/Qwen3-Embedding-0.6B').to("cuda")
+    return vae#, model, tokenizer
+
+#vae, model, tokenizer = load_models()
+vae = load_models()
+
+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)
+    
+# ---------------- 3️⃣ Трансформации ----------------
+def get_image_transform(min_size=256, max_size=512, step=64):
+    def transform(img, dry_run=False):
+        # Сохраняем исходные размеры изображения
+        original_width, original_height = img.size
+
+        # 0. Ресайз: масштабируем изображение, чтобы максимальная сторона была равна max_size
+        if original_width >= original_height:
+            new_width = max_size
+            new_height = int(max_size * original_height / original_width)
+        else:
+            new_height = max_size
+            new_width = int(max_size * original_width / original_height)
+
+        if new_height < min_size or new_width < min_size:
+            # 1. Ресайз: масштабируем изображение, чтобы минимальная сторона была равна min_size
+            if original_width <= original_height:
+                new_width = min_size
+                new_height = int(min_size * original_height / original_width)
+            else:
+                new_height = min_size
+                new_width = int(min_size * original_width / original_height)
+
+        # 2. Проверка: если одна из сторон превышает max_size, готовимся к обрезке
+        crop_width = min(max_size, (new_width // step) * step)
+        crop_height = min(max_size, (new_height // step) * step)
+
+        # Убеждаемся, что размеры обрезки не меньше min_size
+        crop_width = max(min_size, crop_width)
+        crop_height = max(min_size, crop_height)
+        
+        # Если запрошен только предварительный расчёт размеров
+        if dry_run:
+            return crop_width, crop_height
+        
+        # Конвертация в RGB и ресайз
+        img_resized = img.convert("RGB").resize((new_width, new_height), Image.LANCZOS)
+        
+        # Определение координат обрезки (обрезаем с учетом вотермарок - треть сверху)
+        top = (new_height - crop_height) // 3
+        left = 0
+        
+        # Обрезка изображения
+        img_cropped = img_resized.crop((left, top, left + crop_width, top + crop_height))
+        
+        # Сохраняем итоговые размеры после всех преобразований
+        final_width, final_height = img_cropped.size
+        
+        # тензор
+        img_tensor = ToTensor()(img_cropped)
+        img_tensor = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])(img_tensor)
+        return img_tensor, img_cropped, final_width, final_height
+
+    return transform
+
+# ---------------- 4️⃣ Функции обработки ----------------
+def last_token_pool(last_hidden_states: torch.Tensor,
+                    attention_mask: torch.Tensor) -> torch.Tensor:
+    # Определяем, есть ли left padding
+    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
+    if left_padding:
+        return last_hidden_states[:, -1]
+    else:
+        sequence_lengths = attention_mask.sum(dim=1) - 1
+        batch_size = last_hidden_states.shape[0]
+        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]
+
+def encode_texts_batch(texts, tokenizer, model, device="cuda", max_length=150, normalize=False):
+    with torch.inference_mode():
+        # Токенизация
+        batch = tokenizer(
+            texts,
+            return_tensors="pt",
+            padding="max_length",
+            truncation=True,
+            max_length=max_length
+        ).to(device)
+
+        # Прогон через модель
+        #outputs = model(**batch)
+
+        # Пулинг по last token
+        #embeddings = last_token_pool(outputs.last_hidden_state, batch["attention_mask"])
+
+        # L2-нормализация (опционально, обычно нужна для семантического поиска)
+        #if normalize:
+        #    embeddings = F.normalize(embeddings, p=2, dim=1)
+
+        # Прогон через базовую модель (внутри CausalLM)
+        outputs = model.model(**batch, output_hidden_states=True)
+
+        # Берем последний слой (эмбеддинги всех токенов)
+        hidden_states = outputs.hidden_states[-1]  # [B, L, D]
+
+        # Можно применить нормализацию по каждому токену (как в CLIP)
+        if normalize:
+            hidden_states = F.normalize(hidden_states, p=2, dim=-1)
+
+    return hidden_states.cpu().numpy() # embeddings.unsqueeze(1).cpu().numpy()
+
+def clean_label(label):
+    label = label.replace("Image 1", "").replace("Image 2", "").replace("Image 3", "").replace("Image 4", "").replace("The image depicts ","").replace("The image presents ","").replace("The image features ","").replace("The image portrays ","").replace("The image is ","").strip()
+    if label.startswith("."):
+        label = label[1:].lstrip()
+    return label
+
+def process_labels_for_guidance(original_labels, prob_to_make_empty=0.01):
+    """
+    Обрабатывает список меток для classifier-free guidance.
+
+    С вероятностью prob_to_make_empty:
+    - Метка в первом списке заменяется на пустую строку.
+    - К метке во втором списке добавляется префикс "zero:".
+
+    В противном случае метки в обоих списках остаются оригинальными.
+
+    """
+    labels_for_model = []
+    labels_for_logging = []
+
+    for label in original_labels:
+        if random.random() < prob_to_make_empty:
+            labels_for_model.append("")  # Заменяем на пустую строку для модели
+            labels_for_logging.append(f"zero: {label}") # Добавляем префикс для логгирования
+        else:
+            labels_for_model.append(label) # Оставляем оригинальную метку для модели
+            labels_for_logging.append(label) # Оставляем оригинальную метку для логгирования
+
+    return labels_for_model, labels_for_logging
+
+def encode_to_latents(images, texts):
+    transform = get_image_transform(min_size, max_size, step)
+    
+    try:
+        # Обработка изображений (все одинакового размера)
+        transformed_tensors = []
+        pil_images = []
+        widths, heights = [], []
+        
+        # Применяем трансформацию ко всем изображениям
+        for img in images:
+            try:
+                t_img, pil_img, w, h = transform(img)
+                transformed_tensors.append(t_img)
+                pil_images.append(pil_img)
+                widths.append(w)
+                heights.append(h)
+            except Exception as e:
+                print(f"Ошибка трансформации: {e}")
+                continue
+
+        if not transformed_tensors:
+            return None
+
+        # Создаём батч
+        batch_tensor = torch.stack(transformed_tensors).to(device, dtype)
+        if batch_tensor.ndim==5:
+            batch_tensor = batch_tensor.unsqueeze(2)  # [B, C, 1, H, W]
+        
+        # Кодируем батч
+        with torch.no_grad():
+            posteriors = vae.encode(batch_tensor).latent_dist.mode()
+            latents = (posteriors - shift_factor) / scaling_factor
+        
+        latents_np = latents.to(dtype).cpu().numpy()
+
+        # Обрабатываем тексты
+        text_labels = [clean_label(text) for text in texts]
+
+        model_prompts, text_labels = process_labels_for_guidance(text_labels, empty_share)                
+        #embeddings = encode_texts_batch(model_prompts, tokenizer, model)
+
+        return {
+            "vae": latents_np,
+            #"embeddings": embeddings,
+            "text": text_labels,
+            "width": widths,
+            "height": heights
+        }
+        
+    except Exception as e:
+        print(f"Критическая ошибка в encode_to_latents: {e}")
+        raise
+        
+
+# ---------------- 5️⃣ Обработка папки с изображениями и текстами ----------------
+def process_folder(folder_path, limit=None):
+    """
+    Рекурсивно обходит указанную директорию и все вложенные директории,
+    собирая пути к изображениям и соответствующим текстовым файлам.
+    """
+    image_paths = []
+    text_paths = []
+    width = []
+    height = []
+    transform = get_image_transform(min_size, max_size, step)
+    
+    # Используем os.walk для рекурсивного обхода директорий
+    for root, dirs, files in os.walk(folder_path):
+        for filename in files:
+            # Проверяем, является ли файл изображением
+            if filename.lower().endswith((".jpg", ".jpeg", ".png")):
+                image_path = os.path.join(root, filename)
+                try:
+                    img = Image.open(image_path)
+                except Exception as e:
+                    print(f"Ошибка при открытии {image_path}: {e}")
+                    os.remove(image_path)
+                    text_path = os.path.splitext(image_path)[0] + ".txt"
+                    if os.path.exists(text_path):
+                        os.remove(text_path)
+                    continue
+                # Применяем трансформацию только для получения размеров
+                w, h = transform(img, dry_run=True)
+                # Формируем путь к текстовому файлу
+                text_path = os.path.splitext(image_path)[0] + ".txt"
+                
+                # Добавляем пути, если текстовый файл существует
+                if os.path.exists(text_path) and min(w, h)>0:
+                    image_paths.append(image_path)
+                    text_paths.append(text_path)
+                    width.append(w)  # Добавляем в список
+                    height.append(h)  # Добавляем в список
+                    
+                    # Проверяем ограничение на количество
+                    if limit and limit>0 and len(image_paths) >= limit:
+                        print(f"Достигнут лимит в {limit} изображений")
+                        return image_paths, text_paths, width, height
+    
+    print(f"Найдено {len(image_paths)} изображений с текстовыми описаниями")
+    return image_paths, text_paths, width, height
+    
+def process_in_chunks(image_paths, text_paths, width, height, chunk_size=10000, batch_size=1):
+    total_files = len(image_paths)
+    start_time = time.time()
+    chunks = range(0, total_files, chunk_size)
+    
+    for chunk_idx, start in enumerate(chunks, 1):
+        end = min(start + chunk_size, total_files)
+        chunk_image_paths = image_paths[start:end]
+        chunk_text_paths = text_paths[start:end]
+        chunk_widths = width[start:end] if isinstance(width, list) else [width] * len(chunk_image_paths)
+        chunk_heights = height[start:end] if isinstance(height, list) else [height] * len(chunk_image_paths)
+        
+        # Чтение текстов
+        chunk_texts = []
+        for text_path in chunk_text_paths:
+            try:
+                with open(text_path, 'r', encoding='utf-8') as f:
+                    text = f.read().strip()
+                chunk_texts.append(text)
+            except Exception as e:
+                print(f"Ошибка чтения {text_path}: {e}")
+                chunk_texts.append("")
+        
+        # Группируем изображения по размерам
+        size_groups = {}
+        for i in range(len(chunk_image_paths)):
+            size_key = (chunk_widths[i], chunk_heights[i])
+            if size_key not in size_groups:
+                size_groups[size_key] = {"image_paths": [], "texts": []}
+            size_groups[size_key]["image_paths"].append(chunk_image_paths[i])
+            size_groups[size_key]["texts"].append(chunk_texts[i])
+        
+        # Обрабатываем каждую группу размеров отдельно
+        for size_key, group_data in size_groups.items():
+            print(f"Обработка группы с размером {size_key[0]}x{size_key[1]} - {len(group_data['image_paths'])} изображений")
+            
+            group_dataset = Dataset.from_dict({
+                "image_path": group_data["image_paths"],
+                "text": group_data["texts"]
+            })
+            
+            # Теперь можно использовать указанный batch_size, т.к. все изображения одного размера
+            processed_group = group_dataset.map(
+                lambda examples: encode_to_latents(
+                    [Image.open(path) for path in examples["image_path"]],
+                    examples["text"]
+                ),
+                batched=True,
+                batch_size=batch_size,
+                #remove_columns=["image_path"],
+                desc=f"Обработка группы размера {size_key[0]}x{size_key[1]}"
+            )
+            
+            # Сохраняем результаты группы
+            group_save_path = f"{save_path}_temp/chunk_{chunk_idx}_size_{size_key[0]}x{size_key[1]}"
+            processed_group.save_to_disk(group_save_path)
+            clear_cuda_memory()
+            elapsed = time.time() - start_time
+            processed = (chunk_idx - 1) * chunk_size + sum([len(sg["image_paths"]) for sg in list(size_groups.values())[:list(size_groups.values()).index(group_data) + 1]])
+            if processed > 0:
+                remaining = (elapsed / processed) * (total_files - processed)
+                elapsed_str = str(timedelta(seconds=int(elapsed)))
+                remaining_str = str(timedelta(seconds=int(remaining)))
+                print(f"ETA: Прошло {elapsed_str}, Осталось {remaining_str}, Прогресс {processed}/{total_files} ({processed/total_files:.1%})")
+
+# ---------------- 7️⃣ Объединение чанков ----------------
+def combine_chunks(temp_path, final_path):
+    """Объединение обработанных чанков в финальный датасет"""
+    chunks = sorted([
+        os.path.join(temp_path, d) 
+        for d in os.listdir(temp_path) 
+        if d.startswith("chunk_")
+    ])
+    
+    datasets = [load_from_disk(chunk) for chunk in chunks]
+    combined = concatenate_datasets(datasets)
+    combined.save_to_disk(final_path)
+    
+    print(f"✅ Датасет успешно сохранен в: {final_path}")
+
+    
+
+# Создаем временную папку для чанков
+temp_path = f"{save_path}_temp"
+os.makedirs(temp_path, exist_ok=True)
+
+# Получаем список файлов
+image_paths, text_paths, width, height = process_folder(folder_path,limit)
+print(f"Всего найдено {len(image_paths)} изображений")
+
+# Обработка с чанкованием
+process_in_chunks(image_paths, text_paths, width, height, chunk_size=20000, batch_size=batch_size)
+
+# Удаление  папки
+try:
+    shutil.rmtree(folder_path)
+    print(f"✅ Папка {folder_path} успешно удалена")
+except Exception as e:
+    print(f"⚠️ Ошибка при удалении папки: {e}")
+    
+# Объединение чанков в финальный датасет
+combine_chunks(temp_path, save_path)
+
+# Удаление временной папки
+try:
+    shutil.rmtree(temp_path)
+    print(f"✅ Временная папка {temp_path} успешно удалена")
+except Exception as e:
+    print(f"⚠️ Ошибка при удалении временной папки: {e}")

.ipynb_checkpoints/sdxs_create_flux-checkpoint.ipynb

@@ -0,0 +1,600 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "6bf71a1a-1bf0-42c7-8709-6686e8d2f46c",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "test unet\n",
+      "Количество параметров: 798780960\n",
+      "Output shape: torch.Size([1, 32, 60, 48])\n",
+      "UNet2DConditionModel(\n",
+      "  (conv_in): Conv2d(32, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "  (time_proj): Timesteps()\n",
+      "  (time_embedding): TimestepEmbedding(\n",
+      "    (linear_1): Linear(in_features=128, out_features=512, bias=True)\n",
+      "    (act): SiLU()\n",
+      "    (linear_2): Linear(in_features=512, out_features=512, bias=True)\n",
+      "  )\n",
+      "  (down_blocks): ModuleList(\n",
+      "    (0): DownBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=128, bias=True)\n",
+      "          (norm2): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (1): DownBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (2): CrossAttnDownBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-1): 2 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 512, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-1): 2 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=512, out_features=4096, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2048, out_features=512, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=512, out_features=512, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (3): CrossAttnDownBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-1): 2 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 1024, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-3): 4 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1024, out_features=8192, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4096, out_features=1024, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (up_blocks): ModuleList(\n",
+      "    (0): CrossAttnUpBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-2): 3 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 1024, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-3): 4 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1024, out_features=8192, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4096, out_features=1024, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 2048, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(2048, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(2048, 1024, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1536, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1536, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1536, 1024, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (1): CrossAttnUpBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-2): 3 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 512, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-1): 2 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=512, out_features=4096, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2048, out_features=512, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=512, out_features=512, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1536, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1536, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1536, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1024, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 768, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(768, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(768, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (2): UpBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 768, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(768, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(512, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 384, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(384, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (3): UpBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 384, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(384, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=128, bias=True)\n",
+      "          (norm2): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(384, 128, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1-2): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=128, bias=True)\n",
+      "          (norm2): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (mid_block): UNetMidBlock2DCrossAttn(\n",
+      "    (attentions): ModuleList(\n",
+      "      (0): Transformer2DModel(\n",
+      "        (norm): GroupNorm(32, 1024, eps=1e-06, affine=True)\n",
+      "        (proj_in): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "        (transformer_blocks): ModuleList(\n",
+      "          (0-3): 4 x BasicTransformerBlock(\n",
+      "            (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn1): Attention(\n",
+      "              (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn2): Attention(\n",
+      "              (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm3): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "            (ff): FeedForward(\n",
+      "              (net): ModuleList(\n",
+      "                (0): GEGLU(\n",
+      "                  (proj): Linear(in_features=1024, out_features=8192, bias=True)\n",
+      "                )\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "                (2): Linear(in_features=4096, out_features=1024, bias=True)\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "        )\n",
+      "        (proj_out): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "      )\n",
+      "    )\n",
+      "    (resnets): ModuleList(\n",
+      "      (0-1): 2 x ResnetBlock2D(\n",
+      "        (norm1): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "        (conv1): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "        (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "        (dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (nonlinearity): SiLU()\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (conv_norm_out): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "  (conv_act): SiLU()\n",
+      "  (conv_out): Conv2d(128, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      ")\n",
+      "Output shape: torch.Size([1, 32, 60, 48])\n"
+     ]
+    }
+   ],
+   "source": [
+    "config_sdxs = {\n",
+    "    # === Основные размеры и каналы ===\n",
+    "    \"in_channels\": 32,               # Количество входных каналов (совместимость с 16-канальным VAE)\n",
+    "    \"out_channels\": 32,              # Количество выходных каналов (симметрично in_channels)\n",
+    "    \"center_input_sample\": False,    # Отключение центрирования входных данных (стандарт для диффузионных моделей)\n",
+    "    \"flip_sin_to_cos\": True,         # Автоматическое преобразование sin/cos в эмбеддингах времени (для стабильности)\n",
+    "    \"freq_shift\": 0,                 # Сдвиг частоты (0 - стандартное значение для частотных эмбеддингов)\n",
+    "\n",
+    "    # === Архитектура блоков ===\n",
+    "    \"down_block_types\": [            # Типы блоков энкодера (иерархия обработки):\n",
+    "        \"DownBlock2D\",\n",
+    "        \"DownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "    ],\n",
+    "    \"mid_block_type\": \"UNetMidBlock2DCrossAttn\",  # Центральный блок с cross-attention (бутылочное горлышко сети)\n",
+    "    \"up_block_types\": [              # Типы блоков декодера (восстановление изображения):\n",
+    "        \"CrossAttnUpBlock2D\",\n",
+    "        \"CrossAttnUpBlock2D\",        \n",
+    "        \"UpBlock2D\",\n",
+    "        \"UpBlock2D\",\n",
+    "    ],\n",
+    "    \"only_cross_attention\": False,   # Использование как cross-attention, так и self-attention\n",
+    "\n",
+    "    # === Конфигурация каналов ===\n",
+    "    \"block_out_channels\": [128, 256, 512, 1024], \n",
+    "    \"layers_per_block\": 2, # Число слоев в блоках\n",
+    "    \"downsample_padding\": 1,          # Паддинг при уменьшении разрешения\n",
+    "    \"mid_block_scale_factor\": 1.0,    # Усиление сигнала в центральном блоке\n",
+    "\n",
+    "    # === Нормализация ===\n",
+    "    \"norm_num_groups\": 32,            # Число групп для GroupNorm (оптимально для стабильности)\n",
+    "    \"norm_eps\": 1e-05,                # Эпсилон для нормализации (стандартное значение)\n",
+    "\n",
+    "    # === Cross-Attention ===\n",
+    "    \"cross_attention_dim\": 1024,      # Размерность текстовых эмбеддинго\n",
+    "    \n",
+    "    \"transformer_layers_per_block\": [1, 1, 2, 4],  # Число трансформерных слоев (уменьшение с глубиной)\n",
+    "    \"attention_head_dim\": [2, 4, 8, 16],  # Размерность головы внимания \n",
+    "    \"dual_cross_attention\": False,    # Отключение двойного внимания (упрощение архитектуры)\n",
+    "    \"use_linear_projection\": True,    # Изменено на True для лучшей организации памяти\n",
+    "\n",
+    "    # === ResNet Блоки ===\n",
+    "    \"resnet_time_scale_shift\": \"default\",  # Способ интеграции временных эмбеддингов\n",
+    "    \"resnet_skip_time_act\": False,    # Отключение активации в skip-соединениях\n",
+    "    \"resnet_out_scale_factor\": 1.0,   # Коэффициент масштабирования выхода ResNet\n",
+    "\n",
+    "    # === Временные эмбеддинги ===\n",
+    "    \"time_embedding_type\": \"positional\",  # Тип временных эмбеддингов (стандартный подход)\n",
+    "\n",
+    "    # === Свертки ===\n",
+    "    \"conv_in_kernel\": 3,              # Ядро входной свертки (баланс между рецептивным полем и параметрами)\n",
+    "    \"conv_out_kernel\": 3,             # Ядро выходной свертки (симметрично входной)\n",
+    "}\n",
+    "\n",
+    "if 1:\n",
+    "    checkpoint_path = \"/workspace/sdxs/sdxs_flux\"#\"sdxs\"\n",
+    "    import torch\n",
+    "    from diffusers import UNet2DConditionModel\n",
+    "    print(\"test unet\")\n",
+    "    new_unet = UNet2DConditionModel(**config_sdxs).to(\"cuda\", dtype=torch.float16)\n",
+    "\n",
+    "    assert all(ch % 32 == 0 for ch in new_unet.config[\"block_out_channels\"]), \"Каналы должны быть кратны 32\"\n",
+    "    num_params = sum(p.numel() for p in new_unet.parameters())\n",
+    "    print(f\"Количество параметров: {num_params}\")\n",
+    "\n",
+    "    # Генерация тестового латента (640x512 в latent space)\n",
+    "    test_latent = torch.randn(1, 32, 60, 48).to(\"cuda\", dtype=torch.float16)  # 60x48 ≈ 512px\n",
+    "    timesteps = torch.tensor([1]).to(\"cuda\", dtype=torch.float16)\n",
+    "    encoder_hidden_states = torch.randn(1, 77, 1024).to(\"cuda\", dtype=torch.float16)\n",
+    "    \n",
+    "    with torch.no_grad():\n",
+    "        output = new_unet(\n",
+    "            test_latent, \n",
+    "            timesteps, \n",
+    "            encoder_hidden_states\n",
+    "        ).sample\n",
+    "        \n",
+    "    print(f\"Output shape: {output.shape}\")  \n",
+    "    new_unet.save_pretrained(checkpoint_path)\n",
+    "    print(new_unet)\n",
+    "    del new_unet\n",
+    "    torch.cuda.empty_cache()\n",
+    "    print(f\"Output shape: {output.shape}\") \n",
+    "    # Количество параметров: 1228819856 2042991632 1889899536"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a81bca2d-d96d-4794-90f4-b54ea8682b52",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

.ipynb_checkpoints/train_flux-checkpoint.py

@@ -0,0 +1,802 @@
+#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
+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/datasets/mjnj"
+project = "sdxs_flux"
+batch_size = 32
+base_learning_rate = 4e-5 #2.7e-5
+min_learning_rate = 9e-6 #2.7e-5
+num_epochs = 10
+sample_interval_share = 10
+cfg_dropout = 0.9
+max_length = 192
+use_wandb = True
+use_comet_ml = False
+save_model = True
+use_decay = True
+fbp = False
+optimizer_type = "adam8bit"
+torch_compile = False
+unet_gradient = True
+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 = 1.01
+warmup_percent = 0.01
+percentile_clipping = 95 #96 #97
+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":   1.5,
+    "mae":   0.5,
+}
+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.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
+        #self.window = window_steps
+        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)
+
+# --------------------------- Ин��циализация 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("vae1x", torch_dtype=dtype).to("cpu").eval()
+vae = AutoencoderKLFlux2.from_pretrained("black-forest-labs/FLUX.2-dev",subfolder="vae",torch_dtype=dtype).to(device).eval()
+tokenizer = AutoTokenizer.from_pretrained("tokenizer")
+text_model = AutoModel.from_pretrained("text_encoder").to(device).eval()
+
+# --- [UPDATED] Функция кодирования текста (с маской и пулингом) ---
+def encode_texts(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)
+        
+        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
+
+from diffusers import FlowMatchEulerDiscreteScheduler
+num_train_timesteps = 1000
+scheduler = FlowMatchEulerDiscreteScheduler(num_train_timesteps=num_train_timesteps)
+
+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 пр��ведем 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:
+    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
+
+            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)
+            # берём t из [0, 1]
+            #t = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
+            u = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
+            t = torch.sigmoid(torch.randn_like(u))
+            
+            # 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()
+            
+            # --- 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()
+                
+            accelerator.backward(total_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)
+                    if grad_val is not None:
+                        grad = float(grad_val)
+                        print("grad_val is None")
+                    else:
+                        grad = 0.0 
+                    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!")

config.json

@@ -0,0 +1,78 @@
+{
+  "_class_name": "UNet2DConditionModel",
+  "_diffusers_version": "0.36.0",
+  "_name_or_path": "sdxs_flux",
+  "act_fn": "silu",
+  "addition_embed_type": null,
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": null,
+  "attention_head_dim": [
+    2,
+    4,
+    8,
+    16
+  ],
+  "attention_type": "default",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    1024
+  ],
+  "center_input_sample": false,
+  "class_embed_type": null,
+  "class_embeddings_concat": false,
+  "conv_in_kernel": 3,
+  "conv_out_kernel": 3,
+  "cross_attention_dim": 1024,
+  "cross_attention_norm": null,
+  "down_block_types": [
+    "DownBlock2D",
+    "DownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "dropout": 0.0,
+  "dual_cross_attention": false,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "in_channels": 32,
+  "layers_per_block": 2,
+  "mid_block_only_cross_attention": null,
+  "mid_block_scale_factor": 1.0,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "out_channels": 32,
+  "projection_class_embeddings_input_dim": null,
+  "resnet_out_scale_factor": 1.0,
+  "resnet_skip_time_act": false,
+  "resnet_time_scale_shift": "default",
+  "reverse_transformer_layers_per_block": null,
+  "sample_size": null,
+  "time_cond_proj_dim": null,
+  "time_embedding_act_fn": null,
+  "time_embedding_dim": null,
+  "time_embedding_type": "positional",
+  "timestep_post_act": null,
+  "transformer_layers_per_block": [
+    1,
+    1,
+    2,
+    4
+  ],
+  "up_block_types": [
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D",
+    "UpBlock2D",
+    "UpBlock2D"
+  ],
+  "upcast_attention": false,
+  "use_linear_projection": true
+}

dataset_flux.py

@@ -0,0 +1,413 @@
+# pip install flash-attn --no-build-isolation
+import torch
+import os
+import gc
+import numpy as np
+import random
+import json
+import shutil
+import time
+
+from datasets import Dataset, load_from_disk, concatenate_datasets
+from diffusers import AutoencoderKL,AutoencoderKLWan,AutoencoderKLFlux2
+from torchvision.transforms import Resize, ToTensor, Normalize, Compose, InterpolationMode, Lambda
+from transformers import AutoModel, AutoImageProcessor, AutoTokenizer, AutoModelForCausalLM
+from typing import Dict, List, Tuple, Optional, Any
+from PIL import Image
+from tqdm import tqdm
+from datetime import timedelta
+
+# ---------------- 1️⃣ Настройки ----------------
+dtype = torch.float32
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+batch_size = 5
+min_size = 192 #384 #320 #192 #256 #192
+max_size = 320 #768 #640 #384 #256 #384
+step = 64 #64
+empty_share = 0.0
+limit = 0
+# Основная процедура обработки
+folder_path = "/workspace/mjnj" #alchemist"
+save_path = "/workspace/sdxs/datasets/mjnj" #"alchemist"
+os.makedirs(save_path, exist_ok=True)
+
+# Функция для очистки CUDA памяти
+def clear_cuda_memory():
+    if torch.cuda.is_available():
+        used_gb = torch.cuda.max_memory_allocated() / 1024**3
+        print(f"used_gb: {used_gb:.2f} GB")
+        torch.cuda.empty_cache()
+        gc.collect()
+
+# ---------------- 2️⃣ Загрузка моделей ----------------
+def load_models():
+    print("Загрузка моделей...")
+    #vae = AutoencoderKL.from_pretrained("AiArtLab/sdxs",subfolder="vae1x",torch_dtype=dtype).to(device).eval()
+    vae = AutoencoderKLFlux2.from_pretrained("black-forest-labs/FLUX.2-dev",subfolder="vae",torch_dtype=dtype).to(device).eval()
+
+    #model_name = "Qwen/Qwen3-0.6B"
+    #tokenizer = AutoTokenizer.from_pretrained(model_name)
+    #model = AutoModelForCausalLM.from_pretrained(
+    #    model_name,
+    #    torch_dtype=dtype,
+    #    device_map=device
+    #).eval()
+    #tokenizer = AutoTokenizer.from_pretrained('Qwen/Qwen3-Embedding-0.6B', padding_side='left')
+    #model = AutoModel.from_pretrained('Qwen/Qwen3-Embedding-0.6B').to("cuda")
+    return vae#, model, tokenizer
+
+#vae, model, tokenizer = load_models()
+vae = load_models()
+
+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)
+    
+# ---------------- 3️⃣ Трансформации ----------------
+def get_image_transform(min_size=256, max_size=512, step=64):
+    def transform(img, dry_run=False):
+        # Сохраняем исходные размеры изображения
+        original_width, original_height = img.size
+
+        # 0. Ресайз: масштабируем изображение, чтобы максимальная сторона была равна max_size
+        if original_width >= original_height:
+            new_width = max_size
+            new_height = int(max_size * original_height / original_width)
+        else:
+            new_height = max_size
+            new_width = int(max_size * original_width / original_height)
+
+        if new_height < min_size or new_width < min_size:
+            # 1. Ресайз: масштабируем изображение, чтобы минимальная сторона была равна min_size
+            if original_width <= original_height:
+                new_width = min_size
+                new_height = int(min_size * original_height / original_width)
+            else:
+                new_height = min_size
+                new_width = int(min_size * original_width / original_height)
+
+        # 2. Проверка: если одна из сторон превышает max_size, готовимся к обрезке
+        crop_width = min(max_size, (new_width // step) * step)
+        crop_height = min(max_size, (new_height // step) * step)
+
+        # Убеждаемся, что размеры обрезки не меньше min_size
+        crop_width = max(min_size, crop_width)
+        crop_height = max(min_size, crop_height)
+        
+        # Если запрошен только предварительный расчёт размеров
+        if dry_run:
+            return crop_width, crop_height
+        
+        # Конвертация в RGB и ресайз
+        img_resized = img.convert("RGB").resize((new_width, new_height), Image.LANCZOS)
+        
+        # Определение координат обрезки (обрезаем с учетом вотермарок - треть сверху)
+        top = (new_height - crop_height) // 3
+        left = 0
+        
+        # Обрезка изображения
+        img_cropped = img_resized.crop((left, top, left + crop_width, top + crop_height))
+        
+        # Сохраняем итоговые размеры после всех преобразований
+        final_width, final_height = img_cropped.size
+        
+        # тензор
+        img_tensor = ToTensor()(img_cropped)
+        img_tensor = Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])(img_tensor)
+        return img_tensor, img_cropped, final_width, final_height
+
+    return transform
+
+# ---------------- 4️⃣ Функции обработки ----------------
+def last_token_pool(last_hidden_states: torch.Tensor,
+                    attention_mask: torch.Tensor) -> torch.Tensor:
+    # Определяем, есть ли left padding
+    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
+    if left_padding:
+        return last_hidden_states[:, -1]
+    else:
+        sequence_lengths = attention_mask.sum(dim=1) - 1
+        batch_size = last_hidden_states.shape[0]
+        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]
+
+def encode_texts_batch(texts, tokenizer, model, device="cuda", max_length=150, normalize=False):
+    with torch.inference_mode():
+        # Токенизация
+        batch = tokenizer(
+            texts,
+            return_tensors="pt",
+            padding="max_length",
+            truncation=True,
+            max_length=max_length
+        ).to(device)
+
+        # Прогон через модель
+        #outputs = model(**batch)
+
+        # Пулинг по last token
+        #embeddings = last_token_pool(outputs.last_hidden_state, batch["attention_mask"])
+
+        # L2-нормализация (опционально, обычно нужна для семантического поиска)
+        #if normalize:
+        #    embeddings = F.normalize(embeddings, p=2, dim=1)
+
+        # Прогон через базовую модель (внутри CausalLM)
+        outputs = model.model(**batch, output_hidden_states=True)
+
+        # Берем последний слой (эмбеддинги всех токенов)
+        hidden_states = outputs.hidden_states[-1]  # [B, L, D]
+
+        # Можно применить нормализацию по каждому токену (как в CLIP)
+        if normalize:
+            hidden_states = F.normalize(hidden_states, p=2, dim=-1)
+
+    return hidden_states.cpu().numpy() # embeddings.unsqueeze(1).cpu().numpy()
+
+def clean_label(label):
+    label = label.replace("Image 1", "").replace("Image 2", "").replace("Image 3", "").replace("Image 4", "").replace("The image depicts ","").replace("The image presents ","").replace("The image features ","").replace("The image portrays ","").replace("The image is ","").strip()
+    if label.startswith("."):
+        label = label[1:].lstrip()
+    return label
+
+def process_labels_for_guidance(original_labels, prob_to_make_empty=0.01):
+    """
+    Обрабатывает список меток для classifier-free guidance.
+
+    С вероятностью prob_to_make_empty:
+    - Метка в первом списке заменяется на пустую строку.
+    - К метке во втором списке добавляется префикс "zero:".
+
+    В противном случае метки в обоих списках остаются оригинальными.
+
+    """
+    labels_for_model = []
+    labels_for_logging = []
+
+    for label in original_labels:
+        if random.random() < prob_to_make_empty:
+            labels_for_model.append("")  # Заменяем на пустую строку для модели
+            labels_for_logging.append(f"zero: {label}") # Добавляем префикс для логгирования
+        else:
+            labels_for_model.append(label) # Оставляем оригинальную метку для модели
+            labels_for_logging.append(label) # Оставляем оригинальную метку для логгирования
+
+    return labels_for_model, labels_for_logging
+
+def encode_to_latents(images, texts):
+    transform = get_image_transform(min_size, max_size, step)
+    
+    try:
+        # Обработка изображений (все одинакового размера)
+        transformed_tensors = []
+        pil_images = []
+        widths, heights = [], []
+        
+        # Применяем трансформацию ко всем изображениям
+        for img in images:
+            try:
+                t_img, pil_img, w, h = transform(img)
+                transformed_tensors.append(t_img)
+                pil_images.append(pil_img)
+                widths.append(w)
+                heights.append(h)
+            except Exception as e:
+                print(f"Ошибка трансформации: {e}")
+                continue
+
+        if not transformed_tensors:
+            return None
+
+        # Создаём батч
+        batch_tensor = torch.stack(transformed_tensors).to(device, dtype)
+        if batch_tensor.ndim==5:
+            batch_tensor = batch_tensor.unsqueeze(2)  # [B, C, 1, H, W]
+        
+        # Кодируем батч
+        with torch.no_grad():
+            posteriors = vae.encode(batch_tensor).latent_dist.mode()
+            latents = (posteriors - shift_factor) / scaling_factor
+        
+        latents_np = latents.to(dtype).cpu().numpy()
+
+        # Обрабатываем тексты
+        text_labels = [clean_label(text) for text in texts]
+
+        model_prompts, text_labels = process_labels_for_guidance(text_labels, empty_share)                
+        #embeddings = encode_texts_batch(model_prompts, tokenizer, model)
+
+        return {
+            "vae": latents_np,
+            #"embeddings": embeddings,
+            "text": text_labels,
+            "width": widths,
+            "height": heights
+        }
+        
+    except Exception as e:
+        print(f"Критическая ошибка в encode_to_latents: {e}")
+        raise
+        
+
+# ---------------- 5️⃣ Обработка папки с изображениями и текстами ----------------
+def process_folder(folder_path, limit=None):
+    """
+    Рекурсивно обходит указанную директорию и все вложенные директории,
+    собирая пути к изображениям и соответствующим текстовым файлам.
+    """
+    image_paths = []
+    text_paths = []
+    width = []
+    height = []
+    transform = get_image_transform(min_size, max_size, step)
+    
+    # Используем os.walk для рекурсивного обхода директорий
+    for root, dirs, files in os.walk(folder_path):
+        for filename in files:
+            # Проверяем, является ли файл изображением
+            if filename.lower().endswith((".jpg", ".jpeg", ".png")):
+                image_path = os.path.join(root, filename)
+                try:
+                    img = Image.open(image_path)
+                except Exception as e:
+                    print(f"Ошибка при открытии {image_path}: {e}")
+                    os.remove(image_path)
+                    text_path = os.path.splitext(image_path)[0] + ".txt"
+                    if os.path.exists(text_path):
+                        os.remove(text_path)
+                    continue
+                # Применяем трансформацию только для получения размеров
+                w, h = transform(img, dry_run=True)
+                # Формируем путь к текстовому файлу
+                text_path = os.path.splitext(image_path)[0] + ".txt"
+                
+                # Добавляем пути, если текстовый файл существует
+                if os.path.exists(text_path) and min(w, h)>0:
+                    image_paths.append(image_path)
+                    text_paths.append(text_path)
+                    width.append(w)  # Добавляем в список
+                    height.append(h)  # Добавляем в список
+                    
+                    # Проверяем ограничение на количество
+                    if limit and limit>0 and len(image_paths) >= limit:
+                        print(f"Достигнут лимит в {limit} изображений")
+                        return image_paths, text_paths, width, height
+    
+    print(f"Найдено {len(image_paths)} изображений с текстовыми описаниями")
+    return image_paths, text_paths, width, height
+    
+def process_in_chunks(image_paths, text_paths, width, height, chunk_size=10000, batch_size=1):
+    total_files = len(image_paths)
+    start_time = time.time()
+    chunks = range(0, total_files, chunk_size)
+    
+    for chunk_idx, start in enumerate(chunks, 1):
+        end = min(start + chunk_size, total_files)
+        chunk_image_paths = image_paths[start:end]
+        chunk_text_paths = text_paths[start:end]
+        chunk_widths = width[start:end] if isinstance(width, list) else [width] * len(chunk_image_paths)
+        chunk_heights = height[start:end] if isinstance(height, list) else [height] * len(chunk_image_paths)
+        
+        # Чтение текстов
+        chunk_texts = []
+        for text_path in chunk_text_paths:
+            try:
+                with open(text_path, 'r', encoding='utf-8') as f:
+                    text = f.read().strip()
+                chunk_texts.append(text)
+            except Exception as e:
+                print(f"Ошибка чтения {text_path}: {e}")
+                chunk_texts.append("")
+        
+        # Группируем изображения по размерам
+        size_groups = {}
+        for i in range(len(chunk_image_paths)):
+            size_key = (chunk_widths[i], chunk_heights[i])
+            if size_key not in size_groups:
+                size_groups[size_key] = {"image_paths": [], "texts": []}
+            size_groups[size_key]["image_paths"].append(chunk_image_paths[i])
+            size_groups[size_key]["texts"].append(chunk_texts[i])
+        
+        # Обрабатываем каждую группу размеров отдельно
+        for size_key, group_data in size_groups.items():
+            print(f"Обработка группы с размером {size_key[0]}x{size_key[1]} - {len(group_data['image_paths'])} изображений")
+            
+            group_dataset = Dataset.from_dict({
+                "image_path": group_data["image_paths"],
+                "text": group_data["texts"]
+            })
+            
+            # Теперь можно использовать указанный batch_size, т.к. все изображения одного размера
+            processed_group = group_dataset.map(
+                lambda examples: encode_to_latents(
+                    [Image.open(path) for path in examples["image_path"]],
+                    examples["text"]
+                ),
+                batched=True,
+                batch_size=batch_size,
+                #remove_columns=["image_path"],
+                desc=f"Обработка группы размера {size_key[0]}x{size_key[1]}"
+            )
+            
+            # Сохраняем результаты группы
+            group_save_path = f"{save_path}_temp/chunk_{chunk_idx}_size_{size_key[0]}x{size_key[1]}"
+            processed_group.save_to_disk(group_save_path)
+            clear_cuda_memory()
+            elapsed = time.time() - start_time
+            processed = (chunk_idx - 1) * chunk_size + sum([len(sg["image_paths"]) for sg in list(size_groups.values())[:list(size_groups.values()).index(group_data) + 1]])
+            if processed > 0:
+                remaining = (elapsed / processed) * (total_files - processed)
+                elapsed_str = str(timedelta(seconds=int(elapsed)))
+                remaining_str = str(timedelta(seconds=int(remaining)))
+                print(f"ETA: Прошло {elapsed_str}, Осталось {remaining_str}, Прогресс {processed}/{total_files} ({processed/total_files:.1%})")
+
+# ---------------- 7️⃣ Объединение чанков ----------------
+def combine_chunks(temp_path, final_path):
+    """Объединение обработанных чанков в финальный датасет"""
+    chunks = sorted([
+        os.path.join(temp_path, d) 
+        for d in os.listdir(temp_path) 
+        if d.startswith("chunk_")
+    ])
+    
+    datasets = [load_from_disk(chunk) for chunk in chunks]
+    combined = concatenate_datasets(datasets)
+    combined.save_to_disk(final_path)
+    
+    print(f"✅ Датасет успешно сохранен в: {final_path}")
+
+    
+
+# Создаем временную папку для чанков
+temp_path = f"{save_path}_temp"
+os.makedirs(temp_path, exist_ok=True)
+
+# Получаем список файлов
+image_paths, text_paths, width, height = process_folder(folder_path,limit)
+print(f"Всего найдено {len(image_paths)} изображений")
+
+# Обработка с чанкованием
+process_in_chunks(image_paths, text_paths, width, height, chunk_size=20000, batch_size=batch_size)
+
+# Удаление  папки
+try:
+    shutil.rmtree(folder_path)
+    print(f"✅ Папка {folder_path} успешно удалена")
+except Exception as e:
+    print(f"⚠️ Ошибка при удалении папки: {e}")
+    
+# Объединение чанков в финальный датасет
+combine_chunks(temp_path, save_path)
+
+# Удаление временной папки
+try:
+    shutil.rmtree(temp_path)
+    print(f"✅ Временная папка {temp_path} успешно удалена")
+except Exception as e:
+    print(f"⚠️ Ошибка при удалении временной папки: {e}")

diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:41dfe79753fadad9ba8e148ac73e1b61f23c7a1146cece5147c0041d613298bd
+size 3195253456

sdxs_create_flux.ipynb

@@ -0,0 +1,600 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "6bf71a1a-1bf0-42c7-8709-6686e8d2f46c",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "test unet\n",
+      "Количество параметров: 798780960\n",
+      "Output shape: torch.Size([1, 32, 60, 48])\n",
+      "UNet2DConditionModel(\n",
+      "  (conv_in): Conv2d(32, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "  (time_proj): Timesteps()\n",
+      "  (time_embedding): TimestepEmbedding(\n",
+      "    (linear_1): Linear(in_features=128, out_features=512, bias=True)\n",
+      "    (act): SiLU()\n",
+      "    (linear_2): Linear(in_features=512, out_features=512, bias=True)\n",
+      "  )\n",
+      "  (down_blocks): ModuleList(\n",
+      "    (0): DownBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=128, bias=True)\n",
+      "          (norm2): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (1): DownBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(128, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (2): CrossAttnDownBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-1): 2 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 512, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-1): 2 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=512, out_features=4096, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2048, out_features=512, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=512, out_features=512, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(256, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "      (downsamplers): ModuleList(\n",
+      "        (0): Downsample2D(\n",
+      "          (conv): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (3): CrossAttnDownBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-1): 2 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 1024, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-3): 4 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1024, out_features=8192, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4096, out_features=1024, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(512, 1024, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (up_blocks): ModuleList(\n",
+      "    (0): CrossAttnUpBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-2): 3 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 1024, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-3): 4 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=1024, out_features=8192, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=4096, out_features=1024, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0-1): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 2048, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(2048, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(2048, 1024, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1536, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1536, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "          (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1536, 1024, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (1): CrossAttnUpBlock2D(\n",
+      "      (attentions): ModuleList(\n",
+      "        (0-2): 3 x Transformer2DModel(\n",
+      "          (norm): GroupNorm(32, 512, eps=1e-06, affine=True)\n",
+      "          (proj_in): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (transformer_blocks): ModuleList(\n",
+      "            (0-1): 2 x BasicTransformerBlock(\n",
+      "              (norm1): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn1): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm2): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (attn2): Attention(\n",
+      "                (to_q): Linear(in_features=512, out_features=512, bias=False)\n",
+      "                (to_k): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_v): Linear(in_features=1024, out_features=512, bias=False)\n",
+      "                (to_out): ModuleList(\n",
+      "                  (0): Linear(in_features=512, out_features=512, bias=True)\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                )\n",
+      "              )\n",
+      "              (norm3): LayerNorm((512,), eps=1e-05, elementwise_affine=True)\n",
+      "              (ff): FeedForward(\n",
+      "                (net): ModuleList(\n",
+      "                  (0): GEGLU(\n",
+      "                    (proj): Linear(in_features=512, out_features=4096, bias=True)\n",
+      "                  )\n",
+      "                  (1): Dropout(p=0.0, inplace=False)\n",
+      "                  (2): Linear(in_features=2048, out_features=512, bias=True)\n",
+      "                )\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "          (proj_out): Linear(in_features=512, out_features=512, bias=True)\n",
+      "        )\n",
+      "      )\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1536, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1536, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1536, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(1024, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 768, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(768, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=512, bias=True)\n",
+      "          (norm2): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(768, 512, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (2): UpBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 768, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(768, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 512, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(512, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (2): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 384, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=256, bias=True)\n",
+      "          (norm2): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(384, 256, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "      (upsamplers): ModuleList(\n",
+      "        (0): Upsample2D(\n",
+      "          (conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (3): UpBlock2D(\n",
+      "      (resnets): ModuleList(\n",
+      "        (0): ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 384, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(384, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=128, bias=True)\n",
+      "          (norm2): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(384, 128, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "        (1-2): 2 x ResnetBlock2D(\n",
+      "          (norm1): GroupNorm(32, 256, eps=1e-05, affine=True)\n",
+      "          (conv1): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (time_emb_proj): Linear(in_features=512, out_features=128, bias=True)\n",
+      "          (norm2): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "          (dropout): Dropout(p=0.0, inplace=False)\n",
+      "          (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "          (nonlinearity): SiLU()\n",
+      "          (conv_shortcut): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1))\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (mid_block): UNetMidBlock2DCrossAttn(\n",
+      "    (attentions): ModuleList(\n",
+      "      (0): Transformer2DModel(\n",
+      "        (norm): GroupNorm(32, 1024, eps=1e-06, affine=True)\n",
+      "        (proj_in): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "        (transformer_blocks): ModuleList(\n",
+      "          (0-3): 4 x BasicTransformerBlock(\n",
+      "            (norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn1): Attention(\n",
+      "              (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "            (attn2): Attention(\n",
+      "              (to_q): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_k): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_v): Linear(in_features=1024, out_features=1024, bias=False)\n",
+      "              (to_out): ModuleList(\n",
+      "                (0): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "              )\n",
+      "            )\n",
+      "            (norm3): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)\n",
+      "            (ff): FeedForward(\n",
+      "              (net): ModuleList(\n",
+      "                (0): GEGLU(\n",
+      "                  (proj): Linear(in_features=1024, out_features=8192, bias=True)\n",
+      "                )\n",
+      "                (1): Dropout(p=0.0, inplace=False)\n",
+      "                (2): Linear(in_features=4096, out_features=1024, bias=True)\n",
+      "              )\n",
+      "            )\n",
+      "          )\n",
+      "        )\n",
+      "        (proj_out): Linear(in_features=1024, out_features=1024, bias=True)\n",
+      "      )\n",
+      "    )\n",
+      "    (resnets): ModuleList(\n",
+      "      (0-1): 2 x ResnetBlock2D(\n",
+      "        (norm1): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "        (conv1): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (time_emb_proj): Linear(in_features=512, out_features=1024, bias=True)\n",
+      "        (norm2): GroupNorm(32, 1024, eps=1e-05, affine=True)\n",
+      "        (dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (conv2): Conv2d(1024, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      "        (nonlinearity): SiLU()\n",
+      "      )\n",
+      "    )\n",
+      "  )\n",
+      "  (conv_norm_out): GroupNorm(32, 128, eps=1e-05, affine=True)\n",
+      "  (conv_act): SiLU()\n",
+      "  (conv_out): Conv2d(128, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n",
+      ")\n",
+      "Output shape: torch.Size([1, 32, 60, 48])\n"
+     ]
+    }
+   ],
+   "source": [
+    "config_sdxs = {\n",
+    "    # === Основные размеры и каналы ===\n",
+    "    \"in_channels\": 32,               # Количество входных каналов (совместимость с 16-канальным VAE)\n",
+    "    \"out_channels\": 32,              # Количество выходных каналов (симметрично in_channels)\n",
+    "    \"center_input_sample\": False,    # Отключение центрирования входных данных (стандарт для диффузионных моделей)\n",
+    "    \"flip_sin_to_cos\": True,         # Автоматическое преобразование sin/cos в эмбеддингах времени (для стабильности)\n",
+    "    \"freq_shift\": 0,                 # Сдвиг частоты (0 - стандартное значение для частотных эмбеддингов)\n",
+    "\n",
+    "    # === Архитектура блоков ===\n",
+    "    \"down_block_types\": [            # Типы блоков энкодера (иерархия обработки):\n",
+    "        \"DownBlock2D\",\n",
+    "        \"DownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "        \"CrossAttnDownBlock2D\",\n",
+    "    ],\n",
+    "    \"mid_block_type\": \"UNetMidBlock2DCrossAttn\",  # Центральный блок с cross-attention (бутылочное горлышко сети)\n",
+    "    \"up_block_types\": [              # Типы блоков декодера (восстановление изображения):\n",
+    "        \"CrossAttnUpBlock2D\",\n",
+    "        \"CrossAttnUpBlock2D\",        \n",
+    "        \"UpBlock2D\",\n",
+    "        \"UpBlock2D\",\n",
+    "    ],\n",
+    "    \"only_cross_attention\": False,   # Использование как cross-attention, так и self-attention\n",
+    "\n",
+    "    # === Конфигурация каналов ===\n",
+    "    \"block_out_channels\": [128, 256, 512, 1024], \n",
+    "    \"layers_per_block\": 2, # Число слоев в блоках\n",
+    "    \"downsample_padding\": 1,          # Паддинг при уменьшении разрешения\n",
+    "    \"mid_block_scale_factor\": 1.0,    # Усиление сигнала в центральном блоке\n",
+    "\n",
+    "    # === Нормализация ===\n",
+    "    \"norm_num_groups\": 32,            # Число групп для GroupNorm (оптимально для стабильности)\n",
+    "    \"norm_eps\": 1e-05,                # Эпсилон для нормализации (стандартное значение)\n",
+    "\n",
+    "    # === Cross-Attention ===\n",
+    "    \"cross_attention_dim\": 1024,      # Размерность текстовых эмбеддинго\n",
+    "    \n",
+    "    \"transformer_layers_per_block\": [1, 1, 2, 4],  # Число трансформерных слоев (уменьшение с глубиной)\n",
+    "    \"attention_head_dim\": [2, 4, 8, 16],  # Размерность головы внимания \n",
+    "    \"dual_cross_attention\": False,    # Отключение двойного внимания (упрощение архитектуры)\n",
+    "    \"use_linear_projection\": True,    # Изменено на True для лучшей организации памяти\n",
+    "\n",
+    "    # === ResNet Блоки ===\n",
+    "    \"resnet_time_scale_shift\": \"default\",  # Способ интеграции временных эмбеддингов\n",
+    "    \"resnet_skip_time_act\": False,    # Отключение активации в skip-соединениях\n",
+    "    \"resnet_out_scale_factor\": 1.0,   # Коэффициент масштабирования выхода ResNet\n",
+    "\n",
+    "    # === Временные эмбеддинги ===\n",
+    "    \"time_embedding_type\": \"positional\",  # Тип временных эмбеддингов (стандартный подход)\n",
+    "\n",
+    "    # === Свертки ===\n",
+    "    \"conv_in_kernel\": 3,              # Ядро входной свертки (баланс между рецептивным полем и параметрами)\n",
+    "    \"conv_out_kernel\": 3,             # Ядро выходной свертки (симметрично входной)\n",
+    "}\n",
+    "\n",
+    "if 1:\n",
+    "    checkpoint_path = \"/workspace/sdxs/sdxs_flux\"#\"sdxs\"\n",
+    "    import torch\n",
+    "    from diffusers import UNet2DConditionModel\n",
+    "    print(\"test unet\")\n",
+    "    new_unet = UNet2DConditionModel(**config_sdxs).to(\"cuda\", dtype=torch.float16)\n",
+    "\n",
+    "    assert all(ch % 32 == 0 for ch in new_unet.config[\"block_out_channels\"]), \"Каналы должны быть кратны 32\"\n",
+    "    num_params = sum(p.numel() for p in new_unet.parameters())\n",
+    "    print(f\"Количество параметров: {num_params}\")\n",
+    "\n",
+    "    # Генерация тестового латента (640x512 в latent space)\n",
+    "    test_latent = torch.randn(1, 32, 60, 48).to(\"cuda\", dtype=torch.float16)  # 60x48 ≈ 512px\n",
+    "    timesteps = torch.tensor([1]).to(\"cuda\", dtype=torch.float16)\n",
+    "    encoder_hidden_states = torch.randn(1, 77, 1024).to(\"cuda\", dtype=torch.float16)\n",
+    "    \n",
+    "    with torch.no_grad():\n",
+    "        output = new_unet(\n",
+    "            test_latent, \n",
+    "            timesteps, \n",
+    "            encoder_hidden_states\n",
+    "        ).sample\n",
+    "        \n",
+    "    print(f\"Output shape: {output.shape}\")  \n",
+    "    new_unet.save_pretrained(checkpoint_path)\n",
+    "    print(new_unet)\n",
+    "    del new_unet\n",
+    "    torch.cuda.empty_cache()\n",
+    "    print(f\"Output shape: {output.shape}\") \n",
+    "    # Количество параметров: 1228819856 2042991632 1889899536"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a81bca2d-d96d-4794-90f4-b54ea8682b52",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

train_flux.py

@@ -0,0 +1,802 @@
+#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
+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/datasets/mjnj"
+project = "sdxs_flux"
+batch_size = 32
+base_learning_rate = 4e-5 #2.7e-5
+min_learning_rate = 9e-6 #2.7e-5
+num_epochs = 10
+sample_interval_share = 10
+cfg_dropout = 0.9
+max_length = 192
+use_wandb = True
+use_comet_ml = False
+save_model = True
+use_decay = True
+fbp = False
+optimizer_type = "adam8bit"
+torch_compile = False
+unet_gradient = True
+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 = 1.01
+warmup_percent = 0.01
+percentile_clipping = 95 #96 #97
+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":   1.5,
+    "mae":   0.5,
+}
+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.buffers = {k: deque(maxlen=window_steps) for k in self.ratios.keys()}
+        #self.window = window_steps
+        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)
+
+# --------------------------- Ин��циализация 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("vae1x", torch_dtype=dtype).to("cpu").eval()
+vae = AutoencoderKLFlux2.from_pretrained("black-forest-labs/FLUX.2-dev",subfolder="vae",torch_dtype=dtype).to(device).eval()
+tokenizer = AutoTokenizer.from_pretrained("tokenizer")
+text_model = AutoModel.from_pretrained("text_encoder").to(device).eval()
+
+# --- [UPDATED] Функция кодирования текста (с маской и пулингом) ---
+def encode_texts(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)
+        
+        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
+
+from diffusers import FlowMatchEulerDiscreteScheduler
+num_train_timesteps = 1000
+scheduler = FlowMatchEulerDiscreteScheduler(num_train_timesteps=num_train_timesteps)
+
+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 пр��ведем 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:
+    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
+
+            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)
+            # берём t из [0, 1]
+            #t = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
+            u = torch.rand(latents.shape[0], device=latents.device, dtype=latents.dtype)
+            t = torch.sigmoid(torch.randn_like(u))
+            
+            # 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()
+            
+            # --- 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()
+                
+            accelerator.backward(total_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)
+                    if grad_val is not None:
+                        grad = float(grad_val)
+                        print("grad_val is None")
+                    else:
+                        grad = 0.0 
+                    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!")